저자: Alex · 외국인
요약 보기
엔비디아는 GPU 시장 점유율(개별 GPU 83%)과 CUDA 생태계, 텐서 코어 및 AI 특화 칩 전략으로 컴퓨팅 시장을 지배하며, 데이터센터 매출이 게임 매출보다 2배 빠르게 성장해 향후 주력 매출원이 될 수 있다고 주장. AI 하드웨어 지출이 2030년 1.7조 달러 규모로 성장할 것으로 전망. 메타버스 관련 옴니버스 플랫폼과 슈퍼컴퓨터 시장 점유율(93%)도 강조.
다듬은 글
it's tough to think of a more important industry today than computing or when we depend on more in our daily lives microchips are used in countless applications from cell phones and solar panels to satellites and self-driving cars this year alone the global chip industry could bring in around 600 billion dollars and by 2030 spending on artificial intelligence ships alone could close in on 2 trillion dollars per year the competition for this market has been turned up to 11.
with multiple trillion dollar companies like amazon microsoft and google all starting to design their own chips but it could be that none of these tech giants will sit on the computing throne actually the number one contender weighs in at just under 600 billion dollars of market cap which is one quarter of the size of microsoft the company i'm talking about is nvidia and there are three key reasons why it might dominate this market over the next decade your time is valuable so let's get right into it nvidia is known for their graphics processing units or gpus when most people think of nvidia's graphics cards they probably think of nerds playing computer games that's for a pretty good reason as of late last year nvidia had a dominant 83 percent of the market share for discrete gpus with amd owning the other 17 that means that nvidia sells five out of every six graphics cards if you look at all gpus sold last year instead of just separate graphics cards amd and intel's chip shipments were both down by double digits due to a global chip shortage ouch but nvidia's overall shipments were up eight percent here's the thing even though nvidia is dominating that market it's not the only one they operate in over the past few years nvidia has branched out from gaming into areas like data science artificial intelligence modeling and simulation and even edge computing that last one is important because the smaller nvidia can make its chips the closer to the edge of the connected world they can get devices at the edge are much cooler than just cell phones they're things like drones and delivery robots and autonomous cars pretty much everything i talk about on this channel but you don't win the chip market by making the smallest chips you win it by making the best solutions to real problems for the best price and if you want to understand how nvidia is doing that you need to know what kinds of problems gpus solve that cpus aren't really built for in my eight years as a researcher at mit i worked on a lot of programs and prototypes that use gpus for a wide variety of tasks here are some of my biggest takeaways gpus are really good at something called linear algebra instead of working with single numbers this special area of math works with sets of numbers called vectors matrices and tensors if that sounds absolutely terrible to you that's because it is i still have nightmares about this stuff from grad school despite being the stuff of nightmares this special math is the best way to solve many similar problems in parallel for example if i'm playing a video game and i want to move around every surface of every object in the scene needs to adjust depending on my avatar's position and orientation instead of doing that math problem object by object for each player in a scene developers can just build digital environments then those graphics get loaded into a gpu that can translate rotate and scale everything depending on the positions and the pointing vectors that you give it for each player well it turns out that solving many similar problems in parallel this way is useful for way more than just gaming think about an autonomous vehicle that needs to process many cameras and radar feeds as it navigates through a real environment that's kind of like a video game except you have eyes all over the body and the danger is very real wait is this just an episode about all of my nightmares think about all the graphical applications this type of math can be used for drones in the sky telescopes trying to find a specific object facial recognition social media filters the list goes on and on here's the thing tensor math isn't just useful for graphics you can use this kind of math to solve many kinds of problems you can build a matrix of how genes express themselves in a wide variety of diseases for medical diagnostics or you can solve a bunch of valuation models at once for every stock on the market how about trying to solve the same hard cryptographic problem using many different guesses at the same time that's crypto mining what about scoring pieces of content for different users based on their interests and showing them whatever scores the highest in the matrix that's a content recommender system like netflix or spotify the company with the best parallel number crunching solution is going to win the lion's share of the computing market for all these different verticals that need this same type of system one big reason nvidia is outselling competitors like amd five to one is how they handle more complex and power-hungry computing tasks like ray tracing reflections and upscaling graphics at a high level there's a big fundamental difference in how nvidia and amd build their chips amd puts a physical ray accelerator onto each core of their gpus it's kind of like a supercharger but for computer chips instead of cars i'm just kidding i don't know anything about cars nvidia's specialized ray tracing cores are physically separate from their other gpu cores nvidia has another special type of core called the tensor core because they focus on doing this tensor math very very fast at the cost of accuracy then nvidia fixes those inaccuracies separately using deep learning super sampling or dlss which is a major selling point for their products for dlss they train an artificial intelligence model on high resolution scans of every video game then apply that ai to adjust the output of the tensor cores to a player's version of that same game nvidia's secret sauce is pushing the right limits of what can be done with specialized hardware while offloading as much work as possible to specialized software then using ai to optimize the output of both which is useful everywhere from video games to real-life camera data to simulating physics experiments to ai applications like speech and object recognition nvidia also has cuda which is their coding language and development platform cuda is how scientists and researchers can use nvidia's gpus for their own applications like my projects back at mit cuda is a massive ecosystem with a huge community of developers kind of like apple swift for ios apps new projects and new developers tend to flock to where there's already an established developer community well cuda is designed specifically for nvidia devices and it's up to other device makers like amd to make it work with their chips if you're finding value in this research so far let me know in the comments below i think you'll also find a lot of value by going to public.com ticker symbol u and claiming a free slice of stock worth up to 70 when you fund your account i'll leave a link to that exclusive offer for you in the description below nvidia's specialized tensor computing chips ai enhanced computing software and programmer ecosystem are why they dominate the gpu market but they're also the fundamental building blocks for winning the enterprise high performance computing market this is a tough capability to try and build yourself so it makes sense for a business to connect to a pre-built scalable solution and customize it to their needs for years