This week, we’ll be featuring a series of shows recorded from Strange Loop, a great developer-focused conference that takes place every year right in my backyard! The conference is a multi-disciplinary melting pot of developers and thinkers across a variety of fields, and we’re happy to be able to bring a bit of it to those of you who couldn’t make it in person! In this show I speak with Soumith Chintala, a Research Engineer in the Facebook AI Research Lab (FAIR). Soumith joined me at Strange Loop before his talk on Pytorch, the deep learning framework. In this talk we discuss the market evolution of deep learning frameworks and tools, different approaches to programming deep learning frameworks, Facebook’s motivation for investing in Pytorch, and much more. This was a fun interview, I hope you enjoy! The notes for this show can be found at For series information, visit
United States


00:00:10hello and welcome to another episode of Twin we'll talk the podcast right interview interesting people doing interesting things and machine learning and artificial intelligence I am your host Sam Cherrington a big thanks to everyone who participated in last week's twin mall online meet up and to Kevin T.
00:00:28from signup for presenting you can find the slides for his presentation in the meet ups like channel as well as in this week's show notes are final meet up of the year will be held on Wednesday December thirteenth make sure to bring your thoughts on the top machine
00:00:43learning and they I stories for twenty seventeen for our discussion segment for the main presentation prior to Mull taught guest Bruno Gunn solve is we'll be discussing the paper understanding deep learning requires rethinking generalization by she longs Jiang from MIT and Google brain and others you can find
00:01:04more details and register at twelve Millay I dot com slash media if you receive my newsletter you already know this but trouble is growing and we're looking for an energetic and passionate community manager to help expand our programs this position can be remote but if you happen to
00:01:21be in Saint Louis all the better if you're interested please reach out to me for additional details I should mention that if you don't already get my newsletter you are really missing out and should visit to Malaysia I dot com slash newsletter to sign up now the show
00:01:37you're about to hear is part of our strange loop twenty seventeen series brought to you by our friends at next SOSUS makes us this is a company focused on making machine learning more easily accessible to enterprise developers their machine learning Apion meets developers where they're at regardless of
00:01:55their mastery of data science so they can start putting up predictive applications immediately and in their preferred programming language it's a simple is loading your date in selecting the type of problem you want to solve their automated platform trains and select the best model fit for your data
00:02:11and then outputs predictions to learn more about nexus is be sure to check out the first episode in the series at two Molly I dot com slash talks last sixty nine where I speak with co founders Ryan CV and chase in Montgomery be sure to also get your
00:02:28free nexus S. A. P. I. key and discover how to start leveraging machine learning in your next project at next doses dot com slash twinkle in this show I speak with sumit shin tala or research engineer in the Facebook AI research lab sumit join me a strange look
00:02:45before us talk on pi torch the deep learning framework in this conversation we discussed the market evolution of deep learning frameworks different approaches to programming deep learning frameworks Facebook's motivations for investing in PI torch and much more this is a fun interview and I hope you enjoy it
00:03:03and now on to the show uhhuh Hey everyone I am here at the strange loop conference in Saint Louis and I'm with sumit chin tala who is %HESITATION research engineer at fair the Facebook ad I research lab and to meet is giving a talk here at the conference
00:03:27tomorrow and graciously agreed to spend some time with us to talk a little bit about what he's up to and about as talks %HESITATION welcomes me hi Sam nice to be here strange of for the first time absolutely absolutely in welcome to Saint Louis so one we get
00:03:43started by having you tell us a little bit about your background and how you got into a I research sure let's see about eighty years ago I I wanted to be a digital artist trying to make Rita facts for movies and stuff I did an internship there and
00:04:03as soon as I went into the internship a week and I realize that I was terrible at this was not an art by any standards and then I had to find second choices in life and then I was looking at my interests and one of the things that
00:04:18struck me was how decent programmer since a young age and I I kind of like the whole computer vision angle object detection dinner you show a picture and some machine shows tells you %HESITATION there's a person here there's a cat here that just fascinated me so I went
00:04:38down that line I did a small internship at a research lab in India at this place called triple ID and like there I did a little bit of random stuff you know as an undergrad you explore all can the thing and then I tried a little bit of
00:04:56face the taxes and a little bit of taking a bunch of pictures off of monument and then stitching them together and three D. okay and I had no idea what I was doing this how to put together things based on various tools and snippets and you know how
