This week is NVIDIA’s GTC, or GPU Technology Conference, and they likely should have changed the name to ATC because this year – it is all about AI.
To say this show was rich in big announcements and technology would be an understatement. The advances in autonomous cars, robotics, graphics tools (the Adobe fireside chat was amazing), and conversational AI were amazing.
But what stood out was a new AI-focused ARM-based CPU called Grace, for Admiral Grace Hopper (I once took a computing class from her), one of the most amazing people I’ve ever met. And this part seems to live up to Grace’s name because it could revolutionize HPC and Supercomputers.
It is slated to be the core technology in the future Swiss National Supercomputing Centre’s new ALPS supercomputer, which is expected to exceed today’s fastest supercomputer by 10x when it comes online in 2023. But they also indicated that a variant of this technology would target PCs, effectively throwing down the gauntlet to the x86 folks.
Let’s talk about the Grace CPU this week and how it could change the HPC, Supercomputer, and PC landscape.
ARM vs. X86
Generally, when you have a war, you fight over real estate. In technology wars, you fight over customers, and the war between X86 and ARM has been going on for some time. The initial battlefield was the smartphone market, and X86 was forced out of that segment mainly due to not approaching it soon enough with a ground-up redesign focused on battery life rather than performance.
ARM and X86 coexist in tablets, but ARM generally has a higher volume, thanks mainly to the iPod and Amazon Fire Tablets. In PCs, X86 is dominant, but ARM, thanks to Qualcomm and Microsoft, has been making inroads, and in Chromebooks, ARM has also found considerable success tough X86 plays there as well.
Much like smartphones are ARM’s safe space, servers were the same for X86 as ARM solutions essentially hit that opportunity and bounced. For ARM to penetrate, they needed one of the big players (HPE, Dell, IBM, Lenovo) to take up the technology aggressively, and they didn’t. But should one of the X86 companies in servers move on this opportunity, it could be a game-changer. Initially, AMD tried to do this through acquisition, but AMD had primarily been forced out of the data center and didn’t have the clout needed to succeed.
But NVIDIA is well-positioned in both HPC (High-Performance Computing) and AI. They are well positioned in the data center and very credible when talking about extreme computing performance. It’s their thing.
They have brought forward a tuned solution that, on paper, better integrates with their enterprise GPU solutions to provide a massive performance bump. It does this by avoiding the PCI Express bottleneck inherent in earlier solutions with a 30x improvement in bandwidth.
This effort is on top of the work they are doing with AWS to increase the performance of their Gravitron II part and Marvell to created an SDK for hyper-converged edge servers. Oh, and they also have a somewhat related project with MediaTek for high-performance PCs and Notebooks.
Prepare For Disruption
As we come out of the pandemic, enterprises and cloud providers are rethinking their computing infrastructure to deal with the new Work From Home reality. We were already facing pretty significant disruption as data centers evolved to embrace high-performance edge computing.
This change will already drive many new server purchases from buyers who will want to push the envelope on performance to assure their installations don’t become prematurely obsolete. This period is ideal for introducing a new HPC CPU because buyers are more open to change and new architectures to deal with the new normal. And that new normal is heavily AI-focused.
If this part performs as advertised, it could really shake up the HPC and Supercomputer landscape, and once validated there, moving through the cloud and even into the PC market should be a relative piece of cake. Or, put another way, if NVIDIA ARM takes the cloud as we shift more and more workloads to the cloud, it could become the dominant technology not only in the data center but on the desktop.
The potential for disruption is off the chart, but neither Intel nor AMD are likely to take this challenge sitting down, so the only inevitable outcome is we are about to see a race for more performance we haven’t seen since the birth of X86.
But both AMD and Intel have advanced GPUs now, and both are capable of blending their technologies to come up with higher performance parts; they’ll have to execute fast or watch NVIDIA take over the data center market. All three companies are unusually well run when making this anyone’s race, but NVIDIA the company to watch because they are driving the change.