저자: Alex · 외국인
요약 보기
엔비디아는 매출·마진 강세로 AI 칩 점유율 70% 이상을 보유. AMD MI300은 출시가 늦고 Transformer 엔진이 없어 2위 대안으로 평가. 인텔 Gaudi 2/3는 가격 경쟁력 있는 도전자로 제시되며 Stability AI 등 고객 확보. 구글·MS·아마존은 자체 칩을 개발하지만 엔비디아와 보완 관계로 파트너십 유지. 단기적으로 엔비디아의 강세가 이어질 것으로 봄.
다듬은 글
there's no doubt that Nvidia is on the hottest Financial streak in its 30-year history with revenues doubling year over year margins over 70 percent and a rock solid balance sheet it's no wonder that Nvidia stock has tripled year to date and with over a 70 market share for AI chips there's really only one question left to ask is NVIDIA truly Untouchable like Tesla or will competition kill nvidia's margins and their market cap in the process your time is valuable so let's get right into it AMD has been competing with Nvidia in the GPU market for many years now so once Nvidia started shipping their h100 gpus and their hgx 8 GPU systems it was only a matter of time before AMD announced a competing product line and that's exactly what they did with their Mi 300X the Mi 300 systems will come as a single accelerator or on an 8 GPU board called their Instinct platform similar to nvidia's hgx platform I actually think that amd's Mi 300 is an impressive chip especially given how quickly AMD announced it but it has some serious challenges in front of it first nvidia's h100 systems are already shipping in large volumes while amd's mi-300s won't even start shipping until at least the end of the year and second mi-300s don't have Transformer engines in them like nvidia's h100s do that's important because Transformers are used in a wide variety of AI applications from large language models to image processing techniques like object detection nvidia's Transformer engine roughly triples the h100s performance for large language models and on top of that they recently announced a piece of software called tensor RT llm which boosts GPU performance by another 100 for inference specifically so nvidia's chips currently have a big weed in terms of training AI models and an absolutely massive lead when it comes to running them but that doesn't mean that amd's chips will be dead on arrival especially if their price correctly instead I think there'll be the clear second choice after nvidia's h100s at least until Nvidia replaces them with their three nanometer Blackwell chips and if Nvidia is still Supply constrained into next year the extra demand could be amd's to capture and we can't talk about Nvidia and AMD without also talking about Intel Intel just hosted their big innovation Day event last week and of course AI was a Hot Topic Intel has been saying there are really two markets for AI processing one for large-scale infrastructure and AI training and another for much smaller workloads including AI inference for AI training which is the market we're focused on today Intel currently has their Gowdy 2 accelerators which were introduced last year to compete with nvidia's gpus while the h100s are built using a five nanometer process Gaudi 2 is still using seven nanometer technology and while Intel's chips under performing videos by anywhere from nine percent in servers to 28 in offline applications that's actually a small enough Gap to think of these chips as real competition for NVIDIA especially if intel is willing to price them in a way that beats Nvidia in terms of performance per dollar Intel also showed that eight Gowdy 2 chips outperformed 8 Nvidia h100s by over 30 percent when it comes to training multimodal AI models meaning models that can interpret some combination of text images audio video and so on multimodal AI models are already a big part of the generative AI market today so this could end up being a pretty big deal for Intel and Gaudi 2 already has at least one big generative AI customer so far stability AI which is the company Behind Popular generative models like stable diffusion is building a 4000 node Gowdy cluster specifically for AI training here's a quick clip from cnbc's closing bell where Intel CEO Pat gelsinger shares exactly how they plan to compete with Nvidia and how their own Advanced chip production plans are going so far down body you you announced for Gowdy 2 that's your existing chip a pretty big customer instability AI so that's heartening but Dowdy 3 is the one that's really supposed to close the gap with Nvidia is it on schedule how soon in calendar 24 can we expect it yeah so I say overall you know Gowdy 2 Gaudi 3 next year we also disclose Falcon Shores here we have a robust road map for 23 24 and 25.