00:05:13you called programmers stack overflow lot see this takes if it's an entire put together something and I was guilty as charged again and then I got an opportunity to come to CMU and Pittsburgh okay just doing more exploration trying to figure out what I want to do there
00:05:34I got to play a robot soccer so program relies to play soccer honestly honestly nicely you you basically like flies to allied with your program and then did they just play soccer and I've seen pictures of a variety of scenarios of these the humanoid ones yes the ones
00:05:55that I've worked on where these human I'd one okay they were called gnocchi robots and a %HESITATION Q. I a and they were I mean I came in with the expectation that the figure out so much in robotics we must be pretty good and we couldn't make them
00:06:14stand properly and they would walk and they would fall down and that's that's the state of the art if you could make them walk following that we probably all seen the video the Boston robotics yeah like tries to turn the door hacked the doorknob and and that is
00:06:30a is a so that makes shocked me because I was like oh robotics was so much more advanced than it is and and also I I saw opportunity there like %HESITATION that this bill is still kind of %HESITATION pen and and we were trying to do the whole
00:06:50algorithm of their %HESITATION by playing soccer with vision like %HESITATION can you identify where the ball is and walk towards it and stuff and that part as well was very very rudimentary not working very well saying it would still look for in our other orange pixel and your
00:07:07image and how to walk training that takes a little bigger and even something as stupid as that it wasn't working right that's pretty funny so I did a bunch of random stuff there and I had a good time at CNN I decided that I want to go go
00:07:28into artificial intelligence computer vision about X. and then I applied to a bunch of places CMU being one of them and I didn't get accepted anywhere and I was looking for a late late applications and stuff and then also combined with someone who they're like where there was
00:07:46computer vision and I saw this web page by this guy a colony on the cause and he had their legs Janke then pay some it on take a Texan stuff going on but I was like Hey it's and I you I'll give it a shot Clinton my you
00:08:05and a couple of other places in a last minute because I got rejected from my top top schools they wanted to go to no offense young I love the store and I got accepted and what you got and it wasn't super happy with myself because I know I
00:08:23can do so it's better if I worked harder and then it gets even more hilarious I come into and why you I am only on the front of like Hey I did a little bit over here and there at CMU and applied the I want to work with
00:08:38you try to do more object detection research and and he replied immunity has like Hey lets me once again here that's made this day this time no Jahn's in his office and it's like Hey listen do you know anything about it near on that for you know what
00:08:55I knew my kind of research I like nope only the turnaround or XO lines and I have no idea what would your day and say he went on to explain to me how neural networks work this is in two thousand ten you know that folks were in hot
00:09:12Anjelica and had a lot of time on his hat though that relationship and he introduced me to one of his PhD students peer sermon a who is now a good will and here and I I was peers like understudy I worked on many things there like I am
00:09:30fluent in your on that for you bill it like your bill does declining Famer call AB learn and I was kind of holding on to that and that made me understand more about how your networks work also not be stronger on the engineering side of things that's roughly
00:09:46how entered the field and my two years at NYU be published one paper and another paper on the work we did got published a and another conference and then in two thousand twelve man I graduated I couldn't find a job in deep learning while well December was in
00:10:07the whole of the turning blue started so twenty tell man I I was gonna go accept the job at Amazon as a test in there while you might wanna waste the last two years of my life is so frustrating I I was trying not to give up I
00:10:23is still super interested in the field that you know you have practical constraints right right you you need to think of all these things yeah so in the last minute I think I had to accept down is on offer by Saturday and on Wednesday yon lake and beyond
00:10:42at that that day I don't even remember why and ID on the day he was like %HESITATION where you're going and I told them I couldn't find a job anywhere else and go to Amazon is like %HESITATION just yesterday one of the company that co funded god fresh
00:10:56funding and they're looking to hire engineers so that is a conversation that happened on Wednesday I went and gave my interviews and Thursday in Princeton New Jersey and on Friday I signed for museum me and they were doing music and deep learning on sons okay so that is
00:11:15a company that was like my guess is am kind of thing now it was it was basically you you wanna know if someone's playing music you should be able to transcribe it to live on the sheet and someone takes a picture of sheet then you have to be
00:11:31able to play it back for them so it was like a full cycle I wanna hear I wanna play so it is like a tool for musicians and doesn't exist for guitar tab do you have any idea there it should and they also they do but it's not
00:11:48a solved problem to the couple many play multiple notes together if you placing alone at the time it's very easy but if you play like five or six cards at a time lady company understanding which of those exactly maps to what you played okay and still like ongoing
