WEBVTT Kind: captions Language: en 00:00:12.280 --> 00:00:14.070 align:start position:0% hello<00:00:12.519> everybody<00:00:12.960> I'm<00:00:13.080> Alexi<00:00:13.440> krau<00:00:13.960> the 00:00:14.070 --> 00:00:14.080 align:start position:0% hello everybody I'm Alexi krau the 00:00:14.080 --> 00:00:16.630 align:start position:0% hello everybody I'm Alexi krau the founder<00:00:14.400> and<00:00:14.559> organizer<00:00:15.120> of<00:00:15.320> beera<00:00:15.719> AI<00:00:16.480> the 00:00:16.630 --> 00:00:16.640 align:start position:0% founder and organizer of beera AI the 00:00:16.640 --> 00:00:18.510 align:start position:0% founder and organizer of beera AI the longest<00:00:17.000> running<00:00:17.520> deepest<00:00:17.920> technical 00:00:18.510 --> 00:00:18.520 align:start position:0% longest running deepest technical 00:00:18.520 --> 00:00:20.870 align:start position:0% longest running deepest technical biggest<00:00:19.000> AI<00:00:19.240> metup<00:00:19.600> in<00:00:19.680> the<00:00:19.800> world<00:00:20.439> running<00:00:20.760> in 00:00:20.870 --> 00:00:20.880 align:start position:0% biggest AI metup in the world running in 00:00:20.880 --> 00:00:23.390 align:start position:0% biggest AI metup in the world running in the<00:00:21.000> Bay<00:00:21.199> Area<00:00:21.480> continuously<00:00:22.000> since<00:00:22.400> 2015 00:00:23.390 --> 00:00:23.400 align:start position:0% the Bay Area continuously since 2015 00:00:23.400 --> 00:00:26.589 align:start position:0% the Bay Area continuously since 2015 live<00:00:24.000> on<00:00:24.160> location<00:00:24.760> AWS<00:00:25.320> ji<00:00:25.720> LOF<00:00:26.240> and<00:00:26.400> here 00:00:26.589 --> 00:00:26.599 align:start position:0% live on location AWS ji LOF and here 00:00:26.599 --> 00:00:28.830 align:start position:0% live on location AWS ji LOF and here with<00:00:26.760> us<00:00:27.039> we<00:00:27.199> have<00:00:27.480> Michael<00:00:27.840> going<00:00:28.480> who<00:00:28.640> is<00:00:28.760> the 00:00:28.830 --> 00:00:28.840 align:start position:0% with us we have Michael going who is the 00:00:28.840 --> 00:00:33.229 align:start position:0% with us we have Michael going who is the engineering<00:00:29.359> lead<00:00:30.160> at<00:00:30.400> neural<00:00:31.160> magic<00:00:32.160> uh<00:00:32.439> and 00:00:33.229 --> 00:00:33.239 align:start position:0% engineering lead at neural magic uh and 00:00:33.239 --> 00:00:36.590 align:start position:0% engineering lead at neural magic uh and uh<00:00:33.520> they<00:00:33.640> are<00:00:33.760> the<00:00:33.879> maintainers<00:00:34.480> of<00:00:35.120> VM<00:00:36.120> uh 00:00:36.590 --> 00:00:36.600 align:start position:0% uh they are the maintainers of VM uh 00:00:36.600 --> 00:00:40.310 align:start position:0% uh they are the maintainers of VM uh welcome<00:00:36.960> and<00:00:37.280> please<00:00:37.480> tell<00:00:37.640> us<00:00:37.840> what<00:00:38.120> VM<00:00:38.760> is<00:00:39.760> VM 00:00:40.310 --> 00:00:40.320 align:start position:0% welcome and please tell us what VM is VM 00:00:40.320 --> 00:00:43.709 align:start position:0% welcome and please tell us what VM is VM is<00:00:40.559> the<00:00:40.920> is<00:00:41.160> the<00:00:42.000> most<00:00:42.360> used<00:00:43.039> open-source 00:00:43.709 --> 00:00:43.719 align:start position:0% is the is the most used open-source 00:00:43.719 --> 00:00:45.869 align:start position:0% is the is the most used open-source inference<00:00:44.160> server<00:00:44.840> uh<00:00:45.000> around<00:00:45.239> the<00:00:45.360> world 00:00:45.869 --> 00:00:45.879 align:start position:0% inference server uh around the world 00:00:45.879 --> 00:00:48.590 align:start position:0% inference server uh around the world it's<00:00:46.320> uh<00:00:46.520> loved<00:00:46.920> for<00:00:47.199> its<00:00:47.840> you<00:00:47.960> know<00:00:48.440> uh 00:00:48.590 --> 00:00:48.600 align:start position:0% it's uh loved for its you know uh 00:00:48.600 --> 00:00:51.430 align:start position:0% it's uh loved for its you know uh support<00:00:48.879> of<00:00:49.079> many<00:00:49.320> models<00:00:50.199> inference 00:00:51.430 --> 00:00:51.440 align:start position:0% support of many models inference 00:00:51.440 --> 00:00:54.510 align:start position:0% support of many models inference optimizations<00:00:52.440> uh<00:00:52.719> quantization<00:00:53.440> methods<00:00:54.399> uh 00:00:54.510 --> 00:00:54.520 align:start position:0% optimizations uh quantization methods uh 00:00:54.520 --> 00:00:55.950 align:start position:0% optimizations uh quantization methods uh and<00:00:54.640> of<00:00:54.760> course<00:00:55.079> constant<00:00:55.399> research<00:00:55.840> that 00:00:55.950 --> 00:00:55.960 align:start position:0% and of course constant research that 00:00:55.960 --> 00:00:58.029 align:start position:0% and of course constant research that gets<00:00:56.199> cont<00:00:56.640> contributed<00:00:57.120> into 00:00:58.029 --> 00:00:58.039 align:start position:0% gets cont contributed into 00:00:58.039 --> 00:01:01.349 align:start position:0% gets cont contributed into it<00:00:59.039> and<00:00:59.559> uh<00:01:00.039> why<00:01:00.120> is<00:01:00.199> it<00:01:00.320> so<00:01:00.519> widely<00:01:00.840> used<00:01:01.199> and 00:01:01.349 --> 00:01:01.359 align:start position:0% it and uh why is it so widely used and 00:01:01.359 --> 00:01:03.069 align:start position:0% it and uh why is it so widely used and why<00:01:01.440> is<00:01:01.519> it<00:01:01.640> so<00:01:01.800> Loft<00:01:02.280> like<00:01:02.440> it<00:01:02.559> has<00:01:02.680> a<00:01:02.800> fiber 00:01:03.069 --> 00:01:03.079 align:start position:0% why is it so Loft like it has a fiber 00:01:03.079 --> 00:01:05.670 align:start position:0% why is it so Loft like it has a fiber Comm<00:01:03.519> on<00:01:03.680> it<00:01:04.559> yeah<00:01:04.720> I<00:01:04.799> mean<00:01:05.000> we<00:01:05.159> have<00:01:05.400> you<00:01:05.519> know 00:01:05.670 --> 00:01:05.680 align:start position:0% Comm on it yeah I mean we have you know 00:01:05.680 --> 00:01:08.230 align:start position:0% Comm on it yeah I mean we have you know sort<00:01:05.880> of<00:01:06.520> many<00:01:06.840> companies<00:01:07.280> and<00:01:07.479> many<00:01:07.720> academic 00:01:08.230 --> 00:01:08.240 align:start position:0% sort of many companies and many academic 00:01:08.240 --> 00:01:10.910 align:start position:0% sort of many companies and many academic institutions<00:01:09.040> that<00:01:09.479> build<00:01:09.799> on<00:01:10.000> to<00:01:10.200> it<00:01:10.720> uh<00:01:10.840> you 00:01:10.910 --> 00:01:10.920 align:start position:0% institutions that build on to it uh you 00:01:10.920 --> 00:01:12.550 align:start position:0% institutions that build on to it uh you know<00:01:11.080> it's<00:01:11.280> it's<00:01:11.400> a<00:01:11.560> really<00:01:11.840> good<00:01:12.159> thing<00:01:12.400> when 00:01:12.550 --> 00:01:12.560 align:start position:0% know it's it's a really good thing when 00:01:12.560 --> 00:01:15.070 align:start position:0% know it's it's a really good thing when you<00:01:12.680> see<00:01:13.640> uh<00:01:13.799> you<00:01:13.880> know<00:01:14.040> dozens<00:01:14.400> of<00:01:14.600> papers 00:01:15.070 --> 00:01:15.080 align:start position:0% you see uh you know dozens of papers 00:01:15.080 --> 00:01:18.109 align:start position:0% you see uh you know dozens of papers every<00:01:15.360> year<00:01:16.119> based<00:01:16.520> off<00:01:16.680> of<00:01:16.920> VM<00:01:17.439> or<00:01:17.680> comparing 00:01:18.109 --> 00:01:18.119 align:start position:0% every year based off of VM or comparing 00:01:18.119 --> 00:01:20.670 align:start position:0% every year based off of VM or comparing to<00:01:18.320> VM<00:01:19.320> uh<00:01:19.439> you<00:01:19.520> know<00:01:19.920> uh<00:01:20.079> against<00:01:20.320> some<00:01:20.479> new 00:01:20.670 --> 00:01:20.680 align:start position:0% to VM uh you know uh against some new 00:01:20.680 --> 00:01:22.910 align:start position:0% to VM uh you know uh against some new method<00:01:21.000> that<00:01:21.119> they<00:01:21.280> came<00:01:21.520> off<00:01:21.720> of<00:01:22.560> uh<00:01:22.680> which<00:01:22.799> is 00:01:22.910 --> 00:01:22.920 align:start position:0% method that they came off of uh which is 00:01:22.920 --> 00:01:25.149 align:start position:0% method that they came off of uh which is usually<00:01:23.200> made<00:01:23.439> as<00:01:23.520> a<00:01:23.640> fork<00:01:24.079> a<00:01:24.240> VM<00:01:24.840> with<00:01:25.000> some 00:01:25.149 --> 00:01:25.159 align:start position:0% usually made as a fork a VM with some 00:01:25.159 --> 00:01:27.190 align:start position:0% usually made as a fork a VM with some new<00:01:25.360> idea<00:01:25.720> put<00:01:25.880> in<00:01:26.360> and<00:01:26.439> so<00:01:26.600> this<00:01:26.720> is<00:01:26.880> allowed 00:01:27.190 --> 00:01:27.200 align:start position:0% new idea put in and so this is allowed 00:01:27.200 --> 00:01:28.910 align:start position:0% new idea put in and so this is