but as you say it's the warm-up for Gowdy 3 next year we have a first silicon out of Fab it's now in packaging looking very healthy so far so we'd say the road map is on track and the ml perf scores you know the the key machine learning Benchmark today hey it's showing us we are very competitive with the market leader Nvidia and obviously Gaudi 3 will be a big step up you know from there and I disclosed the first details on that today but we're gaining momentum right now and the Market's starting to realize there's another opportunity here for AI leadership in the industry now manufacturing right because these five nodes in four years for those who aren't you know Fab Geeks that's moving at about twice the pace that Intel used to move so are you on Pace yeah in the five nodes we first Intel seven done that's you know Sapphire Rapids Alder Lake Raptor Lake the pro the PC products that are now in volume intel four are meteor Lake product or today we announced the brand for that called Intel Core Ultra right you know with the new AI capabilities so that's now ramping in volume so we'll say two are done and we've showed the server products for next year what we call our Granite Rapids and Sierra Forest and they're on our head of schedule you know so we'd say the next two uh are on track and then we showed the first uh Intel 20a and this is super important because this is the new transistor the new backside power you know these are the breakthroughs but the culmination is 18a right and what we said today is hey we'll soon be releasing the final design rules for that in the very near future so that the industry can start and I'll be sending my first product in the early part of next year for manufacturing so it's on track so two done three on track we're feeling really good about getting back to process leadership okay doing that from a manufacturing base in the U.S and Europe This is spectacular I've actually been really impressed with Intel's chips lately Pat gelsinger also revealed that they're already working on Gowdy 3 which will be built using a five nanometer process and have twice as much compute power 50 more bandwidth and 50 more memory capacity as well so we'll have to wait and see when these chips actually come out and at what price but Intel's AI chips could play a big role in their comeback Story by helping them capitalize on nvidia's Supply constraints as well so if they can do that and catch up to tsmc in terms of Chip fabrication technology but without all the risks of being so close to China I could definitely see a lot of institutional and Retail investors putting Intel back into their portfolios after all that's why Warren Buffett invested in tsmc over Intel in the first place only to sell out completely a few months later by the way if you want a great app for stock market research and investing check out MooMoo this is the best app for tracking institutional investors if I click on the US markets page and then on institutional tracker I can check major investing firms like Berkshire Hathaway and darkenvest Warren Buffett's top 10 Holdings account for over 85 percent of his portfolio which makes sense because he's always talking about the power of concentration and if I scroll down I can see which positions he's increased and decreased new positions and which positions he sold out of I can also see all of his individual Holdings Apple Bank of America American Express Coca-Cola and Chevron are still his top five positions but I can always follow Berkshire to get notified if anything changes this is a total game changer for finding great comp companies and investing at great prices and if you sign up right now you can get up to 16 free stocks including a full share of Tesla or Google all you need to do is download the app using my link below keep your funds at that level for at least 60 days and enjoy your free stocks but this offer ends soon so make sure to get started today and a big thanks to MooMoo and to you for supporting the channel alright chip makers like Intel and AMD aren't the only companies trying to take market share away from nvidia's profitable AI Empire other trillion dollar Titans of Industry are trying to defend their own margins by designing their own ad chips as well let's start with Google weighing in at a 1.7 trillion dollar market cap Google search business stands to lose the most from this current generative AI Revolution to combat that Google unveiled an arsenal of AI offerings including their chat bot Bard their large language model Palm 2 and their fourth generation tensor Processing Unit or TPU earlier this year year Google estimated that over 90 percent of their AI training tasks used their own tpus and Google has multiple clusters of over 4 000 tpus that can be combined to create a super computer thanks to custom-made Optical switches that connect them all around the world that's important because large language models like Palm 2 and gpt4 have grown too big to be stored on single chips in fact Palm was trained across 8 000 tpus over a span of 50 days Google has been focused on advancing chip to Chip and computer to computer connections for years now as a result Google's Computing nodes can easily be reconfigured in real time to optimize performance avoid nodes that have issues or ones that are undergoing maintenance and it makes it easy for Google to upgrade this worldwide super Computing Network one piece at a time that's pretty incredible especially considering it's been operational since 2020 but when it comes to generative