00:12:08problem so I spent time at museum me for a couple years we're building out in the mobile stuff I mean we wanted the whole thing to run on phones okay I was training on that for Exxon on sheet music and we call this music optical recognition Mar and
00:12:28then the company kind of had to fold at some point in the meanwhile like why does it means a me I also started actively maintaining torch which was the defining fair market that was one of the bigger ones had that time and I went to the wanted to
00:12:45get out of me as a name because we're not sure how you know the business side of things was going and then beyond started at Facebook six months before that man and they were using towards hazard mandatory for Americans so they needed good engineer as to you know
00:13:02maintain and then double torch and so I by the time I was joining Facebook I was the only man to north toward service perfect they just got the engineer who will help like had decided things okay so I came to Facebook and there were so many smart people
00:13:22I just learned so much from I was also interested in research and so I ended up going down this path of generative adversarial networks where we are trying to synthesize images so Daniel network kind of just synthesizes images from nothing or legal training or for what use cases
00:13:43this was more of an unsupervised learning use case so the answer for as learning you one of the things you do is generation and the motivation there is that yes you can generate something you have generally good concepts about how that process works so the motivation is that
00:14:05if we can do a really good demand generation neural network we can take parts of that new one network and bootstrap other new on that for which are doing computer vision tasks to get better profile so we could take parts of this no network and then make it
00:14:21to work on a different task lake dog which is class classification and without having as much data you would still get very good accuracy so that's the whole on supply semi supervised learning motivation I worked on a few things on the address so network side on and then
00:14:38coming to in this was back in twenty twelve twenty thirteen now twenty fourteen I join Facebook %HESITATION got it got anything in the twenty sixteen Gans in general have been more you know past two three years I think yeah yeah he fifteen highway sixty yeah coming to %HESITATION
00:14:57what I'm gonna talk about tomorrow what happened was towards has been an aging design in general it's it's been seven years since the previous release of towards cannot so it was becoming more inflexible you nice to feel changes there's this concert for researcher use the tools that they
00:15:16have available to them best and they push those tools to the limit and the new tools come that till then again make the researchers more flexible and exploring new things sued towards was reaching its limits of flexibility so we wanted to double the new tool and so we
00:15:36worked on and for four years started off as an intern project and then we can't kept are looking at and we released it earlier this year it's called high towards and it's what I'm gonna be talking about tomorrow it's strange though they'll be talking about high toured how
00:15:54it came about what engineering challenges we face python generally is %HESITATION is it fairly slow language but it's the most popular language for machine learning for you know statistics like all kinds of things so the most obvious choice to build something in was python because all of the
00:16:15users were familiar with and we have a huge ecosystem variant entries see for yeah use only tutorials and it's very easy to learn and stuff yeah so you had that outside but the downside was deep learning is one of those high performance computing spaces aren't every every second
00:16:36every millisecond matters but python is slow like how do you make some a package as really fast but taking a constrain that the users want to use it from pipe so generally like how we work on these challenges in various ways I'm just gonna talk about that so
00:16:56many things we did was we moved the most critical parts and to see the made a large part of the implementation lock free and let's make sure not the kind of breeze by these topics to dive into some of these sure but you know one of the things
00:17:12that is maybe an interesting place to start is and I've talked about this I think possibly on the podcast definitely in my newsletter just the the idea that it's actually kind of interesting here your story and how you know in in a lot of ways is like all
00:17:28about timing and miss timing and timing windows and things like that like I think I tore just kind of popped up on the scene if you well at a time when I think a lot of people that crown tenser flows like the heir apparent to the deep learning
00:17:44framework world right and you know I wonder if he just hearing your story of like your experiences with you know the timing cycle of of machine learning and deep learning like if that influences your perspective on this kind of the Evelyn the market evolution of you know tools
00:18:03and you know where you see what you see the opportunity is for origin is cannot we think things are going right so tends to flow popped up was a December twenty fifty in that thing it took the whole deep learning world by a bang and Google put so
00:18:22much effort evangelizing tens of low I mean from Sundar Pichai too like pretty much everyone was like Hey this is times inflow this the one who is going to be a life in next few years right so dad and you know they're huge T. and they've been putting
