[HN Gopher] An Interview with Nvidia CEO Jensen Huang About AI's... ___________________________________________________________________ An Interview with Nvidia CEO Jensen Huang About AI's iPhone Moment Author : pps Score : 84 points Date : 2023-03-25 15:26 UTC (7 hours ago) (HTM) web link (stratechery.com) (TXT) w3m dump (stratechery.com) | crop_rotation wrote: | Nvidia seems in a really great position. They are to AI what | Intel was to the PC (and unlike the PC era there is not a single | Microsoft here who controls the entire ecosystem). CUDA still has | no alternative. Yes Google has TPUs but outside of Google NVIDIA | still dominates and enjoys the network effects (support in all | kinds of libs and framework). They face the same problems that | Intel faced, as in the market just wants the software and if the | software works the same, the hardware is replaceable. It will be | interesting to see how they adapt. | TechnicolorByte wrote: | Interesting comparison with Intel in the PC market: | | > the market just wants the software and if the software works | the same, the hardware is replaceable | | And likely to be true given how much competition is heating up | in the AI hardware space. Granted, many of these competitors | and startups especially have existed for years and haven't made | much of a dent. Even Google's TPU doesn't seem _that_ much | better than Nvidia's stuff based on their limited MLPerf score | releases. Maybe this "iPhone moment" for AI will change that | and force competitors to finally put some real effort in it. | | As for Nvidia, looks like they are trying to adapt by selling | their own enterprise software solutions such as Omniverse and | their AI model customization stuff. Will be interesting to see | if they can transform into more of a software solutions | provider going forward. | Giorgi wrote: | Why is this text though? Is there video? | jaflo wrote: | What's wrong with text? | TechnicolorByte wrote: | Looks like there's a paid podcast available. I'm sure there's | some TTS apps that could take in the text and make it | consumable that way, though. | visarga wrote: | NaturalReaders is a good one | tmsh wrote: | Lots of great quotes like: "Inference will be the | way software is operated in the future. Inference is simply a | piece of software that was written by a computer instead of a | piece of software that was written by a human and every computer | will just run inference someday." | ww520 wrote: | You can argue compilers have been doing that. And JIT was the | next step to enable runtime rewriting of software. AI directed | JIT or compilers probably will be next. | bottlepalm wrote: | Man how lucky is Nvidia? crypto winds down just as AI is ramps | up. | PartiallyTyped wrote: | Nvidia is part of the reason why it happened and why it | happened _now_ , they didn't get lucky by any stretch of the | imagination wrt to AI. | greatpostman wrote: | I wouldn't call it luck, just a great product that's enabling | societal change | ralphc wrote: | NVidia got lucky, at least in the long term. They spent years | perfecting SIMD to make spaceships blow up and, by | coincidence, that's the same technology that enabled coin | mining then deep learning. | orangepanda wrote: | Is Nvidia taking advantage of the crypto/AI trends, or enabling | them? | rchaud wrote: | they're selling shovels in a gold rush, that's the best | business to be in. | danpalmer wrote: | They have been investing in this for a long time, CUDA is 15 | years old. That said, they have plenty of competition now and | that's only going to increase. | | - Apple have got their own chips now for mobile AI applications | for iPhones. | | - Google have got their own chips for Android and servers. | | - AMD are gaining on Intel in the server space, and have their | own GPUs, being able to sell CPUs and GPUs that complement each | other may be a good strategy, plus AMD have plenty of their own | experience with OpenCL. | noogle wrote: | How much of a moat is CUDA? | | It's indeed ages beyond any of their competitors. However, | most ML/DS people interact with CUDA via a higher-level | framework. In recent years this community has consolidated | around a few (and even only one platform, PyTorch) framework. | For some reason AMD had not invested in platform backends, | but there is no network effect or a vendor lock-in to hinder | a shift from CUDA to ROCm if it is supported equally well. | floatngupstream wrote: | There is an enormous investment beside the training side. | Once you have your model, you still need to run it. This is | where Triton, TensorRT, and handcrafted CUDA kernels as | plugins come in. There is no equivalent on ROCm for this | (MIGraphX is not close). | A4ET8a8uTh0 wrote: | Yeah... but have you tried using AMD GPUs for any LLMs for | example? All the interesting stuff that is publicly released | is for Nvidia. I would love to be able to focus on AMD since | Nvidia has been adding some anti-user features lately. | pier25 wrote: | > _Google have