When GPU leader Nvidia announced Project Denver — its ARM-based foray into the server CPU world — on Wednesday, the news sent shockwaves through the IT world. There had long been talk that Nvidia wanted into the server CPU business, but most industry watchers believed it would go the x86 route. Instead, it chose ARM, which arguably shakes up the market far more than just another x86 option would have.
Before delving into how Project Denver affects other vendors, it’s worth looking into why Nvidia chose to go this route. According to Director of Technical Marketing Nick Stam, with whom I spoke during a meeting at CES, legal issues were a big reason. Namely, Intel wouldn’t license the necessary x86 patents to Nvidia, and the company wasn’t too keen on buying an existing x86 licensee in order to obtain them. Enter ARM.
Its dominance in mobile devices means it is the processor of the future (Nvidia’s Tegra mobile processor, for example, is ARM-based), and ARM’s forthcoming Cortex A15 server design is mighty compelling from a power-performance standpoint. Planned server/desktop support from Microsoft and a number of Linux vendors is the very necessary icing on the cake. With the operating systems in place, Stam believes applications — the key to success for any IT platform — will follow. Additionally, Stam noted, Project Denver is targeting high-performance machines (from desktops to supercomputers) in which Nvidia’s Tesla GPU has already proven itself and will continue to handle more workloads; the CPU is playing second fiddle to the GPU.
So, who needs to figure out a game plan for combating the Nvidia processor once it’s available (likely in 2013, when Nvidia’s Tesla GPU reaches the “Maxwell” generation)?
It goes without saying that Intel and AMD have the most to lose from Nvidia’s planned server processor. Their x86 processors currently dominate in desktops, servers and supercomputers of all types. Nvidia, then will directly affect at least a portion of their businesses — especially those like gaming and HPC, where parallel processing reigns supreme. Further, both Intel and AMD also are pushing their own hybrid CPU-GPU chips (Sandy Bridge and Fusion, respectively), and an Nvidia GPU-CPU option could be a threat to those emerging businesses.
Of course, Intel and AMD can take care of themselves. As I noted in my recent “Infrastructure 2011: 5 Trends Not to Expect” piece, Intel, especially, has plenty of money to buy its way into the ARM space if it feels compelled to do so, either via acquisition or by becoming a licensee. Or, if it’s energy efficiency that makes ARM so popular, perhaps Intel will use its low-power Atom processors as the foundation for a new lineup of server chips. Then there is the overwhelming market dominance factor: Will server makers, ISVs and/or developers actually be willing to make the switch from their tried and true x86 processors?
As the startup that’s presently the face of ARM-based servers — and, in fact, ARM Holding’s chosen one — one might think Calxeda has a lot to fear from Nvidia moving in on its turf. But that might not be the case. Calxeda appears focused on delivering energy-efficient servers for web data centers, which don’t necessarily fall into Nvidia’s higher-performance scope. Plus, if Calxeda delivers a product this year, as expected, it will have almost two years to build momentum and customers before Project Denver becomes a product.
ARM-licensee Marvell also has plans to sell its own server processor based on ARM’s Cortex-A15 design, and it could be a real loser if Nvidia’s chip lives up to expectations. Not only is Marvell a new name in server processors, but (unlike Calxeda) it’s also without a server line in which to package its processors. However, assuming Marvell goes after relatively low-performance web applications and takes advantage of its time-to-market advantage over Nvidia (Marvell’s chips are slated to ship well before 2013), it might be able to carve out its own niche and leave the high-performance business for Nvidia. Cost also will be a consideration — it seems only logical that Nvidia’s GPU-CPU hybrid will cost more than Marvell’s CPU-only chip.
Tilera is an interesting case because its many-core processors are neither x86- nor ARM-based (although they are RISC-based), and, not being a server maker itself, it relies on finding OEM partners, just like Nvidia will. However, like Calxeda, Tilera isn’t necessarily targeting high-performance workloads with its data center ambitions and, in fact, is positioning itself as an ideal, low-power embedded processor. Where it might get stung, though, is in partnerships with server makers such as SGI, which presently incorporates Tilera processors into high-end systems with the purpose of offloading certain tasks to those processors to lower power consumption. The CPU component of Nvidia’s forthcoming chip will be low-power and possibly higher-performance thanks to its ARM foundation, but will come packaged with the Tesla GPU horsepower. That and the proven Nvidia brand name could make it a compelling addition to high-end systems, either in place of or in addition to x86 processors, to the detriment of Tilera.
SeaMicro, which sells an ultra-low-power server packed with 512 Intel Atom processors, could be out of luck if ARM catches on in servers, regardless who’s pushing it. If ARM’s Cortex-A15 design can deliver energy savings comparable to Atom while also delivering better performance — the Cortex-A15 is a quad-core System on a Chip, after all — without requiring 512 processors per server, then perhaps Atom, in its current state, won’t ever catch on in servers. Nvidia’s Project Denver just ups the ARM ante — there’s no way an Atom-only approach can touch ARM plus the already supercomputer-proven Tesla GPU in performance. Even if Nvidia has no plans to compete with SeaMicro directly, its ARM strategy validates the architecture and makes options like Calxeda and Marvell — which will compete with SeaMicro — look a lot better.