00:18:41green effort into generally making sure everyone are covered by tens of low there's a data scientist or declining researcher or like a production engineer and like what Google did was amazing they raise the bar for engineering for the thirty fifth mark still far high until then like if
00:19:04you think about it towards the end %HESITATION cafe these were the dominant deploying the fortunes of the were all like the %HESITATION started as one man grad student projects and you could totally see that in the the quality control was really bad they were not plan properly if
00:19:27you want to install any of those back in two thousand fourteen you would spend the day maybe more did you install that dependency this seven see the ten so flow it made the engineering bar very high they're like we're not **** around here right we want to make
00:19:46the best product out there for people and they went with a piano style programming model yeah so the tennis all programming model is very very low level which means if you want to write something like on a collision their letter then you'd spend writing so much bother plate
00:20:09cut and also like the ten service tomorrow it's because symbolic which means that it like you create a graph and then you run it's later the problem with that is if you want to debug anything you'd have to use the tolling given by tens of flow like you
00:20:27can for example debugger code by itself you have to run your model in this other virtual machine and then you can set break points in that virtual machine using tens of lows on tools meeting because it's not standard language Jack's python exam in the case of pie torch
00:20:46it's being interpreted by some virtual machine and knows how to read this graph and that is scheduled execution against this graph and so the only way to develop you know the bank to blogging and other tooling for it is via that part that is okay so the up
00:21:04sides to it are that this virtual machine can be as big or small as possible you can ship into phones you can do anything the downside is that now where you define your network was in your python VM and you're running in that we're gonna differently and there's
00:21:18this disconnect so as a double for you always have to keep thinking about how the behavior something in the tens of lowers your machine maps back to what you wrote in the python code with anyone anyone kind of one enterprise side that has a job experience you know
00:21:35nose hair that clay how different the write once run anywhere is from the practice of needing to know the internals of your heap size and garbage collection strategies and stuff like that exactly but doesn't it also give you our lease that the ecosystem is not an individual developer
00:21:52or the flexibility to decouple the V. M. from the year Kobe so meaning you know you can write your code using tenser flow and python and but have the VM written in you know go or whatever the fast concurrent highly concurrent language flavor of the day is an
00:22:09even port that over to you know distributed models are H. B. C. year yes the upside is that the VM can now be written and rewritten and many other things the downside is that you have to have a build the whole ecosystem Iran your V. M. so that
00:22:27users don't feel deprived by two yeah antenna was also like that that's the whole programming model symbolic where you define as a mother model and then you compile it and then you run it right the torch model has always been imperative which is you don't really have a
00:22:43separate VM is declared things and you don't even like you just like you just ride one plus two and it just executed like there's no like separation between declaration and execution yeah we wanted to extend the same thing to pi towards so you just write arbitrary imperative code
00:23:01and that is your new lan network itself and also lends itself to I think at least in the data science community there's a lot of popularity and flexibility around like Jupiter notebooks as a farmer you are having an imperative execution lends itself to Raymond %HESITATION yeah you can
00:23:19you can basically seen your execution as it goes you can reference things and all that yeah that's one of the biggest upside okay so %HESITATION yeah well they're building pie towards we wanted to continue the torch model imperative style the dynamic nature of things and we wanted to
00:23:37build it in the way that it also reaches a very high bar opens now yeah so that's been our core philosophy and so for Google yeah I think the way most people has interpreted their strategy behind diving deep into the tents are flow is you know they foresee
00:23:59this world where a I. workloads you know whether they're training or inference work clothes are gonna drop a ton of compute you know their business model their non advertising business model is you know heavily geared towards providing compute via the Google cloud and so they can own the
00:24:18model in which you know they can basically on the V. M. for a I and and they can be the best place to run a out workloads right what's the Facebook motivation for investing so heavily and I torch and and tooling is just not to be you know
00:24:33controlled by Google or is there more to it so Facebook motivation is two fold for Facebook a research which is a hundred odd researchers and in general for the community of A. I. we have a single point agenda at fair which is to try to solve a at
00:24:55and for that we're building the best tools out there and we keep them open just because there's nothing secret or we want to build the best possibly and I and we keep publishing about how the kind of make progress our motivation is not clear on the business model