got their own chips for Android_ | | What chips for Android? | mastax wrote: | I'm not convinced AMD has a good AI play. (Disclaimer: I hold | a long position in AMD). | | AI hardware seems like it can be much simpler than GPGPUs, | given the successful implementations by many companies | including small startups. | | AI hardware _software_ seems like it is extremely difficult. | Making a simple programmer and ops interface over a massively | parallel, distributed, memory bandwidth constrained system | that needs to be able to compile and run high performance | custom code out of customers ' shifting piles of random | python packages. | | AMD has continuously struggled at (2) and hasn't seemed to | recommit to doing it properly. AMD certainly has silicon | design expertise, but given (1) I don't think that is enough. | | Xilinix is in interesting alternative path for products or | improving AMDs software/devex. I'm not sure what to expect | from that, yet. | jacooper wrote: | AMD GPUs are a joke for anything but gaming | sofixa wrote: | Yet. Their GPUs were not great for years, but have managed | to catch up (and even overperform for their price point), | so other workloads are yet to come. | Gigachad wrote: | Still waiting for rocm support on the 5700xt which they | kept promising was ready any day now. | jacooper wrote: | Unfortunately with the way ROCm is developed, and how its | only intended for specific GPUs, I doubt it. | danpalmer wrote: | At the moment maybe, but owning the CPU may allow for | better integration. | | Imagine if AMD launch a new bus implementation from CPU to | GPU. That's not something Nvidia can do by themselves. | Maybe Nvidia buys Intel and does it though! | imwithstoopid wrote: | Nvidia is in an excellent position - they have CUDA as you | point out, and they are moving that into server room | dominance in this application space | | Google has TPUs but have these even made a tiny dent in | Nvidia's position? | | I assume anything Apple is cooking is using Nvidia in the | server room already | | Intel seems completely absent from this market | | AMD seems content to limit its ambitions to punching Intel | | its Nvidia's game to lose at this point...I wonder when they | start moving in the other direction and realize they have the | power to introduce their own client platform (I secretly wish | they would try to mainstream a linux laptop running on Nvidia | ARM but obviously this is just a fantasy) | | if anything, I think Huang may not be ambitious enough! | losteric wrote: | > I assume anything Apple is cooking is using Nvidia in the | server room already | | For training, sure. For inference, Apple has been in a | solid competitive position since M1. LLaMa, Stable | Diffusion, etc, can all run on consumer devices that my | tech-illiterate parents might own. | smoldesu wrote: | LLaMa and Stable Diffusion will run on almost any device | with 4gb of free memory. | rolenthedeep wrote: | > AMD seems content to limit its ambitions to punching | Intel | | What's the deal with that anyway? A lot of people want a | real alternative to Nvidia, and AMD just... Doesn't care? | | I guess we'll have to wait for intel to release something | like CUDA and _then_ AMD will finally do something about | the GPGPU demand. | roenxi wrote: | I was wondering the same thing and thinking about it. | | When AMD bought ATI they viewed the GPU as a potential | differentiator on CPUs. They've invested a lot of effort | into CPU-GPU fusion with their APU products. That has the | potential to start paying off in a big way sometime - | especially if they figure our how to fuse high end GPU | and CPU and just offer a GPGPU chip to everyone. I can | see why AMD might put their bets here. | | But the trade off was that Nvidia put a lot of effort in | doing linear algebra quickly and easily on their GPUs and | AMD doesn't have a response to that. Especially since | they probably strategised on BLAS on an APU. But it turns | out there were a lot of benefits to fast BLAS and Nvidia | is making all the money from that. | | In short, Nvidia solved a simpler problem that turned out | to be really valuable, it would take AMD a long time to | organise to do the same thing and it may be a misfit in | their strategy. Hence ROCm sucks and I'm not part of the | machine learning revolution. :( | potatolicious wrote: | > _" Google has TPUs but have these even made a tiny dent | in Nvidia's position?"