AI 2020 was basically the Stone Age Google's 4th generation tpus are more power efficient than nvidia's a100 chips but far less performant than nvidia's current h100s and while Google could announce a new TPU to compete with nvidia's h100s or their upcoming three nanometer Blackwell chips it might not happen so soon about a month ago at Google's Cloud next event they announced a big partnership with Nvidia in which Google Cloud will use nvidia's AI hardware and software stack to help power their next generation of AI applications we're going to put Nvidia dgx cloud in gcp dgx cloud is nvidia's AI supercomputer this is where we do our AI research this is where we optimize our incredible software stacks and our foundational platform vertex AI provides the foundation for how everyone consumes our generative AI llms it's being integrated and running on nvidia's latest h100 gpus and many of Google's own models and products also heavily rely on Nvidia we're not just offering customers we're using it internally in developing our foundational technology today we use Nvidia gpus to build our next Generation Ai and serve models for our customers and users but Google cloud is only the third biggest cloud infrastructure provider with about an 11 share of the market Microsoft's Azure has captured twice as much of the cloud market and AWS has captured almost three times what Google has so it's important to know whether they'll be partnering with Nvidia as well or trying to compete with them Microsoft has been reportedly working on its own five nanometer AI chip codenamed Athena to reduce its Reliance on nvidia's Hardware to train large language models they've actually been working on this chip in secret since 2019 and some Microsoft and openai employees are already testing how well these chips will work with gpt4 it's not clear if Microsoft will make their Athena chips available for Azure customers or just keep them in-house for big projects but either way I don't really think these chips will be direct replacements for nvidia's gpus instead they could act as more of a buffer for Microsoft's internal employees and maybe Azure clients who don't need the full power of Nvidia ships and can use a much less powerful but much more efficient alternative after all Nvidia and Microsoft also announced a huge partnership last year in which Azure will use tens of thousands of Nvidia gpus so however Microsoft's Athena chip ends up being used it's probably more complementary to nvidia's chips than a straight up competitor the same is true for Amazon Amazon web services launched their arm-based graviton server chips back in 2018 but those are designed to compete more with AMD and Intel CPUs not nvidia's gpus AWS does have two homegrown AI Focus chips today the first is called inference Asia which was launched in 2019 as the name implies the inferentia chips focus on AI inference so they really compete with lower cost High throughput chips on the other hand their trainium chips were released in 2021 and they're built to have a good price to Performance ratio for AI training in my opinion AWS isn't really trying to compete with Nvidia either as the biggest provider of cloud computing and AI infrastructure AWS provides access to many different kinds of machines for a wide variety of computational workloads the name of the game here is cost optimization so both AWS and their clients want to run their workloads on the best machine for the job without taking up any more resources than they need and when that job is training large machine learning models or building generative AI applications AWS has no problem relying on nvidia's h100 gpus if they're homegrown inferencia and trainium chips aren't the right tools for the job just for completeness AWS does already offer for instances powered by Intel's Gaudi 2 chips and are actively considering amd's Mi 300 accelerators as well if and when AWS does end up releasing their next generation of chips I think they'll stick to the same Playbook focusing on chips that save them power and money in specific areas where their current offerings are either overbooked or underutilized rather than trying to completely cut out a specific chip maker so Google Microsoft and Amazon all seem to be happy to keep partnering with Nvidia instead of trying to compete with them in the AI chip market and that makes sense because it's hard to justify any one cloud provider competing with all of nvidia's research and development for a product that will end up staying fully in-house after all it's not like Google Cloud will start using Amazon's inferentia or trainium chips anytime soon which really limits their total addressable Market compared to third-party chip makers that can go after every cloud service provider and third-party chip makers are actually coming out with competitive AI chips amd's Mi 300 systems and Intel's Gaudi 2 and Gaudi 3 chips could make their way into a lot of data centers in the near future the problem is that both the Mi 300 and the Gaudi 3 chips are still at least a few quarters away so for now we should expect Nvidia to keep having blowout quarters as they capture as much of the high performance generative AI Data Center Market as possible between their h100 gpus and advanced networking Technologies Cuda and their AI software libraries and long-standing Partnerships with every major cloud provider Nvidia is currently the only game in town but there's one more massive AI breakthrough that you need to know about so make sure to check out bol
Comments 0