00:25:12so much it's more like Hey we're trying to solve this very challenging problem and you see this manifest in various ways in the Facebook product as the secondhand effect like the property they're not sitting with the high researchers and say how do we improve Facebook at the product
00:25:31Facebook I research the people are independently work hang on there but as the building publish these things the product teams look at our research and the like %HESITATION you can implement the staying in our product in this way and that is just gonna be better products experience for
00:25:51everyone like some examples are we've had the accessibility interface improved quite a bit recently about a year year and a half ago where now if you're if you're a user who is blind or near line you can you face but like you can basically touch Facebook as as
00:26:09you would any to tell you what's going on before you test a pitcher into the ditch and just not tell it to say a picture posted by this person okay but if you touch a pitcher now it actually tells you %HESITATION it's a pitcher where a ball is
00:26:23playing with the cat and it's it's very descriptive and similarly were trying to do the same for videos as well okay that's one manifestation others are where you want to break language barriers so let's have five hundred friends like some of them I met in various places in
00:26:42traps and make it one of my good friends right so huge post in Chinese I a still want to be able to know what it's about so we have to translate feature admitted into Facebook all powered by Facebook research on you see these manifest in various areas under
00:26:59product you is I think we you see that of all this while like if I remember correctly the Facebook products only relatively recently switched to switch the way they did translation to an oral translation I think those a blog post on the front lawn maybe two months ago
00:27:17so very interesting okay so your talk so kind of walk us through the the main point your talks basically you year at the high level you've got pie towards it's inherently built around python but you need to find ways to overcome the limitations of python kinda how you
00:27:37do that was at the main thrust of the talk or yeah so it's mostly just the general engineering talk about predators okay son I give like in our view of high towards how it works okay strains of audience doesn't necessarily know about defining computations right more all over
00:27:54so right on it's like a lot of diversity in programming models the use and like strange to pass everyone under the sun yeah so which is part of what makes a great awesome I like it is my first fan club I've just seen the session so far and
00:28:10they look in the sessions lined up for tomorrow it's it's awesome so interesting factoid the guy who founded strange loop kind and Alex Miller trees and grew out of a beat up he had here town called lambda lounge it looked at by functional which is an add a
00:28:29start up that I was that years ago this is probably eight years ago now %HESITATION he basically hosted and debated this this me that they would be in an office I'd be sitting there late at night try to finish my work and hearing people talk about my own
00:28:44ads and stuff I just had no clue nice I didn't realize since Lewis had had like a big community and like I had like this is all new to so I'll be talking a little bit about high towards start off the Denali how to relate it to other
00:29:04things now and then a little bit of deep learning workloads to general challenges why they're very different from let's make those in turn like how how should people think about by torch and relating it to things that they not sure let's say you're I mean the most common
00:29:20thing everyone knows about his house right so let's look at Java script if you're trying to build a compiler for Java script the things you must care about is Java script a code which is very bright and she has a lot of control so when it's very very
00:29:39the cast insensitive if your billing very high performance compiler for Java script you will build it innovative that you'll try to optimize for like branch prediction then and like try to get faces and do trace compilation you're twenty saying like for example if you look at rooms javascript
00:29:59compiler the eight yeah VA it's it's facing Janet and or you can look at legit as another example is that other tracing **** and they do %HESITATION they'll quickly traced through upcoming code and then if something's compile little bill primetime could generate really quickly and these are all
00:30:18in the order of nanoseconds even because they're very very small computations and they're very bright and she and doing something like low place staying or standard auction these are the things that would really go well let such workloads okay with deep learning what you do is you do
00:30:38operations on tensors let's say like an eventual matrices and usually are doing a competition between a and B. and N. B. are not too integers their two thousand integers are two hundred thousand teachers %HESITATION is %HESITATION is this an intentional matrix right and and when you're doing these
00:30:57calculations you're basically doing a lot of multiplications of multiplications or like any kind of point why is a reduction or some kind of convolvulus moving window kind of operation these are the most common things and it sorry so when you're trying to make these like something like this
00:31:17more efficient you look at how fast you can how fast like how much you can paralyze each of these operations individually and it turns out almost all of these operations are banned it bond so these operations can run as fast as how how fast your memory manner at