_ | | This seems unknowable without Google's internal data. The | salient question is: "how many Nvidia GPUs would Google | have bought if they didn't have TPUs?" | | The answer is probably "a lot", but realistically we don't | know how many TPUs are deployed internally and how many | Nvidia GPUs it displaced. | paulmd wrote: | Tesla and the Dojo architecture is another interesting | one - that's another Jim Keller project and frankly Dojo | may be a little underappreciated given how everything | Keller touches tends to turn into gold. | | https://www.nextplatform.com/2022/08/23/inside-teslas- | innova... | | Much like Google, I think Tesla realized this is a | capability they need, and at the scales they operate, | it's cheaper than buying a whole bunch of NVIDIA product. | jitl wrote: | Tegra & later Shield were attempts to get closer to full | end user platform. The Nintendo Switch is their most | successful such device -- with a 2-year old Tegra SKU at | launch. But going full force into consumer tech is a | distraction for them right now. Even the enthusiast | graphics market, which should be high margin, is losing | their interest. They make much more selling to the big | enterprise customer CEO Jensen mentions in the open | paragraph. | echelon wrote: | Gamers are going to be so pissed. They subsided the | advance in GPU compute and will now be ignored for the | much more lucrative enterprise AI customers. | | Nvidia is making the right call, of course. | anonylizard wrote: | Have they ever considered that the subsidy goes the other | way? The margins on an A100 card is probably 100% higher | than a RTX4090. Gaming industry is also like THE first | industry to be revolutionized by AI. Current stuff like | DLSS and AI-accelerated path tracing are mere toys | compared to what will come. | | Nvidia will not give up gaming. When every gamer has a | Nvidia card, every potential AI developer to spring up | from those gamers, will use Nvidia by default. It also | helps gaming GPUs are still lucrative. | dleslie wrote: | It's OK, gaming is also having its AI moment. | | I fully expect future rendering techniques to lean | heavily on AI for the final scene. NeRF, diffusion | models, et cetera are the thin end of the wedge. | smoldesu wrote: | Gamers are in heaven right now. Used 30-series cards are | cheap as dirt, keeping the pressure on Intel/AMD/Apple to | price their GPUs competitively. The 40-series cards are a | hedged bet against anything their competitors can develop | - manufactured at great cost on TSMC's 4nm node and | priced out-of-reach for most users. Still, it's clear | that Nvidia isn't holding out their best stuff, just | charging exorbitant amounts for it. | layoric wrote: | Where are these cheap as dirt 30 series? A 10gb 3080 is | still over $500 usd used ($750 aud) when I've looked. | When did secondhand GPUs that still cost the same as a | brand new PS5 start to be considered cheap? | my123 wrote: | A PS5 is _significantly_ slower than a 3080, it's more | RTX 2070 tier. | capableweb wrote: | > I assume anything Apple is cooking is using Nvidia in the | server room already | | I wouldn't be so quick at assuming this. Apple already ship | ML-capable chips in consumer products, and they've designed | and built revolutionary CPUs in modern time. I'm of course | not sure about it, but I have a feeling they are gonna | introduce something that kicks up the notch on the ML side | sooner or later, the foundation for doing something like | that is already in place. | imwithstoopid wrote: | Apple has no present experience in building big servers | (they had experience at one point, but all those people | surely moved on) | | Mac Minis don't count | | Sure, they are super rich and could just buy their way | into the space...but so far they are really far behind in | all things AI with Siri being a punchline at this point | | if anything, Apple proves that money alone isn't enough | capableweb wrote: | I'm no Apple fan-boy at all (closer to the opposite) so | it pains me a bit to say, but they have a proven track- | record of having zero experience in something, then | releasing something really good in that industry. | | The iPhone was their first phone, and it really kicked in | the smartphone race into high gear. Same for the Apple | Silicon processor. And those are just two relatively | recent examples. | danieldk wrote: | The iPhone had a lot of prehistory in Apple, from Newton | to iPod. Apple Silicon alo has a long history, starting | with the humble beginnings as the Apple A4 in 2010, which | relied on Samsung's Hummingbird for the CPU and PowerVR | for the GPU (plus they acquired PA Semi in 2008). | | So both are not very good examples, because they build up | experience over long periods. | capableweb wrote: | > So both are not very good examples, because they build | up experience over long periods. | | They are examples of something they could similarly do | for the Apple Neural Engine but in a bigger scale in the | future. They have experience deploying it in a smaller | scale/different versions, they would just have to apply | it in bigger scale in order to be able to compete with | NVIDIA. | rchiang wrote: | To be fair, Apple released their iPhone after building | iPods for 6 years. So, it's not like they had zero | experience with handheld devices at the time. | | Also, while Apple did create their first chip (at least | of their current families) in 2007, they did acquire 150 | or so engineers when they bought PA Semi in 2008. So, | that gave them a leg up compared to building a chip team | completely