allowed for<00:01:27.360> a<00:01:27.439> lot<00:01:27.560> of<00:01:27.799> Rapid<00:01:28.079> development<00:01:28.479> and<00:01:28.640> Rapid 00:01:28.910 --> 00:01:28.920 align:start position:0% for a lot of Rapid development and Rapid 00:01:28.920 --> 00:01:31.149 align:start position:0% for a lot of Rapid development and Rapid integration<00:01:29.960> because<00:01:30.320> these<00:01:30.759> you<00:01:30.880> know 00:01:31.149 --> 00:01:31.159 align:start position:0% integration because these you know 00:01:31.159 --> 00:01:32.630 align:start position:0% integration because these you know promising<00:01:31.600> research<00:01:32.000> results<00:01:32.320> can<00:01:32.479> be 00:01:32.630 --> 00:01:32.640 align:start position:0% promising research results can be 00:01:32.640 --> 00:01:35.870 align:start position:0% promising research results can be quickly<00:01:33.320> validated<00:01:34.520> reproduced<00:01:35.520> and 00:01:35.870 --> 00:01:35.880 align:start position:0% quickly validated reproduced and 00:01:35.880 --> 00:01:39.190 align:start position:0% quickly validated reproduced and integrated<00:01:36.520> in<00:01:37.520> um<00:01:37.920> benefiting<00:01:38.360> everybody 00:01:39.190 --> 00:01:39.200 align:start position:0% integrated in um benefiting everybody 00:01:39.200 --> 00:01:42.350 align:start position:0% integrated in um benefiting everybody and<00:01:39.360> you're<00:01:39.960> gu<00:01:40.200> based<00:01:40.479> out<00:01:40.600> of<00:01:40.759> Boston<00:01:41.720> uh<00:01:42.119> yes 00:01:42.350 --> 00:01:42.360 align:start position:0% and you're gu based out of Boston uh yes 00:01:42.360 --> 00:01:44.389 align:start position:0% and you're gu based out of Boston uh yes yeah<00:01:42.520> we're<00:01:42.759> originally<00:01:43.119> an<00:01:43.280> MIT<00:01:43.640> startup<00:01:44.079> MIT 00:01:44.389 --> 00:01:44.399 align:start position:0% yeah we're originally an MIT startup MIT 00:01:44.399 --> 00:01:49.590 align:start position:0% yeah we're originally an MIT startup MIT start<00:01:44.880> I<00:01:44.960> wanted<00:01:45.200> to<00:01:45.360> ask<00:01:45.520> you<00:01:46.079> this<00:01:46.719> uh<00:01:47.719> so<00:01:48.719> MIT 00:01:49.590 --> 00:01:49.600 align:start position:0% start I wanted to ask you this uh so MIT 00:01:49.600 --> 00:01:52.270 align:start position:0% start I wanted to ask you this uh so MIT and<00:01:49.840> Sil<00:01:50.280> Valley<00:01:50.759> do<00:01:50.920> you<00:01:51.079> see<00:01:51.320> any<00:01:51.560> cultural 00:01:52.270 --> 00:01:52.280 align:start position:0% and Sil Valley do you see any cultural 00:01:52.280 --> 00:01:54.069 align:start position:0% and Sil Valley do you see any cultural differences<00:01:53.000> is<00:01:53.200> being<00:01:53.600> kind<00:01:53.719> of<00:01:53.880> a<00:01:53.960> little 00:01:54.069 --> 00:01:54.079 align:start position:0% differences is being kind of a little 00:01:54.079 --> 00:01:56.029 align:start position:0% differences is being kind of a little bit<00:01:54.240> removed<00:01:55.000> from<00:01:55.159> Sil<00:01:55.399> and<00:01:55.560> well<00:01:55.799> does<00:01:55.920> it 00:01:56.029 --> 00:01:56.039 align:start position:0% bit removed from Sil and well does it 00:01:56.039 --> 00:01:58.310 align:start position:0% bit removed from Sil and well does it help<00:01:56.240> you<00:01:56.399> to<00:01:56.600> focus<00:01:57.360> on 00:01:58.310 --> 00:01:58.320 align:start position:0% help you to focus on 00:01:58.320 --> 00:02:00.950 align:start position:0% help you to focus on science<00:01:59.320> I<00:01:59.439> think<00:01:59.719> uh<00:02:00.200> I<00:02:00.320> think<00:02:00.560> yeah<00:02:00.719> there's 00:02:00.950 --> 00:02:00.960 align:start position:0% science I think uh I think yeah there's 00:02:00.960 --> 00:02:04.069 align:start position:0% science I think uh I think yeah there's definitely<00:02:01.759> a<00:02:01.880> lot<00:02:02.159> less<00:02:02.600> talk<00:02:02.880> about<00:02:03.200> tech<00:02:03.960> uh 00:02:04.069 --> 00:02:04.079 align:start position:0% definitely a lot less talk about tech uh 00:02:04.079 --> 00:02:07.069 align:start position:0% definitely a lot less talk about tech uh over<00:02:04.399> there<00:02:04.840> and<00:02:05.280> uh<00:02:05.479> I<00:02:05.560> think<00:02:06.280> you<00:02:06.399> know<00:02:06.799> I<00:02:06.960> I 00:02:07.069 --> 00:02:07.079 align:start position:0% over there and uh I think you know I I 00:02:07.079 --> 00:02:08.589 align:start position:0% over there and uh I think you know I I haven't<00:02:07.280> been<00:02:07.600> over<00:02:07.799> on<00:02:07.960> the<00:02:08.039> west<00:02:08.200> coast<00:02:08.440> too 00:02:08.589 --> 00:02:08.599 align:start position:0% haven't been over on the west coast too 00:02:08.599 --> 00:02:12.229 align:start position:0% haven't been over on the west coast too much<00:02:09.360> but<00:02:10.039> uh<00:02:10.520> I<00:02:10.640> know<00:02:10.920> from<00:02:11.239> our<00:02:11.560> perspective 00:02:12.229 --> 00:02:12.239 align:start position:0% much but uh I know from our perspective 00:02:12.239 --> 00:02:14.670 align:start position:0% much but uh I know from our perspective it's<00:02:12.440> made<00:02:12.599> it<00:02:12.760> pretty<00:02:13.040> easy<00:02:13.400> to<00:02:14.319> uh<00:02:14.400> you<00:02:14.519> know 00:02:14.670 --> 00:02:14.680 align:start position:0% it's made it pretty easy to uh you know 00:02:14.680 --> 00:02:16.070 align:start position:0% it's made it pretty easy to uh you know hire<00:02:14.920> and<00:02:15.080> find<00:02:15.360> people<00:02:15.560> that<00:02:15.680> are<00:02:15.879> very 00:02:16.070 --> 00:02:16.080 align:start position:0% hire and find people that are very 00:02:16.080 --> 00:02:18.350 align:start position:0% hire and find people that are very grounded<00:02:16.599> in<00:02:17.519> uh<00:02:17.720> high<00:02:17.920> performance 00:02:18.350 --> 00:02:18.360 align:start position:0% grounded in uh high performance 00:02:18.360 --> 00:02:21.070 align:start position:0% grounded in uh high performance Computing<00:02:19.040> and<00:02:19.720> uh<00:02:19.800> you<00:02:19.920> know<00:02:20.080> math<00:02:20.440> software 00:02:21.070 --> 00:02:21.080 align:start position:0% Computing and uh you know math software 00:02:21.080 --> 00:02:22.910 align:start position:0% Computing and uh you know math software libraries<00:02:22.080> uh<00:02:22.200> which<00:02:22.319> has<00:02:22.440> been<00:02:22.599> really<00:02:22.800> good 00:02:22.910 --> 00:02:22.920 align:start position:0% libraries uh which has been really good 00:02:22.920 --> 00:02:24.830 align:start position:0% libraries uh which has been really good for<00:02:23.120> us<00:02:23.319> as<00:02:23.480> we're<00:02:23.680> really<00:02:23.920> focused<00:02:24.239> on 00:02:24.830 --> 00:02:24.840 align:start position:0% for us as we're really focused on 00:02:24.840 --> 00:02:27.229 align:start position:0% for us as we're really focused on performance<00:02:25.840> um<00:02:26.200> you<00:02:26.319> know<00:02:26.840> these<00:02:27.080> these 00:02:27.229 --> 00:02:27.239 align:start position:0% performance um you know these these 00:02:27.239 --> 00:02:28.630 align:start position:0% performance um you know these these models<00:02:27.560> are<00:02:27.760> are<00:02:27.920> mostly<00:02:28.280> Matrix 00:02:28.630 --> 00:02:28.640 align:start position:0% models are are mostly Matrix 00:02:28.640 --> 00:02:30.430 align:start position:0% models are are mostly Matrix multiplications<00:02:29.519> so 00:02:30.430 --> 00:02:30.440 align:start position:0% multiplications so 00:02:30.440 --> 00:02:32.270 align:start position:0% multiplications so uh<00:02:30.640> this<00:02:30.720> is<00:02:30.920> really<00:02:31.120> our<00:02:31.400> Focus<00:02:31.800> as<00:02:31.920> a 00:02:32.270 --> 00:02:32.280 align:start position:0% uh this is really our Focus as a 00:02:32.280 --> 00:02:34.470 align:start position:0% uh this is really our Focus as a inference<00:02:32.720> performance<00:02:33.160> company<00:02:34.040> is<00:02:34.120> it<00:02:34.319> okay 00:02:34.470 --> 00:02:34.480 align:start position:0% inference performance company is it okay 00:02:34.480 --> 00:02:37.390 align:start position:0% inference performance company is it okay if<00:02:34.640> I<00:02:34.959> reveal<00:02:35.440> the<00:02:35.840> secret<00:02:36.160> of<00:02:36.319> AI<00:02:36.599> to<00:02:36.800> a<00:02:37.319> go 00:02:37.390 --> 00:02:37.400 align:start position:0% if I reveal the secret of AI to a go 00:02:37.400 --> 00:02:40.350 align:start position:0% if I reveal the secret of AI to a go ahead<00:02:37.720> so<00:02:38.080> oi<00:02:39.080> is<00:02:39.440> just<00:02:39.680> Matrix 00:02:40.350 --> 00:02:40.360 align:start position:0% ahead so oi is just Matrix 00:02:40.360 --> 00:02:43.030 align:start position:0% ahead so oi is just Matrix multiplications<00:02:41.360> all<00:02:41.560> the<00:02:41.680> way<00:02:41.840> down<00:02:42.560> and 00:02:43.030 --> 00:02:43.040 align:start