nvidia has been designing its chips for cloud-based artificial intelligence applications knowing that this demand was coming today nvidia's chips sit inside the servers of almost every cloud computing provider out there including ibm oracle amazon web services microsoft azure and google cloud even though amazon and google both make their own ai chips i'm pretty sure this played a big role in why regulators didn't allow nvidia's acquisition of arm to go through then nvidia would be the dominant player in gpus and cpus in servers desktops mobile devices and a ton of other different kinds of companies would end up relying on their support speaking of which remember the whole specialized hardware thing well while other companies are making generalized cloud ai chips nvidia is building more specialized chips and software suites for a wide variety of applications like networked supercomputing xero trust cyber security edge ai medical imaging and diagnostics and processing for robotics applications nvidia is building maxine which connects computer vision natural language processing and avatar animation into a real-time conversational robot these are three very different ai applications coming together and interacting in a way that behaves more like a human which can be used as a talking kiosk in malls and restaurants a concierge in hotels and robo taxis or a simultaneous multi-language translator for video conferences this is really next generation stuff and it's powered by artificial intelligence there's actually one more initiative i want to talk about before covering nvidia's financials and that's the omniverse no not the metaverse people actually like this one the omniverse is a virtual platform built for real-time photo realistic simulation architects and engineers often create and work on 3d models of their projects manufacturing companies want to see the real-time status of their factory floors and equipment scientists want to visualize the data they capture whether it's radio waves or weather patterns or something else entirely so nvidia's omniverse platform what's creators designers researchers and engineers connect their existing asset libraries tools and projects so that they can collaborate in a shared space and interact with digital twins of their projects on top of that software providers and developers can build and sell extensions apps and microservices for the omniverse to expand its functionality so knowledge workers across different companies don't always have to reinvent the wheel here's the thing when many different industries can share resources like this the creator economy booms ecosystems like the omniverse will let independent developers and micro studios flourish because they can focus on building one great module or widget without having to build the entire infrastructure needed to support it if the omniverse platform gets big enough there could be a real customer base for this kind of work encouraging more small developers to join the marketplace which kicks off a virtuous cycle of more ready to use assets bringing in more customers which brings in more developers which brings in more assets and so on i really think that the omniverse could end up being something like apple's app store but for enterprise businesses now that we understand the absolute basics of what this incredibly advanced vertically integrated high performance computing company does we can have a productive conversation about its market its finances and its future as a tech nerd and an investor i think this is what really matters today over ninety percent of the total computing done on most big hybrid supercomputers is done on the gpu side and that number could climb to as high as 95 percent in the near future gpus really put in the work of the 146 supercomputer systems surveyed 136 use nvidia's gpus that's a 93 market share of super computers for nvidia and that number has been trending up over time research suggests that by 2030 gpus could be the dominant kind of processors for most new workloads growing at a 21 compound annual growth rate over the next eight years as these kinds of capabilities get cheaper and get adopted on smaller and smaller systems as ai systems get cheaper to build and train more companies will use them in more applications as a result projected hardware spend on ai will grow by a hundred fold by 2030 from 17 billion a year in 2021 to 1.7 trillion dollars that's a 67 compound annual growth rate over the next nine years meaning this market is going to expand a lot for the rest of the decade as a result of that 1.7 trillion dollar spend by 2030 seven trillion dollars in enterprise value could be created in the ai computing space which would generate returns of 57 at a compound annual growth rate during the next nine years of course not all of that belongs to nvidia but we should expect nvidia to keep achieving exponential growth by growing the adoption of their ai hardware their services and their applications over the next decade inside this quickly growing market if you're anything like me you raise an eyebrow when you see such huge projections for growth like this so to see if their story checks out let's take a look at nvidia's latest earnings presentation which came out in the middle of february nvidia's gaming revenue which comes from their desktop and laptop geforce gpus is up 37 year over year look how stable this revenue growth has been even with chip shortages messing up the entire market data center growth is even more extreme showing an 11 quarter over quarter growth or a whopping 71 percent year-over-year growth that was led by compute products and strong demand for nvidia ai in applications like natural language processing deep learning recommendation systems and cloud computing exactly the type of stuff we learned about throughout this episode and that seems to really line up with the crazy 67 compound annual growth rate in ai hardware spend that i just talked about nvidia's data center revenue now makes up 40 of nvidia's total revenue while gaming makes up 46 but their data center revenue is growing twice as fast so i expect we could see these number flip by the next fiscal year and nvidia's primary source of revenue could become their data center solutions let me point out three quick things about the metaverse here since you really can't talk about the metaverse without mentioning nvidia these days first off the metaverse needs nvidia but nvidia does not need the metaverse today nvidia's technologies are found in nearly every supercomputer and with every major cloud provider and they have millions of registered developers around the world nvidia is growing at such a fast rate today because it's solving the computing problems of today and one of those computing problems is for the metaverse meta platforms unveiled its new ai supercomputer called research supercluster which is powered by nvidia's technologies later this year they're going to expand that supercomputer to include 21 times more gpus something that can only happen because nvidia builds its chips with cloud computing in mind and third even if you think the idea of the metaverse is stupid building a connected enterprise marketplace for shared assets tools and research is certainly not the omniverse solves real problems today and allows talented individuals and small teams to operate in a market usually reserved for enterprises it's pretty cool to see a free version of the omniverse being announced for individuals and i'm excited to dive into it and share what i learn after all this is the channel that invests in you if you want to learn more about the advanced technologies driving our future and the companies behind them this video right here is for you either way stay long stay strong symbol yo
Comments 0