00:31:38this okay like how fast you can get it in and out of the sea view because inside the CPU you're just doing a small multiplication or like you know an exponential but it's still much more expensive to get two hundred thousand numbers into the CPU and out of
00:31:53the sea so the way you do optimization when when building says stings let's say you're building a compiler for these things aren't you would try to fuse a bunch of these point was operations bunch of these reduction operations into an inner loop and then what you do is
00:32:13you would get these cancers in instead of doing one operation putting it back out in the result and doing another office and putting in the result back out you try to get the tensor in new seven operations at once and then get get all like the result of
00:32:30the seven of them %HESITATION because that would make it more compute bond rather than bandit interest so taking a step back the analogy to Java script is primarily to say you build these high performance compilation and execution environments by understanding the property of the language that you're working
00:32:48with and optimizing around that and we can do that here by in this example you would take %HESITATION that is you know fundamentally rate and then a very in a red is serial kind of way but maybe paralyze are on full those loose I don't know if that's
00:33:06the right way of thinking about it but said tile loops what's that Talbot handling it's called tiling because you can you can bring the combination to %HESITATION I have two hundred thousand of these okay I'll just make tiles of twenty and send them to like twenty different processor
00:33:21okay interesting wrinkle especially with GPUs when GPU Z. at three thousand cores on your GPU right so you want to like break this competition down and feed those into all those separate processors right and then like get the results back but they don't fundamentally change the band with
00:33:40issue they down yeah so %HESITATION I'm gonna be talking a bit about the jet that to be built and apply towards people to just in time compiler it's also tracing did but itself a very different kind arch facing is not in the order of nanoseconds it's in the
00:33:58order microseconds okay but that's completely fine because general deep learning workloads are in the order second and the kind of organizations classes we write as well are as I said more like fusion and batching by bashing I mean lets you do competition X. Y. Z. but between X.
00:34:20competition accent Z. their shared operations let's X. also does multiply since he also does multiplies so the dancers and while an accent I tenses and all in zero very small what you do is you create a multiply operation combine the two tensors that are involved in there and
00:34:38then after the result comes out separate them so there's a call dynamic bashing yep so we've been writing this jet that's very new for us make as in not not many people have work generally in this direction answer flows building one called excel a it's a compiler and
00:34:57that kind of automation's they're doing as well are very similar in nature like everyone's exploring now these like how to make tense or competition pastor we're just taking the just in time approach and they're taking data time analysis of pros and I think the %HESITATION jet and this
00:35:14probably comes from really Java script as something that is more relevant %HESITATION interactive types of work closed and bats but deep learning is primarily a batch workloads it leads to training part of it depends on how you see it like we want to keep the interactive ness because
00:35:33remember what we talked about people like to keep that interactive python any I bite on on book style right programming model we want people to keep that flexibility really where you need the bike the super high performance every like that's people do interactively their programming AGP is like
00:35:57this is not a okay so the norm in deep learning okay so you're backed by your super powerful GPU okay and I'll talk a little bit about this and and talk tomorrow about how you can you hide %HESITATION centers can just be transferred to the GPU and you
00:36:14operate on that and all the operations are now being done on the GPU with very high performance but you're doing this in a very attractive that it is a little bit of a mindset for folks that have been kind of Amursana tenser flow or and or a batch
00:36:29oriented world where you kind of create this job he sent it off to the cloud or and then to train and then you check on it a few days we are the worst nightmare for hardware developers because that that they're all building solutions there on %HESITATION let's say
00:36:46you build this model before hand and then you give it to our hardware yeah which lets say takes thirty minutes but it will match your model and the most effective way to our hardware and after that thirty minutes is done if you pump any any images and or
00:37:04any any inputs and it'll be like super fast and we're like well that's kinda down with our model we take a different model in every single interests and what are you gonna do right right interesting so speaking of hardware and hardware developers and enough this is something that
00:37:24you're close to it all the Facebook also is very involved in this %HESITATION C. P. open compute project I was involved in the Big Sur and the big basin where you at okay so do you see as I understand this O. C. P. is primarily been oriented around