position:0% multiplications all the way down and 00:02:43.040 --> 00:02:46.110 align:start position:0% multiplications all the way down and matrices<00:02:43.840> need<00:02:44.120> numbers<00:02:44.959> so<00:02:45.159> first<00:02:45.360> we<00:02:45.519> can 00:02:46.110 --> 00:02:46.120 align:start position:0% matrices need numbers so first we can 00:02:46.120 --> 00:02:48.589 align:start position:0% matrices need numbers so first we can basically<00:02:46.480> convert<00:02:47.120> reality<00:02:47.480> to<00:02:47.640> numbers<00:02:48.440> put 00:02:48.589 --> 00:02:48.599 align:start position:0% basically convert reality to numbers put 00:02:48.599 --> 00:02:50.350 align:start position:0% basically convert reality to numbers put them<00:02:48.720> in<00:02:48.879> matrices<00:02:49.360> and<00:02:49.519> then<00:02:49.680> we<00:02:49.879> run<00:02:50.080> a<00:02:50.200> bunch 00:02:50.350 --> 00:02:50.360 align:start position:0% them in matrices and then we run a bunch 00:02:50.360 --> 00:02:52.229 align:start position:0% them in matrices and then we run a bunch of<00:02:50.480> multiplications<00:02:51.400> yeah<00:02:51.680> through<00:02:51.879> tensor 00:02:52.229 --> 00:02:52.239 align:start position:0% of multiplications yeah through tensor 00:02:52.239 --> 00:02:54.470 align:start position:0% of multiplications yeah through tensor course<00:02:52.599> through<00:02:52.800> tens<00:02:53.200> right<00:02:53.360> so<00:02:53.480> this<00:02:53.599> is<00:02:53.720> all 00:02:54.470 --> 00:02:54.480 align:start position:0% course through tens right so this is all 00:02:54.480 --> 00:02:56.869 align:start position:0% course through tens right so this is all you<00:02:54.599> need<00:02:54.760> to<00:02:54.879> know<00:02:55.000> about<00:02:55.319> yeah<00:02:56.040> uh<00:02:56.200> and<00:02:56.720> but 00:02:56.869 --> 00:02:56.879 align:start position:0% you need to know about yeah uh and but 00:02:56.879 --> 00:02:58.910 align:start position:0% you need to know about yeah uh and but this<00:02:57.000> is<00:02:57.120> a<00:02:57.319> complex<00:02:57.680> engineering<00:02:58.319> problem 00:02:58.910 --> 00:02:58.920 align:start position:0% this is a complex engineering problem 00:02:58.920 --> 00:03:00.630 align:start position:0% this is a complex engineering problem right<00:02:59.159> so<00:02:59.319> you<00:02:59.480> mentioned 00:03:00.630 --> 00:03:00.640 align:start position:0% right so you mentioned 00:03:00.640 --> 00:03:02.149 align:start position:0% right so you mentioned uh<00:03:00.760> some<00:03:00.920> of<00:03:01.040> the<00:03:01.159> challenges<00:03:01.720> what<00:03:01.840> are<00:03:02.000> the 00:03:02.149 --> 00:03:02.159 align:start position:0% uh some of the challenges what are the 00:03:02.159 --> 00:03:05.509 align:start position:0% uh some of the challenges what are the kind<00:03:02.280> of<00:03:02.440> hardest<00:03:02.879> problems<00:03:03.879> VM 00:03:05.509 --> 00:03:05.519 align:start position:0% kind of hardest problems VM 00:03:05.519 --> 00:03:07.430 align:start position:0% kind of hardest problems VM solves<00:03:06.519> actually<00:03:06.760> one<00:03:06.840> of<00:03:06.959> the<00:03:07.080> hardest 00:03:07.430 --> 00:03:07.440 align:start position:0% solves actually one of the hardest 00:03:07.440 --> 00:03:11.110 align:start position:0% solves actually one of the hardest problems<00:03:07.799> we<00:03:08.080> have<00:03:08.640> is<00:03:09.120> I<00:03:09.200> would<00:03:09.400> say<00:03:09.959> CPU 00:03:11.110 --> 00:03:11.120 align:start position:0% problems we have is I would say CPU 00:03:11.120 --> 00:03:13.589 align:start position:0% problems we have is I would say CPU overhead<00:03:12.120> uh<00:03:12.440> particularly<00:03:12.959> because<00:03:13.319> a<00:03:13.440> lot 00:03:13.589 --> 00:03:13.599 align:start position:0% overhead uh particularly because a lot 00:03:13.599 --> 00:03:15.670 align:start position:0% overhead uh particularly because a lot of<00:03:13.760> VM<00:03:14.280> is<00:03:14.560> surprisingly<00:03:15.080> enough<00:03:15.280> written<00:03:15.519> in 00:03:15.670 --> 00:03:15.680 align:start position:0% of VM is surprisingly enough written in 00:03:15.680 --> 00:03:18.190 align:start position:0% of VM is surprisingly enough written in Python<00:03:16.280> M<00:03:16.799> even<00:03:17.080> though<00:03:17.440> it's<00:03:17.760> you<00:03:17.879> know<00:03:18.040> a 00:03:18.190 --> 00:03:18.200 align:start position:0% Python M even though it's you know a 00:03:18.200 --> 00:03:21.430 align:start position:0% Python M even though it's you know a very<00:03:18.640> performant<00:03:19.400> Library<00:03:20.400> we<00:03:20.720> choose<00:03:21.159> to 00:03:21.430 --> 00:03:21.440 align:start position:0% very performant Library we choose to 00:03:21.440 --> 00:03:24.509 align:start position:0% very performant Library we choose to continue<00:03:22.400> writing<00:03:23.040> the<00:03:23.200> majority<00:03:23.640> of<00:03:23.879> it<00:03:24.400> uh 00:03:24.509 --> 00:03:24.519 align:start position:0% continue writing the majority of it uh 00:03:24.519 --> 00:03:26.910 align:start position:0% continue writing the majority of it uh the<00:03:24.680> vast<00:03:24.920> majority<00:03:25.280> of<00:03:25.400> it<00:03:25.799> in<00:03:26.080> Python<00:03:26.599> and<00:03:26.720> in 00:03:26.910 --> 00:03:26.920 align:start position:0% the vast majority of it in Python and in 00:03:26.920 --> 00:03:29.190 align:start position:0% the vast majority of it in Python and in pytorch<00:03:27.760> to<00:03:27.959> make<00:03:28.080> it<00:03:28.400> approachable<00:03:29.000> and 00:03:29.190 --> 00:03:29.200 align:start position:0% pytorch to make it approachable and 00:03:29.200 --> 00:03:31.869 align:start position:0% pytorch to make it approachable and accessible<00:03:29.799> for<00:03:29.959> new<00:03:30.319> people<00:03:31.319> uh<00:03:31.439> of<00:03:31.599> course 00:03:31.869 --> 00:03:31.879 align:start position:0% accessible for new people uh of course 00:03:31.879 --> 00:03:34.229 align:start position:0% accessible for new people uh of course this<00:03:32.000> has<00:03:32.239> required<00:03:32.720> us<00:03:33.000> to<00:03:33.760> you<00:03:33.879> know<00:03:34.040> take 00:03:34.229 --> 00:03:34.239 align:start position:0% this has required us to you know take 00:03:34.239 --> 00:03:37.229 align:start position:0% this has required us to you know take advantage<00:03:34.799> of<00:03:35.799> uh<00:03:35.920> sort<00:03:36.080> of<00:03:36.280> Cutting<00:03:36.599> Edge<00:03:37.120> uh 00:03:37.229 --> 00:03:37.239 align:start position:0% advantage of uh sort of Cutting Edge uh 00:03:37.239 --> 00:03:38.869 align:start position:0% advantage of uh sort of Cutting Edge uh technology<00:03:37.840> you<00:03:37.959> know<00:03:38.239> probably<00:03:38.519> from<00:03:38.720> from 00:03:38.869 --> 00:03:38.879 align:start position:0% technology you know probably from from 00:03:38.879 --> 00:03:40.190 align:start position:0% technology you know probably from from pytorch<00:03:39.319> you're<00:03:39.439> familiar<00:03:39.760> with<00:03:39.879> torch 00:03:40.190 --> 00:03:40.200 align:start position:0% pytorch you're familiar with torch 00:03:40.200 --> 00:03:42.990 align:start position:0% pytorch you're familiar with torch compile<00:03:40.959> this<00:03:41.080> has<00:03:41.200> been<00:03:41.319> a<00:03:41.680> great<00:03:42.680> uh<00:03:42.879> you 00:03:42.990 --> 00:03:43.000 align:start position:0% compile this has been a great uh you 00:03:43.000 --> 00:03:45.990 align:start position:0% compile this has been a great uh you know<00:03:43.560> uh<00:03:44.200> advancement<00:03:44.840> that's<00:03:45.080> allowed<00:03:45.360> us<00:03:45.519> to 00:03:45.990 --> 00:03:46.000 align:start position:0% know uh advancement that's allowed us to 00:03:46.000 --> 00:03:48.470 align:start position:0% know uh advancement that's allowed us to sort<00:03:46.200> of<00:03:46.400> get<00:03:46.599> back<00:03:46.760> into<00:03:47.200> traditional<00:03:47.720> graph 00:03:48.470 --> 00:03:48.480 align:start position:0% sort of get back into traditional graph 00:03:48.480 --> 00:03:51.190 align:start position:0% sort of get back into traditional graph compilation<00:03:49.480> uh<00:03:49.640> and<00:03:49.840> and<00:03:50.000> try<00:03:50.159> to<00:03:50.560> you<00:03:50.680> know 00:03:51.190 --> 00:03:51.200 align:start position:0% compilation uh and and try to you know 00:03:51.200 --> 00:03:54.550 align:start position:0% compilation uh and and try to you know uh<00:03:51.360> get<00:03:51.519> some<00:03:51.760> larger<00:03:52.640> graphs<00:03:53.640> uh<00:03:53.879> to<00:03:54.239> to 00:03:54.550 --> 00:03:54.560 align:start position:0% uh get some larger graphs uh to to 00:03:54.560 --> 00:03:56.030 align:start position:0% uh get some larger graphs uh to to prevent<00:03:54.879> the<00:03:55.000> amount<00:03:55.159> of<00:03:55.280> CPU<00:03:55.640> GPU 00:03:56.030 --> 00:03:56.040 align:start position:0% prevent the amount of CPU GPU 00:03:56.040 --> 00:03:57.830 align:start position:0% prevent the amount of CPU GPU communication<00:03:56.519> we<00:03:56.640> have<00:03:56.760> to<00:03:56.879> make<00:03:57.480> but<00:03:57.640> also 00:03:57.830 --> 00:03:57.840 align:start position:0% communication we have to make but also 00:03:57.840 --> 00:03:59.869 align:start position:0% communication we have to make but also it's<00:03:58.040> required<00:03:58.400> us<00:03:58.519> to<00:03:58.640> make<00:03:58.840> sort<00:03:59.000> of<00:03:59.519> uh<00:03:59.799> you 00:03:59.869 --> 00:03:59.879 align:start position:0% it's required us to make sort of uh you 00:03:59.879 --> 00:04:01.990 align:start position:0% it's required us to make sort of uh you know<00:04:00.000> relatively<00:04:00.560> complex 00:04:01.990 --> 00:04:02.000 align:start position:0% know relatively complex 00:04:02.000 --> 00:04:03.710 align:start position:0% know relatively complex multi-process<00:04:03.000> uh<00:04:03.079> you<00:04:03.159> know<00:04:03.360> python 00:04:03.710 --> 00:04:03.720 align:start position:0% multi-process uh you know python 00:04:03.720 --> 00:04:05.309 align:start position:0% multi-process uh you know python architecture<00:04:04.400> with<00:04:04.560> interprocess 00:04:05.309 --> 00:04:05.319 align:start position:0% architecture with interprocess 00:04:05.319 --> 00:04:07.390 align:start position:0% architecture with interprocess communication<00:04:06.319> proper 00:04:07.390 --> 00:04:07.400 align:start position:0% communication proper 00:04:07.400 --> 00:04:10.869 align:start position:0% communication proper asynchronicity<00:04:08.400> um<00:04:08.799> to<00:04:09.000> keep<00:04:09.360> these<00:04:09.840> you<00:04:09.959> know 00:04:10.869 --> 00:04:10.879 align:start position:0% asynchronicity um to keep these you know 00:04:10.879 --> 00:04:15.030 align:start position:0% asynchronicity um to keep these you know uh<00:04:11.720> big<00:04:12.360> very<00:04:12.720> fast<00:04:13.000> gpus<00:04:13.879> fed<00:04:14.640> uh<00:04:14.720> you<00:04:14.840> know 00:04:15.030 --> 00:04:15.040 align:start position:0% uh big very fast gpus fed uh you know 00:04:15.040 --> 00:04:16.509 align:start position:0% uh big very fast gpus fed uh you know this<00:04:15.159> is<00:04:15.319> essentially<00:04:15.680> the<00:04:15.879> goal<00:04:16.239> we<00:04:16.400> just 00:04:16.509 --> 00:04:16.519 align:start position:0% this is essentially the goal we just 00:04:16.519 --> 00:04:18.189 align:start position:0% this is essentially the goal we just want<00:04:16.639> to<00:04:16.759> keep<00:04:17.000> these<00:04:17.160> things<00:04:17.440> at<00:04:17.759> 100% 00:04:18.189 --> 00:04:18.199 align:start position:0% want to keep these things at 100% 00:04:18.199 --> 00:04:20.069 align:start position:0% want to keep these things at 100% utilization<00:04:19.079> and<00:04:19.239> anything<00:04:19.519> we're<00:04:19.720> not<00:04:19.840> doing 00:04:20.069 --> 00:04:20.079 align:start position:0% utilization and anything we're not doing 00:04:20.079 --> 00:04:22.749 align:start position:0% utilization and anything we're not doing on<00:04:20.199> the<00:04:20.320> GPU<00:04:21.199> needs<00:04:21.440> to<00:04:21.639> happen<00:04:22.120> while<00:04:22.280> the<00:04:22.400> GPU 00:04:22.749 --> 00:04:22.759 align:start position:0% on the GPU needs to happen while the GPU 00:04:22.759 --> 00:04:25.550 align:start position:0% on the GPU needs to happen while the GPU is<00:04:22.840> running<00:04:23.360> not<00:04:23.520> while<00:04:23.680> the<00:04:23.840> GP<00:04:24.199> is<00:04:24.560> waiting 00:04:25.550 --> 00:04:25.560 align:start position:0% is running not while the GP is waiting 00:04:25.560 --> 00:04:27.230 align:start position:0% is running not while the GP is waiting uh<00:04:25.720> so<00:04:26.400> you<00:04:26.520> know<00:04:26.680> these<00:04:26.800> are<00:04:26.919> some<00:04:27.040> of<00:04:27.120> the 00:04:27.230 --> 00:04:27.240 align:start position:0% uh so you know these are some of the 00:04:27.240 --> 00:04:29.790 align:start position:0% uh so you know these are some of the biggest<00:04:27.560> concerns<00:04:28.560> uh<00:04:28.880> that<00:04:29.160> that<00:04:29.360> that<00:04:29.479> we 00:04:29.790 --> 00:04:29.800 align:start position:0% biggest concerns uh that that that we 00:04:29.800 --> 00:04:31.990 align:start position:0% biggest concerns uh that that that we have<00:04:29.919> to<00:04:30.080> do<00:04:30.400> to<00:04:30.560> keep<00:04:30.720> the<00:04:30.960> project<00:04:31.759> uh 00:04:31.990 --> 00:04:32.000 align:start position:0% have to do to keep the project uh 00:04:32.000 --> 00:04:34.150 align:start position:0% have to do to keep the project uh successful<00:04:32.840> as<00:04:33.000> all<00:04:33.160> of<00:04:33.280> the<00:04:33.400> new<00:04:33.680> inference 00:04:34.150 --> 00:04:34.160 align:start position:0% successful as all of the new inference 00:04:34.160 --> 00:04:36.029 align:start position:0% successful as all of the new inference optimizations<00:04:35.039> and<00:04:35.199> methods<00:04:35.600> come<00:04:35.759> in 00:04:36.029 --> 00:04:36.039 align:start position:0% optimizations and methods come in 00:04:36.039 --> 00:04:39.029 align:start position:0% optimizations and methods come in through<00:04:36.400> research<00:04:36.840> and<00:04:37.120> through<00:04:38.120> uh<00:04:38.680> uh<00:04:38.880> you 00:04:39.029 --> 00:04:39.039 align:start position:0% through research and through uh uh you 00:04:39.039 --> 00:04:43.350 align:start position:0% through research and through uh uh you know<00:04:39.639> Enterprise<00:04:40.639> and<00:04:41.000> uh<00:04:41.880> uh<00:04:42.880> I<00:04:43.000> mentioned 00:04:43.350 --> 00:04:43.360 align:start position:0% know Enterprise and uh uh I mentioned 00:04:43.360 --> 00:04:45.390 align:start position:0% know Enterprise and uh uh I mentioned that<00:04:43.520> your<00:04:43.680> title<00:04:43.919> is<00:04:44.080> engineering<00:04:44.600> lead<00:04:45.080> and 00:04:45.390 --> 00:04:45.400 align:start position:0% that your title is engineering lead and 00:04:45.400 --> 00:04:48.590 align:start position:0% that your title is engineering lead and so<00:04:46.400> uh<00:04:47.039> building<00:04:47.440> V<00:04:47.919> involves<00:04:48.280> a<00:04:48.360> lot<00:04:48.520> of 00:04:48.590 --> 00:04:48.600 align:start position:0% so uh building V involves a lot of 00:04:48.600 --> 00:04:50.150 align:start position:0% so uh building V involves a lot of engineering<00:04:49.199> can<00:04:49.360> you<00:04:49.520> talk<00:04:49.680> a<00:04:49.759> little<00:04:49.960> bit 00:04:50.150 --> 00:04:50.160 align:start position:0% engineering can you talk a little bit 00:04:50.160 --> 00:04:53.670 align:start position:0% engineering can you talk a little bit about<00:04:50.520> how<00:04:51.080> this<00:04:52.080> uh<00:04:52.720> discipline<00:04:53.440> of 00:04:53.670 --> 00:04:53.680 align:start position:0% about how this uh discipline of 00:04:53.680 --> 00:04:56.469 align:start position:0% about how this uh discipline of engineering<00:04:54.320> is<00:04:54.639> software<00:04:55.120> engineering<00:04:56.120> uh 00:04:56.469 --> 00:04:56.479 align:start position:0% engineering is software engineering uh 00:04:56.479 --> 00:05:01.110 align:start position:0% engineering is software engineering uh knowledge<00:04:57.280> is<00:04:57.520> helping<00:04:57.919> you<00:04:58.280> scale<00:04:59.280> V 00:05:01.110 --> 00:05:01.120 align:start position:0% knowledge is helping you scale V 00:05:01.120 --> 00:05:04.189 align:start position:0% knowledge is helping you scale V yeah<00:05:01.479> I<00:05:01.560> mean<00:05:02.280> uh<00:05:02.720> I<00:05:02.880> have<00:05:03.199> maybe<00:05:03.560> kind<00:05:03.680> of<00:05:03.800> a 00:05:04.189 --> 00:05:04.199 align:start position:0% yeah I mean uh I have maybe kind of a 00:05:04.199 --> 00:05:06.629 align:start position:0% yeah I mean uh I have maybe kind of a vague<00:05:04.560> title<00:05:05.240> uh<00:05:05.360> because<00:05:05.720> I<00:05:05.960> I<00:05:06.039> work<00:05:06.240> on<00:05:06.440> a<00:05:06.520> lot 00:05:06.629 --> 00:05:06.639 align:start position:0% vague title uh because I I work on a lot 00:05:06.639 --> 00:05:09.469 align:start position:0% vague title uh because I I work on a lot of<00:05:06.800> vague<00:05:07.080> things<00:05:07.840> I<00:05:07.919> think<00:05:08.400> uh<00:05:08.720> to<00:05:08.960> to<00:05:09.120> to<00:05:09.280> be 00:05:09.469 --> 00:05:09.479 align:start position:0% of vague things I think uh to to to be 00:05:09.479 --> 00:05:13.469 align:start position:0% of vague things I think uh to to to be in<00:05:09.680> the<00:05:10.000> in<00:05:10.400> the<00:05:11.400> to<00:05:11.560> be<00:05:11.880> really<00:05:12.240> successful<00:05:13.240> in 00:05:13.469 --> 00:05:13.479 align:start position:0% in the in the to be really successful in 00:05:13.479 --> 00:05:16.870 align:start position:0% in the in the to be really successful in