from scratch. | newsclues wrote: | I assumed their server experience is still working in the | iCloud division. | smoldesu wrote: | > Apple already ship ML-capable chips in consumer | products, and they've designed and built revolutionary | CPUs in modern time. | | Has Nvidia not done that too? They shipped ML-capable | consumer hardware before Apple, and have revolutionary | SOCs of their own. On top of that, they have a working | relationship with the server/datacenter market (something | Apple burned) and a team of researchers that basically | wrote the rulebook on modern text and image generation. | Then you factor in CUDA's ubiquity - it runs in cars, | your desktop, your server, your Nintendo Switch - Nvidia | is terrifying right now. | | If the rest of your argument is a feeling that Apple will | turn the tables, I'm not sure I can entertain that | polemic. Apple straight-up doesn't compete in the same | market segment as Nvidia anymore. They cannot release | something that seriously threatens their bottom line. | dividedbyzero wrote: | > They cannot release something that seriously threatens | their bottom line. | | If they manage to move a significant part of ML compute | from datacenter to on-device, and if others follow, that | might hurt Nvidia's bottom line. Big if at this point, | but not unthinkable. | smoldesu wrote: | There are a lot of problems here though. First of all | being that inferencing isn't hard to do - iPhones were | capable of running LLMs before LLaMa and even before it | was accelerated. _Anyone_ can inference a model if they | have enough memory, I think Nvidia is banking on that | part. | | Then there's the issue of model size. You can fit some | pruned models on an iPhone, but it's safe to say the | majority of research and development is going to happen | on easily provisionable hardware running something | standard like Linux or FreeBSD. | | And all this is ignoring the little things, too; training | will still happen in-server, and the CDN required to | distribute these models to a hundred million iPhone users | is not priced attractively. I stand by what I said - | Apple forced themselves into a different lane, and now | Nvidia is taking advantage of it. Unless they intend to | reverse their stance on FOSS and patch up their burned | bridges with the community, Apple will get booted out of | the datacenter like they did with Xserve. | | I'm not against a decent Nvidia competitor (AMD is | amazing) but the game is on lock right now. It would take | a fundamental shift in computing to unseat them, and AI | is the shift Nvidia's prepared for. | KeplerBoy wrote: | why wouldn't they build a relatively small cluster for | training tasks using Nvidia hardware? It's simply the | industry standard, every researcher is familiar with it | and writing a custom back-end for pytorch that scales to | hundreds of nodes is no small task. | | I doubt Apple cares about spending a few hundred million | dollars on A100s as long as they make sure the resulting | models run on billions of apple silicone chips. | danpalmer wrote: | > I assume anything Apple is cooking is using Nvidia in the | server room already | | I don't think Apple's server side is big or interesting. | Far more interesting is the client side, because it's 1bn | devices, and they all run custom Apple silicon for this. | Similarly Google has Tensor chips in end user devices. | | Nvidia doesn't have a story for edge devices like that, and | that could be the biggest issue here for them. | bottlepalm wrote: | They've been investing for a long time, but it's only blown | up in the past year due to recent breakthroughs. Good timing | for Nvidia. | jonas21 wrote: | It's been blowing up since at least 2012-13 when deep | convolutional neural nets started seeing massive success. | bottlepalm wrote: | That's not blowing up. What's happening right now is | blowing up. | jonas21 wrote: | What's going on now is the continuation of a growth trend | that started a decade ago. | jitl wrote: | They nerfed hash rate on their cards multiple times | selectodude wrote: | And drug manufacturers spend a lot of money to nerf their | medications to make them harder to inject. Not all customers | are good customers. | ww520 wrote: | It's not just luck but good strategy. In the past 10 to 15 | years, Nvidia has been leveraging its core GPU to go beyond | gaming video card to expand into different peripheral areas | where massive parallel computing is needed, such as super | computing, cloud computing, animation farm, CAD, visualization, | simulation, car, VR, AI, and Crypto. They have been able to | catch/enable one wave or the other because it's part of their | roadmap. | tpmx wrote: | They have been going pretty much everywhere with their CUDA | runtime. LLMs was a random hit. | | At the same time, it doesn't seem like a great moat - I think | AMD should be able compete pretty soon. ___________________________________________________________________ (page generated 2023-03-25 23:00 UTC)