00:37:40kind of off the shelf stuff like system architecture as opposed to you know board level architecture anything like that but do you see a future where %HESITATION CP takes on like this band with problem in %HESITATION try to make hardware that's more suitable for these kinds of workloads
00:37:57or are those other people's problems I think %HESITATION see if he is is one of those huge nasty wide efforts right I think it's totally possible that under %HESITATION CP you'll get something that's like a custom ASIC and I'm not sure how that will happen in jet yeah
00:38:15because there's so many players and so many people who can contribute to a CP just two days ago I think our yesterday and media release an open back with their log files in his steel files off the chip that does convolutional networks make high performance collision on network
00:38:36based thing is based on the glass on the headline for that is like this is the media's TV you right I mean that was the yeah the click so they basically put it out there you can take them very long inmates deals basically the the code for fabricating
00:38:53these chips are yeah you'll have to still like remapped well log into actual the actual process your your manufacturing under but that's a mechanical process that usually like you then you can usually give you are at steel files to someone and then they'll spend their cut that a
00:39:11company will spend remapping whatever you have to the process most effectively idea it's it's a fairly mechanical process right and the points of now nvidia just leave this open ship yeah and that's totally one of the candidates for example open compute thank open computers all right okay everything
00:39:34is open you can reproduce over this year right so I think they're still not sure about what's gonna come into CP in that direction like I personally don't know the details so we'll see okay so you also mention that parts of the pie torch engine are written in
00:39:57C. makes me think a little bit of I can and just kind of the regular python community there's time remembering correctly there's like pie pie insights on and all these different implementations of the python interpreter is is the same general idea they're not at all hi pie is
00:40:15a replacement for the defaults python implementation which is called see by thought okay so python is a programming language and they have a base implementation of that programming language that the language developers have love bondage is because he bet on it it's written in C. and if you
00:40:36tie python and do your desktop usually that's what it is and site on is just a very cute fail Friday extensions to python like to see by done mostly okay hi pie is a replacement interpreter for C. by it's not written in that can interpret typed up I'm
00:40:57not sure what is written by I don't think it's written and pipe okay the implementation is probably not in python okay bye bye is a just in time interpreter for python is forbidden in park assembly in part C. or whatever okay okay so the idea of being the
00:41:17significance of it is not its implementation but just in time verses yeah not just okay hi towards is just a CPython extensions like all the C. bits we have are the cool and you can find his number hi I'm pie is ninety percent or didn't see yeah okay
00:41:39but it has it's a CPython extensions that is it's not an independent C. library that you can run it's like heavily integrated into the sea python API so quite orders just like a python extension okay these recently been an announcement of you know one of the first like
00:41:58I think kind of broadly publicize hi torch wins if you will is like fast at a %HESITATION I decided to re write all of their course wherein into battle very pleasant surprise I was as soon as fighters came out at them they've tried it and they they find
00:42:16it really effective especially for teaching and the barrier to entry so they switched over to Patterson tens of flow and group supporting them in any way we can yeah yeah it's interesting I think if you know if there's anything that this community benefits from his options and especially
00:42:38options that you know have major you know both major companies behind them pushing them forward but also that are open to community engagement and community contributions and so it's definitely great to see hear from and kind of industry observer point of view that you know we've got you
00:42:57know it what's starting to look like a a second kind of really strong contender at a time when you know again I think a lot of people said oh yeah it's just you know it's center so congrats for your part in that and thanks so much for taking
00:43:13the time to sit down with the ship alright everyone that's our show for today thanks so much for listening and for your continued feedback and support for more information on to me for any of the topics covered in this episode head on over to twin only I dot
00:43:34com slash talks last seventy follow along with our streams live twenty seventeen series visit tamale I dot com slash S. T. loop of course you can send along your feedback or questions via Twitter to add to Molly I or at Sam Cherrington or leave a comment right on
00:43:53the show notes page thanks again to neck so says for their sponsorship of the show check out to Molly I dot com slash talks last sixty nine to hear my interview with the company founders and visit next so says dot com slash twin mall for more information and
00:44:10to try their API for free thanks again for listening and catch you next time

Transcribed by algorithms. Report Errata
Disclaimer: The podcast and artwork embedded on this page are from Sam Charrington, which is the property of its owner and not affiliated with or endorsed by Listen Notes, Inc.


Thank you for helping to keep the podcast database up to date.