the<00:05:13.680> field<00:05:14.039> of<00:05:14.240> performance<00:05:15.039> engineering<00:05:16.039> uh 00:05:16.870 --> 00:05:16.880 align:start position:0% the field of performance engineering uh 00:05:16.880 --> 00:05:19.749 align:start position:0% the field of performance engineering uh I<00:05:17.120> I<00:05:17.240> believe<00:05:17.560> you<00:05:17.680> have<00:05:17.840> to<00:05:18.000> go<00:05:18.280> down<00:05:18.560> to<00:05:18.919> the 00:05:19.749 --> 00:05:19.759 align:start position:0% I I believe you have to go down to the 00:05:19.759 --> 00:05:22.590 align:start position:0% I I believe you have to go down to the assembly<00:05:20.240> level<00:05:20.639> the<00:05:20.759> kernel<00:05:21.240> level<00:05:22.240> uh<00:05:22.400> when 00:05:22.590 --> 00:05:22.600 align:start position:0% assembly level the kernel level uh when 00:05:22.600 --> 00:05:24.469 align:start position:0% assembly level the kernel level uh when you<00:05:22.960> really<00:05:23.240> need<00:05:23.440> to<00:05:23.680> make<00:05:23.800> a<00:05:23.960> really<00:05:24.199> really 00:05:24.469 --> 00:05:24.479 align:start position:0% you really need to make a really really 00:05:24.479 --> 00:05:26.950 align:start position:0% you really need to make a really really tight<00:05:24.759> Loop<00:05:25.759> uh<00:05:25.880> but<00:05:26.039> then<00:05:26.240> also<00:05:26.520> go<00:05:26.720> all<00:05:26.880> the 00:05:26.950 --> 00:05:26.960 align:start position:0% tight Loop uh but then also go all the 00:05:26.960 --> 00:05:29.150 align:start position:0% tight Loop uh but then also go all the way<00:05:27.120> up<00:05:27.280> to<00:05:27.520> the<00:05:27.919> the<00:05:28.120> python<00:05:28.720> the<00:05:28.840> machine 00:05:29.150 --> 00:05:29.160 align:start position:0% way up to the the python the machine 00:05:29.160 --> 00:05:33.830 align:start position:0% way up to the the python the machine learning<00:05:30.120> level<00:05:30.919> uh<00:05:31.080> in<00:05:31.199> order<00:05:31.560> to<00:05:32.360> you<00:05:32.840> know 00:05:33.830 --> 00:05:33.840 align:start position:0% learning level uh in order to you know 00:05:33.840 --> 00:05:35.270 align:start position:0% learning level uh in order to you know understand<00:05:34.360> and<00:05:34.520> connect<00:05:34.800> to<00:05:35.000> the 00:05:35.270 --> 00:05:35.280 align:start position:0% understand and connect to the 00:05:35.280 --> 00:05:38.110 align:start position:0% understand and connect to the applications<00:05:36.280> that<00:05:36.479> these<00:05:36.680> engines<00:05:37.240> are<00:05:37.880> are 00:05:38.110 --> 00:05:38.120 align:start position:0% applications that these engines are are 00:05:38.120 --> 00:05:40.390 align:start position:0% applications that these engines are are ending<00:05:38.400> up<00:05:38.600> just<00:05:38.720> being<00:05:38.960> used<00:05:39.240> for<00:05:40.160> uh<00:05:40.280> you 00:05:40.390 --> 00:05:40.400 align:start position:0% ending up just being used for uh you 00:05:40.400 --> 00:05:42.270 align:start position:0% ending up just being used for uh you know<00:05:40.560> ultimately<00:05:41.440> the<00:05:41.600> worst<00:05:41.880> thing<00:05:42.039> you<00:05:42.160> can 00:05:42.270 --> 00:05:42.280 align:start position:0% know ultimately the worst thing you can 00:05:42.280 --> 00:05:44.710 align:start position:0% know ultimately the worst thing you can do<00:05:42.440> in<00:05:42.600> performance<00:05:43.000> engineering<00:05:43.960> is<00:05:44.360> get<00:05:44.560> the 00:05:44.710 --> 00:05:44.720 align:start position:0% do in performance engineering is get the 00:05:44.720 --> 00:05:47.350 align:start position:0% do in performance engineering is get the wrong<00:05:45.039> Benchmark<00:05:46.039> it's<00:05:46.240> very<00:05:46.479> easy<00:05:46.720> to<00:05:46.919> delude 00:05:47.350 --> 00:05:47.360 align:start position:0% wrong Benchmark it's very easy to delude 00:05:47.360 --> 00:05:49.510 align:start position:0% wrong Benchmark it's very easy to delude yourself<00:05:47.960> that<00:05:48.080> you<00:05:48.199> are<00:05:48.440> performant<00:05:49.199> if<00:05:49.319> you 00:05:49.510 --> 00:05:49.520 align:start position:0% yourself that you are performant if you 00:05:49.520 --> 00:05:50.990 align:start position:0% yourself that you are performant if you create<00:05:49.759> some<00:05:49.960> micro<00:05:50.240> Benchmark<00:05:50.759> and<00:05:50.840> you're 00:05:50.990 --> 00:05:51.000 align:start position:0% create some micro Benchmark and you're 00:05:51.000 --> 00:05:53.029 align:start position:0% create some micro Benchmark and you're like<00:05:51.400> oh<00:05:51.560> I<00:05:51.639> did<00:05:51.800> a<00:05:51.919> great<00:05:52.120> job<00:05:52.319> I'm<00:05:52.440> at<00:05:52.639> 100% 00:05:53.029 --> 00:05:53.039 align:start position:0% like oh I did a great job I'm at 100% 00:05:53.039 --> 00:05:54.749 align:start position:0% like oh I did a great job I'm at 100% utilization<00:05:53.880> but<00:05:54.000> then<00:05:54.120> you<00:05:54.240> get<00:05:54.360> some<00:05:54.520> real 00:05:54.749 --> 00:05:54.759 align:start position:0% utilization but then you get some real 00:05:54.759 --> 00:05:58.070 align:start position:0% utilization but then you get some real world<00:05:55.080> data<00:05:55.919> and<00:05:56.280> it's<00:05:56.759> uh<00:05:57.560> it's<00:05:57.800> very 00:05:58.070 --> 00:05:58.080 align:start position:0% world data and it's uh it's very 00:05:58.080 --> 00:06:00.270 align:start position:0% world data and it's uh it's very heterogeneous<00:05:58.880> you<00:05:59.000> have<00:05:59.120> some<00:05:59.319> very 00:06:00.270 --> 00:06:00.280 align:start position:0% heterogeneous you have some very 00:06:00.280 --> 00:06:02.390 align:start position:0% heterogeneous you have some very uh<00:06:00.400> you<00:06:00.520> know<00:06:00.800> long<00:06:01.160> prompts<00:06:01.520> coming<00:06:01.759> in<00:06:02.000> or<00:06:02.199> or 00:06:02.390 --> 00:06:02.400 align:start position:0% uh you know long prompts coming in or or 00:06:02.400 --> 00:06:04.710 align:start position:0% uh you know long prompts coming in or or lots<00:06:02.639> of<00:06:02.800> images<00:06:03.240> all<00:06:03.400> at<00:06:03.600> once<00:06:03.960> and<00:06:04.479> uh<00:06:04.600> you 00:06:04.710 --> 00:06:04.720 align:start position:0% lots of images all at once and uh you 00:06:04.720 --> 00:06:07.589 align:start position:0% lots of images all at once and uh you have<00:06:04.919> maybe<00:06:05.560> uh<00:06:06.280> you<00:06:06.400> know<00:06:06.840> a<00:06:06.960> bunch<00:06:07.120> of<00:06:07.240> small 00:06:07.589 --> 00:06:07.599 align:start position:0% have maybe uh you know a bunch of small 00:06:07.599 --> 00:06:10.469 align:start position:0% have maybe uh you know a bunch of small ones<00:06:07.960> piling<00:06:08.360> up<00:06:09.160> um<00:06:09.319> you<00:06:09.440> really<00:06:09.680> need<00:06:09.840> to<00:06:10.080> be 00:06:10.469 --> 00:06:10.479 align:start position:0% ones piling up um you really need to be 00:06:10.479 --> 00:06:15.150 align:start position:0% ones piling up um you really need to be sort<00:06:10.680> of<00:06:11.599> uh<00:06:12.599> antagonistic<00:06:13.319> to<00:06:13.639> yourself<00:06:14.639> and 00:06:15.150 --> 00:06:15.160 align:start position:0% sort of uh antagonistic to yourself and 00:06:15.160 --> 00:06:19.070 align:start position:0% sort of uh antagonistic to yourself and uh<00:06:15.280> play<00:06:15.599> Devil's<00:06:15.960> Advocate<00:06:16.800> to<00:06:17.800> uh<00:06:18.319> you<00:06:18.440> know 00:06:19.070 --> 00:06:19.080 align:start position:0% uh play Devil's Advocate to uh you know 00:06:19.080 --> 00:06:20.950 align:start position:0% uh play Devil's Advocate to uh you know understand<00:06:19.280> all<00:06:19.400> of<00:06:19.520> the<00:06:19.599> workloads<00:06:20.599> uh<00:06:20.759> and 00:06:20.950 --> 00:06:20.960 align:start position:0% understand all of the workloads uh and 00:06:20.960 --> 00:06:22.870 align:start position:0% understand all of the workloads uh and and<00:06:21.160> how<00:06:21.360> you<00:06:21.440> can<00:06:21.639> optimize<00:06:22.080> for<00:06:22.280> them<00:06:22.759> and 00:06:22.870 --> 00:06:22.880 align:start position:0% and how you can optimize for them and 00:06:22.880 --> 00:06:24.270 align:start position:0% and how you can optimize for them and also<00:06:23.120> what<00:06:23.240> things<00:06:23.360> are<00:06:23.599> worth<00:06:23.840> optimizing 00:06:24.270 --> 00:06:24.280 align:start position:0% also what things are worth optimizing 00:06:24.280 --> 00:06:25.990 align:start position:0% also what things are worth optimizing for<00:06:24.520> and<00:06:24.720> not<00:06:25.240> because<00:06:25.440> you<00:06:25.560> still<00:06:25.720> want<00:06:25.840> to 00:06:25.990 --> 00:06:26.000 align:start position:0% for and not because you still want to 00:06:26.000 --> 00:06:28.950 align:start position:0% for and not because you still want to