Nvidia enters the SERVER arena

Yes, I am aware ARM licenses its processor cores and doesn’t produce them.

As you know Nvidia is a GPU producer mainly, graphics but also, very important, AI hardware accelerators. Acquiring ARM, like they did today is a very important move. You know, for the upcoming Nvidia Ampere AI accelerator product line, they needed the best server CPU and normally they rely on Intel Xeon. I think for 15 years or so. Because AMD is a direct competitor for Nvidia they'd prefer not to design in a competitor. Unless they have to. And they had to in 2020.

Why? Because the Xeon has PCIe gen 3, AMD Epyc Rome has PCIe gen 4 (double the bandwidth of PCIe gen 3). And Rome has 64 cores versus 56 for Xeon. Performance per Watt per core is better or slightly under the Xeon depending on the benchmark (every vendor picks only favorable benchmarks). And Epyc is better priced than Xeon.

Essentially, today. Nvidia has a path forward for making ARM based servers (ARM is number one now in the top-500 HPC list, Ampere the startup has an 80-core ARM server chip, …). They will focus everything on ARM servers. And take on Intel and AMD in the top market where the big bucks are made.

Remember, in 2018, Intel had 99% market share in the server CPU market! But the market changed fast, the need for supercomputers with a discrete CPU-GPU combo brought AMD back center stage. Today, Intel’s DoE supercomputer contract is late, due to their discrete GPU development (they have none, they announced Xe). Yes, they always had embedded graphics and weren't interested in a discrete GPU chip. Nvidia took care of that. Live and let live. But AMD has landed a few exa-scale contracts in 2019/2020 for their upcoming zen-4 CPU and their corresponding GPU. Today GPU hardware acceleration is required for AI workloads. AMD's Infinity fabric CPU-GPU planned cache coherence (zen-4 probably) will avoid the time consuming CPU copies to DDR and GPU fetches from DDR, the memory will be in the fabric and kept coherent. That is going to impress and thus both Intel and Nvidia had to do something. Meanwhile, AMD made Intel’s market share drop below 90% in 2020 in the server space. And Nvidia knows it missed out on exa-scale because of lack of their own CPU. So they put $40bn on the table. They get a cpu and all info on very important competitors licensing ARM cores. BOOM!

So, recap of 2018: intel was alone, AMD was their biggest competitor in the server space with less than 1% market share. 2020, just two years later, they have two competitors that are extremely motivated to fight for the high margin server market. And one that has double digit market share and a top CPU that has more cores, better off-chip bandwidth, better price and better performance per watt per core. AMD is also back in the GPU game, the upcoming Christmas season will feature their GPU in both Sony Playstation 5 and Xbox Series X. Not Nvidia, AMD. And it is true that it is low margin, big volume business (game consoles) but it is important for visibility and marketing. Now, two years ago, Nvidia and Intel mocked AMD when they were rising fast. But, Nvidia (led by an engineer!) quickly saw the threat and they adapted as fast as they could. Ampere is novel because they jump from 14nm VLSI technology to TSMC 7nm (N7). This is where Apple and AMD are for their chips. Those nodes are denser and better for performance per Watt. And today, they bought ARM and put themselves down as a serious contender. Intel is led by the former CFO of Intel. The problems with their tech nodes kept lingering, the Nervana debacle didn't get much attention (completely replaced by Habana Labs) and the lack of a discrete GPU are now public knowledge. A good thing, we can now voice these facts, we all know the ominous power of a giant protecting its turf.

Conclusion? The big business is the cloud, data-centers and private server farms. FANGMAN, a term from investors, is an acronym for Facebook, Amazon, Apple, Alphabet (Google), Nvidia, Microsoft, ... The big tech names. They are now so cash rich, they can do their own chips. Another threat to the long standing semiconductor companies, they actually are doing that. Apple silicon is hot. But they are at it for a long time. Few people realize how many things are changing at the same time. Things we have never seen before, like Facebook, a pure software social media platform doing hardware design. Whaaaaat? Yes, they are doing chips, they hire ASIC designers and pay them more than what the semiconductor companies are paying. Simply because they can.

Anyway, the main point I'm working towards is this: guess which core most of them license? Hmm, let's see. Amazon Graviton 2? ARM based, interesting. Google’s android ecosystem? Entirely relies on ARM. Oh, really? WWDC2020? Apple kicked out Intel in favor of ARM for Mac. They filed patents to make a kind of LEGO interposer concept to connect ARM chiplets with their hardware accelerators (Siri NLP, Image classification, GPU, face recognition, ...). Their ecosystem will unify around ARM so that the Mac app store gets a kick in the butt. Currently, developers need to make an x86 variant of the iOS apps. Yeah, they don't. So, Mac app store? Dead as dead can be. Native ARM apps will solve that. And Nvidia is now in control! Talking about turning the table!

Now, all stakeholders, especially the trillion dollar club, is not going to stand by and watch. Anti-trust and other stuff will be weaponized. Nvidia acquires tons of crucial competitor info and has control over their license fees. Apple has fought hard against the 5G patent monopoly (Qualcomm), they will need to assure control of their financial future (cost of ARM licenses). So, interesting to see how it will play out, will they do what I think they will do?