keep<00:06:26.160> the<00:06:26.599> software<00:06:27.599> uh<00:06:27.960> maintainable<00:06:28.800> and 00:06:28.950 --> 00:06:28.960 align:start position:0% keep the software uh maintainable and 00:06:28.960 --> 00:06:31.070 align:start position:0% keep the software uh maintainable and and<00:06:29.080> readable<00:06:29.720> and<00:06:29.880> accessible<00:06:30.639> for<00:06:30.840> new 00:06:31.070 --> 00:06:31.080 align:start position:0% and readable and accessible for new 00:06:31.080 --> 00:06:33.629 align:start position:0% and readable and accessible for new people<00:06:31.759> so<00:06:32.440> uh<00:06:32.759> you<00:06:32.880> know<00:06:33.080> it<00:06:33.280> it<00:06:33.400> really 00:06:33.629 --> 00:06:33.639 align:start position:0% people so uh you know it it really 00:06:33.639 --> 00:06:36.430 align:start position:0% people so uh you know it it really requires<00:06:34.400> I<00:06:34.520> think<00:06:35.280> not<00:06:35.479> being<00:06:35.720> afraid<00:06:36.039> to 00:06:36.430 --> 00:06:36.440 align:start position:0% requires I think not being afraid to 00:06:36.440 --> 00:06:38.309 align:start position:0% requires I think not being afraid to understand<00:06:36.599> how<00:06:36.800> computers<00:06:37.280> work<00:06:38.080> uh<00:06:38.160> but 00:06:38.309 --> 00:06:38.319 align:start position:0% understand how computers work uh but 00:06:38.319 --> 00:06:39.670 align:start position:0% understand how computers work uh but also<00:06:38.520> not<00:06:38.680> being<00:06:38.960> afraid<00:06:39.280> of<00:06:39.440> the 00:06:39.670 --> 00:06:39.680 align:start position:0% also not being afraid of the 00:06:39.680 --> 00:06:41.629 align:start position:0% also not being afraid of the applications<00:06:40.680> that<00:06:40.840> people<00:06:41.000> are<00:06:41.120> building<00:06:41.440> on 00:06:41.629 --> 00:06:41.639 align:start position:0% applications that people are building on 00:06:41.639 --> 00:06:43.469 align:start position:0% applications that people are building on top<00:06:41.759> of<00:06:41.880> the<00:06:42.039> computers<00:06:43.000> uh<00:06:43.120> which<00:06:43.240> I<00:06:43.360> think 00:06:43.469 --> 00:06:43.479 align:start position:0% top of the computers uh which I think 00:06:43.479 --> 00:06:45.150 align:start position:0% top of the computers uh which I think has<00:06:43.639> been<00:06:43.840> a<00:06:44.120> a<00:06:44.319> problem<00:06:44.560> since<00:06:44.800> computers 00:06:45.150 --> 00:06:45.160 align:start position:0% has been a a problem since computers 00:06:45.160 --> 00:06:48.029 align:start position:0% has been a a problem since computers have<00:06:45.360> come<00:06:46.080> around<00:06:47.080> makes<00:06:47.319> sense<00:06:47.759> this<00:06:47.880> is 00:06:48.029 --> 00:06:48.039 align:start position:0% have come around makes sense this is 00:06:48.039 --> 00:06:50.390 align:start position:0% have come around makes sense this is great<00:06:48.199> to<00:06:48.360> hear<00:06:48.720> and<00:06:49.240> I<00:06:49.400> come<00:06:49.599> from<00:06:49.960> you<00:06:50.039> know 00:06:50.390 --> 00:06:50.400 align:start position:0% great to hear and I come from you know 00:06:50.400 --> 00:06:53.189 align:start position:0% great to hear and I come from you know strong<00:06:50.880> typed<00:06:51.280> programming<00:06:51.840> background<00:06:52.840> and 00:06:53.189 --> 00:06:53.199 align:start position:0% strong typed programming background and 00:06:53.199 --> 00:06:55.029 align:start position:0% strong typed programming background and at<00:06:53.319> first<00:06:53.680> you<00:06:53.800> know<00:06:54.199> I<00:06:54.280> was<00:06:54.400> a<00:06:54.520> bit<00:06:54.639> skeptical 00:06:55.029 --> 00:06:55.039 align:start position:0% at first you know I was a bit skeptical 00:06:55.039 --> 00:06:57.150 align:start position:0% at first you know I was a bit skeptical about<00:06:55.280> python<00:06:55.919> succeeding<00:06:56.400> in<00:06:56.560> this<00:06:56.720> space 00:06:57.150 --> 00:06:57.160 align:start position:0% about python succeeding in this space 00:06:57.160 --> 00:06:58.950 align:start position:0% about python succeeding in this space but<00:06:57.560> it's<00:06:57.800> become<00:06:58.080> more<00:06:58.240> performant<00:06:58.680> and<00:06:58.840> what 00:06:58.950 --> 00:06:58.960 align:start position:0% but it's become more performant and what 00:06:58.960 --> 00:07:01.670 align:start position:0% but it's become more performant and what I<00:06:59.120> really<00:06:59.360> like<00:06:59.800> I<00:06:59.919> like<00:07:00.160> types<00:07:00.960> so<00:07:01.240> I<00:07:01.440> like 00:07:01.670 --> 00:07:01.680 align:start position:0% I really like I like types so I like 00:07:01.680 --> 00:07:03.950 align:start position:0% I really like I like types so I like things<00:07:01.919> like<00:07:02.120> pantic<00:07:02.599> which<00:07:02.800> gives<00:07:03.080> a<00:07:03.199> schema 00:07:03.950 --> 00:07:03.960 align:start position:0% things like pantic which gives a schema 00:07:03.960 --> 00:07:06.309 align:start position:0% things like pantic which gives a schema and<00:07:04.400> I'm<00:07:04.599> super<00:07:04.840> excited<00:07:05.160> about<00:07:05.560> moo<00:07:06.080> Chris 00:07:06.309 --> 00:07:06.319 align:start position:0% and I'm super excited about moo Chris 00:07:06.319 --> 00:07:08.110 align:start position:0% and I'm super excited about moo Chris lner<00:07:06.800> language<00:07:07.440> what<00:07:07.720> what<00:07:07.840> are<00:07:07.919> your 00:07:08.110 --> 00:07:08.120 align:start position:0% lner language what what are your 00:07:08.120 --> 00:07:10.909 align:start position:0% lner language what what are your thoughts<00:07:08.599> about<00:07:09.039> kind<00:07:09.160> of<00:07:09.840> type<00:07:10.160> Machinery 00:07:10.909 --> 00:07:10.919 align:start position:0% thoughts about kind of type Machinery 00:07:10.919 --> 00:07:12.710 align:start position:0% thoughts about kind of type Machinery help<00:07:11.120> it<00:07:11.360> by<00:07:11.639> be<00:07:11.800> more 00:07:12.710 --> 00:07:12.720 align:start position:0% help it by be more 00:07:12.720 --> 00:07:14.510 align:start position:0% help it by be more performant<00:07:13.720> I<00:07:13.879> that's<00:07:14.000> a<00:07:14.160> really<00:07:14.319> good 00:07:14.510 --> 00:07:14.520 align:start position:0% performant I that's a really good 00:07:14.520 --> 00:07:17.070 align:start position:0% performant I that's a really good question<00:07:14.960> I<00:07:15.000> mean<00:07:15.720> So<00:07:16.319> within<00:07:16.560> neurom<00:07:16.840> magic 00:07:17.070 --> 00:07:17.080 align:start position:0% question I mean So within neurom magic 00:07:17.080 --> 00:07:18.909 align:start position:0% question I mean So within neurom magic and<00:07:17.160> within<00:07:17.400> my<00:07:17.599> background<00:07:18.280> uh<00:07:18.479> we<00:07:18.599> were 00:07:18.909 --> 00:07:18.919 align:start position:0% and within my background uh we were 00:07:18.919 --> 00:07:21.390 align:start position:0% and within my background uh we were heavy<00:07:19.120> users<00:07:19.440> of<00:07:19.599> C++<00:07:20.440> heavy<00:07:20.720> template<00:07:21.080> meta 00:07:21.390 --> 00:07:21.400 align:start position:0% heavy users of C++ heavy template meta 00:07:21.400 --> 00:07:23.830 align:start position:0% heavy users of C++ heavy template meta programming<00:07:22.400> modern<00:07:22.720> features<00:07:23.120> you<00:07:23.199> know<00:07:23.400> we 00:07:23.830 --> 00:07:23.840 align:start position:0% programming modern features you know we 00:07:23.840 --> 00:07:25.869 align:start position:0% programming modern features you know we we<00:07:24.000> we<00:07:24.160> love<00:07:24.440> that<00:07:24.639> so<00:07:24.879> we<00:07:25.000> did<00:07:25.400> give<00:07:25.560> up<00:07:25.680> a<00:07:25.759> lot 00:07:25.869 --> 00:07:25.879 align:start position:0% we we love that so we did give up a lot 00:07:25.879 --> 00:07:28.430 align:start position:0% we we love that so we did give up a lot of<00:07:26.000> that<00:07:26.160> going<00:07:26.319> to<00:07:26.520> python<00:07:26.960> but<00:07:27.440> within<00:07:27.720> VM 00:07:28.430 --> 00:07:28.440 align:start position:0% of that going to python but within VM 00:07:28.440 --> 00:07:30.869 align:start position:0% of that going to python but within VM everything<00:07:28.840> is<00:07:29.039> typed<00:07:29.520> we<00:07:29.599> use<00:07:29.759> mypi<00:07:30.639> and<00:07:30.800> and 00:07:30.869 --> 00:07:30.879 align:start position:0% everything is typed we use mypi and and 00:07:30.879 --> 00:07:33.189 align:start position:0% everything is typed we use mypi and and linting<00:07:31.360> to<00:07:31.599> to<00:07:32.039> you<00:07:32.160> know<00:07:32.520> verify<00:07:32.960> this<00:07:33.080> on 00:07:33.189 --> 00:07:33.199 align:start position:0% linting to to you know verify this on 00:07:33.199 --> 00:07:35.950 align:start position:0% linting to to you know verify this on oursel<00:07:33.520> and<00:07:33.639> it<00:07:33.720> makes<00:07:33.879> it<00:07:34.080> much<00:07:34.319> easier<00:07:34.840> to<00:07:35.840> uh 00:07:35.950 --> 00:07:35.960 align:start position:0% oursel and it makes it much easier to uh 00:07:35.960 --> 00:07:38.869 align:start position:0% oursel and it makes it much easier to uh you<00:07:36.039> know<00:07:36.280> crawl<00:07:36.560> across<00:07:36.800> the<00:07:36.960> code<00:07:37.240> base 00:07:38.869 --> 00:07:38.879 align:start position:0% you know crawl across the code base 00:07:38.879 --> 00:07:42.670 align:start position:0% you know crawl across the code base um<00:07:39.879> you<00:07:40.000> know<00:07:40.400> I<00:07:40.560> I<00:07:40.759> I<00:07:40.840> think<00:07:41.479> projects<00:07:42.000> like<00:07:42.520> uh 00:07:42.670 --> 00:07:42.680 align:start position:0% um you know I I I think projects like uh 00:07:42.680 --> 00:07:45.950 align:start position:0% um you know I I I think projects like uh like<00:07:42.840> modular<00:07:43.800> or<00:07:43.960> or<00:07:44.120> mojo<00:07:44.919> as<00:07:45.039> you<00:07:45.159> said<00:07:45.759> uh 00:07:45.950 --> 00:07:45.960 align:start position:0% like modular or or mojo as you said uh 00:07:45.960 --> 00:07:49.110 align:start position:0% like modular or or mojo as you said uh and<00:07:46.159> and<00:07:46.319> Triton<00:07:47.120> which<00:07:47.280> are<00:07:47.599> these<00:07:48.400> uh<00:07:49.000> you 00:07:49.110 --> 00:07:49.120 align:start position:0% and and Triton which are these uh you 00:07:49.120 --> 00:07:53.149 align:start position:0% and and Triton which are these uh you know<00:07:50.039> not<00:07:50.440> these<00:07:50.639> domain<00:07:51.080> specific<00:07:52.000> languages 00:07:53.149 --> 00:07:53.159 align:start position:0% know not these domain specific languages 00:07:53.159 --> 00:07:56.029 align:start position:0% know not these domain specific languages um<00:07:54.159> are<00:07:54.879> you<00:07:54.960> know<00:07:55.039> for<00:07:55.280> for<00:07:55.440> efficient<00:07:55.759> code 00:07:56.029 --> 00:07:56.039 align:start position:0% um are you know for for efficient code 00:07:56.039 --> 00:07:59.110 align:start position:0% um are you know for for efficient code generation<00:07:57.039> are<00:07:57.240> really<00:07:57.479> interesting<00:07:58.240> uh<00:07:58.479> I<00:07:59.039> I 00:07:59.110 --> 00:07:59.120 align:start position:0% generation are really interesting uh I I 00:07:59.120 --> 00:08:01.790 align:start position:0% generation are really interesting uh I I mean<00:07:59.560> we<00:07:59.720> use<00:08:00.000> Triton<00:08:00.599> a<00:08:00.759> lot<00:08:01.159> pytorch<00:08:01.599> uses 00:08:01.790 --> 00:08:01.800 align:start position:0% mean we use Triton a lot pytorch uses 00:08:01.800 --> 00:08:03.909 align:start position:0% mean we use Triton a lot pytorch uses Triton<00:08:02.120> a<00:08:02.240> lot<00:08:02.879> a<00:08:02.960> lot<00:08:03.080> of<00:08:03.159> the<00:08:03.240> ml<00:08:03.520> Community 00:08:03.909 --> 00:08:03.919 align:start position:0% Triton a lot a lot of the ml Community 00:08:03.919 --> 00:08:08.230 align:start position:0% Triton a lot a lot of the ml Community does<00:08:04.360> as<00:08:04.520> well<00:08:05.159> I<00:08:05.240> think<00:08:05.440> this<00:08:05.560> is<00:08:06.240> uh<00:08:06.599> you<00:08:07.240> know 00:08:08.230 --> 00:08:08.240 align:start position:0% does as well I think this is uh you know 00:08:08.240 --> 00:08:09.670 align:start position:0% does as well I think this is uh you know unfortunately<00:08:08.800> kind<00:08:08.919> of<00:08:09.000> the<00:08:09.159> best<00:08:09.319> thing<00:08:09.520> we 00:08:09.670 --> 00:08:09.680 align:start position:0% unfortunately kind of the best thing we 00:08:09.680 --> 00:08:13.189 align:start position:0% unfortunately kind of the best thing we have<00:08:09.840> at<00:08:09.919> the<00:08:10.039> moment<00:08:10.280> an<00:08:10.440> open<00:08:10.879> source<00:08:11.879> uh<00:08:12.199> I 00:08:13.189 --> 00:08:13.199 align:start position:0% have at the moment an open source uh I 00:08:13.199 --> 00:08:14.550 align:start position:0% have at the moment an open source uh I uh<00:08:13.360> while<00:08:13.520> I<00:08:13.680> really<00:08:13.879> love<00:08:14.080> the<00:08:14.159> language<00:08:14.440> and 00:08:14.550 --> 00:08:14.560 align:start position:0% uh while I really love the language and 00:08:14.560 --> 00:08:17.350 align:start position:0% uh while I really love the language and I<00:08:14.639> wish<00:08:14.759> I<00:08:14.879> could<00:08:15.000> use<00:08:15.120> it<00:08:15.280> more<00:08:15.960> it's<00:08:16.680> it<00:08:16.879> just 00:08:17.350 --> 00:08:17.360 align:start position:0% I wish I could use it more it's it just 00:08:17.360 --> 00:08:20.869 align:start position:0% I wish I could use it more it's it just it's<00:08:17.560> kind<00:08:17.680> of<00:08:17.840> too<00:08:18.120> high<00:08:18.360> level<00:08:19.199> and<00:08:19.879> doesn't 00:08:20.869 --> 00:08:20.879 align:start position:0% it's kind of too high level and doesn't 00:08:20.879 --> 00:08:22.629 align:start position:0% it's kind of too high level and doesn't have<00:08:21.039> enough<00:08:21.319> performance<00:08:21.800> to<00:08:22.199> truly 00:08:22.629 --> 00:08:22.639 align:start position:0% have enough performance to truly 00:08:22.639 --> 00:08:25.629 align:start position:0% have enough performance to truly saturate<00:08:23.280> the<00:08:23.639> you<00:08:23.759> know<00:08:23.960> h100s<00:08:24.680> h20s<00:08:25.440> the 00:08:25.629 --> 00:08:25.639 align:start position:0% saturate the you know h100s h20s the 00:08:25.639 --> 00:08:28.950 align:start position:0% saturate the you know h100s h20s the these<00:08:26.000> really<00:08:27.000> uh<00:08:27.159> you<00:08:27.280> know<00:08:27.479> heavy<00:08:27.720> gpus<00:08:28.720> and 00:08:28.950 --> 00:08:28.960 align:start position:0% these really uh you know heavy gpus and 00:08:28.960 --> 00:08:30.990 align:start position:0% these really uh you know heavy gpus and I<00:08:29.080> I<00:08:29.159> do<00:08:29.560> think<00:08:29.720> something<00:08:30.039> more<00:08:30.440> like<00:08:30.840> like 00:08:30.990 --> 00:08:31.000 align:start position:0% I I do think something more like like 00:08:31.000 --> 00:08:34.389 align:start position:0% I I do think something more like like Mojo<00:08:31.560> that<00:08:32.000> has<00:08:32.120> a<00:08:32.360> sort<00:08:32.560> of<00:08:33.240> progressive<00:08:34.240> uh 00:08:34.389 --> 00:08:34.399 align:start position:0% Mojo that has a sort of progressive uh 00:08:34.399 --> 00:08:36.550 align:start position:0% Mojo that has a sort of progressive uh deepening<00:08:35.399> uh<00:08:35.479> where<00:08:35.599> you<00:08:35.719> can<00:08:35.880> get<00:08:36.080> down<00:08:36.279> to 00:08:36.550 --> 00:08:36.560 align:start position:0% deepening uh where you can get down to 00:08:36.560 --> 00:08:39.230 align:start position:0% deepening uh where you can get down to actually<00:08:36.880> defining<00:08:37.320> you<00:08:37.519> know<00:08:38.279> the<00:08:38.519> the<00:08:38.680> width 00:08:39.230 --> 00:08:39.240 align:start position:0% actually defining you know the the width 00:08:39.240 --> 00:08:41.870 align:start position:0% actually defining you know the the width of<00:08:39.479> of<00:08:39.640> registers<00:08:40.320> or<00:08:40.479> the<00:08:40.560> layouts<00:08:40.959> of<00:08:41.159> things 00:08:41.870 --> 00:08:41.880 align:start position:0% of of registers or the layouts of things 00:08:41.880 --> 00:08:44.310 align:start position:0% of of registers or the layouts of things in<00:08:42.000> a<00:08:42.080> little<00:08:42.240> bit<00:08:42.399> more<00:08:42.599> careful<00:08:43.560> uh<00:08:44.039> uh 00:08:44.310 --> 00:08:44.320 align:start position:0% in a little bit more careful uh uh 00:08:44.320 --> 00:08:46.910 align:start position:0% in a little bit more careful uh uh pathway<00:08:45.320> um<00:08:45.480> is<00:08:45.640> really<00:08:45.839> interesting<00:08:46.519> but<00:08:46.800> one 00:08:46.910 --> 00:08:46.920 align:start position:0% pathway um is really interesting but one 00:08:46.920 --> 00:08:48.470 align:start position:0% pathway um is really interesting but one of<00:08:47.000> the<00:08:47.120> things<00:08:47.360> that<00:08:47.480> we're<00:08:47.720> actually<00:08:48.240> one<00:08:48.360> of 00:08:48.470 --> 00:08:48.480 align:start position:0% of the things that we're actually one of 00:08:48.480 --> 00:08:50.310 align:start position:0% of the things that we're actually one of the<00:08:48.560> biggest<00:08:48.839> users<00:08:49.240> of<00:08:49.560> for<00:08:49.839> our<00:08:50.040> most 00:08:50.310 --> 00:08:50.320 align:start position:0% the biggest users of for our most 00:08:50.320 --> 00:08:52.389 align:start position:0% the biggest users of for our most performing<00:08:50.720> kernels<00:08:51.600> is<00:08:51.800> a<00:08:51.920> library<00:08:52.279> by 00:08:52.389 --> 00:08:52.399 align:start position:0% performing kernels is a library by 00:08:52.399 --> 00:08:54.829 align:start position:0% performing kernels is a library by Nvidia<00:08:52.800> called<00:08:53.040> cutless<00:08:53.640> M<00:08:54.480> uh<00:08:54.600> which<00:08:54.680> is 00:08:54.829 --> 00:08:54.839 align:start position:0% Nvidia called cutless M uh which is 00:08:54.839 --> 00:08:56.470 align:start position:0% Nvidia called cutless M uh which is essentially<00:08:55.200> a<00:08:55.360> code<00:08:55.640> generation<00:08:56.200> it's<00:08:56.320> a 00:08:56.470 --> 00:08:56.480 align:start position:0% essentially a code generation it's a 00:08:56.480 --> 00:08:58.269 align:start position:0% essentially a code generation it's a it's<00:08:56.560> a<00:08:56.640> template<00:08:56.959> meta<00:08:57.279> programming<00:08:57.680> library 00:08:58.269 --> 00:08:58.279 align:start position:0% it's a template meta programming library 00:08:58.279 --> 00:09:00.590 align:start position:0% it's a template meta programming library in<00:08:58.360> C+<00:08:58.839> Plus<00:08:59.519> but<00:08:59.680> it's<00:08:59.800> a<00:08:59.920> code<00:09:00.200> generation 00:09:00.590 --> 00:09:00.600 align:start position:0% in C+ Plus but it's a code generation 00:09:00.600 --> 00:09:02.590 align:start position:0% in C+ Plus but it's a code generation library<00:09:00.959> for<00:09:01.120> Cuda<00:09:01.800> that<00:09:01.959> allows<00:09:02.240> you<00:09:02.399> to 00:09:02.590 --> 00:09:02.600 align:start position:0% library for Cuda that allows you to 00:09:02.600 --> 00:09:06.829 align:start position:0% library for Cuda that allows you to Define<00:09:03.120> these<00:09:03.640> very<00:09:04.040> complex<00:09:05.279> asynchronous 00:09:06.829 --> 00:09:06.839 align:start position:0% Define these very complex asynchronous 00:09:06.839 --> 00:09:09.030 align:start position:0% Define these very complex asynchronous pipelined<00:09:07.839> uh<00:09:08.040> Matrix 00:09:09.030 --> 00:09:09.040 align:start position:0% pipelined uh Matrix 00:09:09.040 --> 00:09:12.870 align:start position:0% pipelined uh Matrix multiplications<00:09:10.079> um<00:09:11.079> in<00:09:11.200> a<00:09:11.399> way<00:09:11.800> that<00:09:12.320> uh<00:09:12.519> is 00:09:12.870 --> 00:09:12.880 align:start position:0% multiplications um in a way that uh is 00:09:12.880 --> 00:09:16.310 align:start position:0% multiplications um in a way that uh is is<00:09:13.880> much<00:09:14.120> easier<00:09:14.440> to<00:09:14.640> do<00:09:14.920> manually<00:09:15.480> than<00:09:15.839> than 00:09:16.310 --> 00:09:16.320 align:start position:0% is much easier to do manually than than 00:09:16.320 --> 00:09:20.829 align:start position:0% is much easier to do manually than than than<00:09:16.440> in<00:09:16.680> Cuda<00:09:17.640> um<00:09:18.160> so<00:09:19.040> uh<00:09:19.399> we<00:09:19.959> we<00:09:20.279> definitely 00:09:20.829 --> 00:09:20.839 align:start position:0% than in Cuda um so uh we we definitely 00:09:20.839 --> 00:09:23.030 align:start position:0% than in Cuda um so uh we we definitely are<00:09:21.200> proponents<00:09:21.560> of<00:09:21.720> code<00:09:21.959> generation<00:09:22.839> we're 00:09:23.030 --> 00:09:23.040 align:start position:0% are proponents of code generation we're 00:09:23.040 --> 00:09:25.150 align:start position:0% are proponents of code generation we're just<00:09:23.200> not<00:09:23.399> a<00:09:23.519> little<00:09:23.680> sure<00:09:24.560> uh<00:09:24.720> we<00:09:24.920> we<00:09:25.040> we 00:09:25.150 --> 00:09:25.160 align:start position:0% just not a little sure uh we we we 00:09:25.160 --> 00:09:25.870 align:start position:0% just not a little sure uh we we we definitely<00:09:25.399> don't<00:09:25.519> want<00:09:25.600> to<00:09:25.680> write 00:09:25.870 --> 00:09:25.880 align:start position:0% definitely don't want to write 00:09:25.880 --> 00:09:27.389 align:start position:0% definitely don't want to write everything<00:09:26.120> in<00:09:26.279> cut<00:09:26.480> lless<00:09:26.760> even<00:09:27.000> though<00:09:27.160> it's 00:09:27.389 --> 00:09:27.399 align:start position:0% everything in cut lless even though it's 00:09:27.399 --> 00:09:28.710 align:start position:0% everything in cut lless even though it's the<00:09:27.640> it's<00:09:27.800> the<00:09:27.959> best<00:09:28.120> thing<00:09:28.240> we<00:09:28.399> have<00:09:28.480> for<00:09:28.600> the 00:09:28.710 --> 00:09:28.720 align:start position:0% the it's the best thing we have for the 00:09:28.720 --> 00:09:30.030 align:start position:0% the it's the best thing we have for the best<00:09:28.920> performance 00:09:30.030 --> 00:09:30.040 align:start position:0% best performance 00:09:30.040 --> 00:09:31.670 align:start position:0% best performance uh<00:09:30.200> but<00:09:30.320> you<00:09:30.399> know<00:09:30.600> Triton<00:09:30.920> and<00:09:31.040> Mojo<00:09:31.399> are 00:09:31.670 --> 00:09:31.680 align:start position:0% uh but you know Triton and Mojo are 00:09:31.680 --> 00:09:34.150 align:start position:0% uh but you know Triton and Mojo are maybe<00:09:31.880> a<00:09:31.959> little<00:09:32.079> bit<00:09:32.200> too<00:09:32.399> high<00:09:32.600> level<00:09:33.600> um 00:09:34.150 --> 00:09:34.160 align:start position:0% maybe a little bit too high level um 00:09:34.160 --> 00:09:36.470 align:start position:0% maybe a little bit too high level um where<00:09:34.959> we<00:09:35.160> we<00:09:35.360> definitely<00:09:35.800> want<00:09:35.959> to<00:09:36.160> invest 00:09:36.470 --> 00:09:36.480 align:start position:0% where we we definitely want to invest 00:09:36.480 --> 00:09:39.509 align:start position:0% where we we definitely want to invest more<00:09:36.640> in<00:09:36.800> Middle<00:09:37.079> grounds<00:09:38.079> uh<00:09:38.279> but<00:09:38.600> you<00:09:38.720> know 00:09:39.509 --> 00:09:39.519 align:start position:0% more in Middle grounds uh but you know 00:09:39.519 --> 00:09:41.790 align:start position:0% more in Middle grounds uh but you know we've<00:09:40.000> uh<00:09:40.240> We've<00:09:40.399> made<00:09:40.600> compilers<00:09:41.200> before<00:09:41.720> uh 00:09:41.790 --> 00:09:41.800 align:start position:0% we've uh We've made compilers before uh 00:09:41.800 --> 00:09:44.670 align:start position:0% we've uh We've made compilers before uh within<00:09:42.040> our<00:09:42.200> company<00:09:42.680> so<00:09:43.240> you<00:09:43.360> know<00:09:43.640> we<00:09:44.120> we 00:09:44.670 --> 00:09:44.680 align:start position:0% within our company so you know we we 00:09:44.680 --> 00:09:46.030 align:start position:0% within our company so you know we we want<00:09:44.800> to<00:09:44.959> encourage<00:09:45.360> other<00:09:45.519> people<00:09:45.680> to<00:09:45.839> work 00:09:46.030 --> 00:09:46.040 align:start position:0% want to encourage other people to work 00:09:46.040 --> 00:09:48.590 align:start position:0% want to encourage other people to work on<00:09:46.200> it<00:09:47.040> sounds<00:09:47.399> great<00:09:47.839> well<00:09:48.120> thank<00:09:48.279> you<00:09:48.399> so 00:09:48.590 --> 00:09:48.600 align:start position:0% on it sounds great well thank you so 00:09:48.600 --> 00:09:50.990 align:start position:0% on it sounds great well thank you so much<00:09:48.880> it's<00:09:49.040> a<00:09:49.160> lot<00:09:49.320> of<00:09:49.440> information<00:09:50.320> and<00:09:50.880> uh 00:09:50.990 --> 00:09:51.000 align:start position:0% much it's a lot of information and uh 00:09:51.000 --> 00:09:52.590 align:start position:0% much it's a lot of information and uh we're<00:09:51.320> really<00:09:51.519> looking<00:09:51.800> forward<00:09:52.279> to<00:09:52.440> your 00:09:52.590 --> 00:09:52.600 align:start position:0% we're really looking forward to your 00:09:52.600 --> 00:09:54.430 align:start position:0% we're really looking forward to your talk<00:09:52.839> and<00:09:53.399> welcome<00:09:53.640> to<00:09:53.800> the<00:09:53.920> community<00:09:54.320> we 00:09:54.430 --> 00:09:54.440 align:start position:0% talk and welcome to the community we 00:09:54.440 --> 00:09:56.269 align:start position:0% talk and welcome to the community we hope<00:09:54.600> to<00:09:54.800> hear<00:09:55.160> more<00:09:55.360> from<00:09:55.560> you<00:09:55.760> guys<00:09:55.959> as<00:09:56.120> you 00:09:56.269 --> 00:09:56.279 align:start position:0% hope to hear more from you guys as you 00:09:56.279 --> 00:09:58.310 align:start position:0% hope to hear more from you guys as you develop<00:09:56.600> your<00:09:56.839> then<00:09:57.360> yeah<00:09:57.600> thank<00:09:57.720> you<00:09:57.880> so<00:09:58.040> much 00:09:58.310 --> 00:09:58.320 align:start position:0% develop your then yeah thank you so much 00:09:58.320 --> 00:10:00.310 align:start position:0% develop your then yeah thank you so much it's<00:09:58.440> a<00:09:58.839> pleasure<00:09:59.079> to<00:09:59.519> contributing<00:09:59.959> to<00:10:00.120> open 00:10:00.310 --> 00:10:00.320 align:start position:0% it's a pleasure to contributing to open 00:10:00.320 --> 00:10:03.320 align:start position:0% it's a pleasure to contributing to open source