Intel (INTC) teased its next generation "Knights Landing" chip at the International Supercomputing Conference. For those of you unfamiliar with the high performance computing wars (I don't blame you - the investment world is gaga over battery powered mobile chips), many supercomputers use GPUs from Nvidia (NVDA) to accelerate highly parallel portions of a given workload. See, GPUs are inherently suited for ultra-parallel workloads given that a GPU is - put very simply - an array of smaller computing units that, when aggregated, can deliver incredible amounts of power-efficient throughput. These accelerators are often paired with standard server CPUs (usually Intel Xeon processors) and the combined system ends up offering an insane amount of computing power. This is a high growth market, and Intel is the leader on the CPU side, while Nvidia is the leader on the co-processor side of things with its "Tesla" line of products.
A Tale Of Two Giants
Intel, however, in a bid to own anything and everything that goes into a data-center, has been developing its own co-processors known as "Xeon Phi." While Nvidia's parts are enhanced/less disabled versions of its highest end PC GPUs, Intel's parts are aimed solely at HPC and are in fact made up of a bunch of small Intel Architecture based cores married to a wide vector unit. Nvidia's advantage here is that it can leverage the same R&D and silicon across multiple areas, but the disadvantage is that Intel does not include the "graphics only" transistors that are irrelevant to high performance computing, which allows the firm to get away with fewer transistors on what is usually a generation ahead process.
At the International Supercomputing Conference, Intel fired a rather potent shot at Nvidia's dominance with the announcement of the first details of its next generation Xeon Phi code-named "Knights Landing":
What exactly does this one slide portend for the future of this space?
Intel Leverages Silicon, Manufacturing Advantages; Nvidia Has Experience, Architecture
So, as expected, "Knights Landing" is built on Intel's 14nm 2nd generation FinFET process, while Nvidia's next generation Tesla will be built on TSMC's (TSM) 20nm HKMG planar process. In terms of die sizes/budgets, Intel's engineers likely have significantly more to work with than the Nvidia engineers on a per unit of area basis as Intel's transistors will be at least a generation ahead of the TSMC ones in this space.
Normally, I would probably say, "well, game over for Nvidia" as it is behind on process technology, but I have always maintained that process merely enables micro-architecture. The current 22nm Xeon Phi, while formidable, still falls behind Nvidia's K20 and K20x built on TSMC's 28nm process. This is an example of where Nvidia's years of experience in the GPGPU space allowed it to design what appears to be a stronger micro-architecture than Intel despite the transistor handicap. While I do expect Intel to accelerate here (particularly as the current Xeon Phi was Intel's first attempt), my conversations with Intel engineers have indicated that even with these structural advantages, Nvidia still manages to do quite a few things better on the micro-architecture side.
However, it is tough to ignore that Intel will be offering standalone Xeon Phi chips (which eliminates the PCIe overhead), and that Intel will be outfitting these parts with likely gigantic amounts of its own home-grown eDRAM (designed for Intel's "Iris" graphics for notebooks). Nvidia has indicated that with its post-Maxwell generation of GPU named "Volta" it would be stacking DRAM on package (but likely not a custom designed eDRAM), which should also be interesting:
In addition, Nvidia's Chief Scientist, Bill Daly, mentioned at a conference some months ago that Nvidia, too, would be integrating the host processor on-die with the GPGPU accelerator, so I expect that Intel's standalone Xeon Phi will be going toe-to-toe with a next generation Tesla part. If you've ever wondered what Nvidia's real goal for its own custom 64-bit ARM (ARMH) processor known as "Project Denver" was, it is likely this integration of a giant GPU with a few of these cores to drive the GPU.
In addition, if you've been paying attention to the industry, Intel has gobbled up all sorts of interconnect technology from the likes of QLogic (QLGC), and Cray (CRAY) in order to bolster its position in the high performance computing space and to eventually integrate such interconnects onto future Xeon/Xeon Phi processors:
Nvidia, a much smaller firm, has been (not so) quietly getting acquainted with Mellanox (MLNX), a high performance computing interconnect specialist that, too, faces the problem of Intel eventually integrating everything it needs onto one chip, thus leaving Mellanox out in the cold. Nvidia is currently a leader in HPC accelerators, so it has the credibility and likely the ability to develop a CPU platform that could see commercial success in the HPC space. Nvidia's "Denver" coupled with the latest GPGPU technology and interconnect/photonics technology from Mellanox could keep Nvidia thriving in this game against the computing world's Goliath.
Who Will "Win"?
My expectation is that Intel and Nvidia end up in heated competition in this space for years to come, with both firms playing to their individual strengths and trying to exploit the weaknesses of the other. While the initial knee-jerk reaction is to assume that Intel's advantages will overpower Nvidia in much the same way that Intel overpowered AMD (AMD) in the PC space, I believe that given Nvidia's excellent management as well as its many years of experience to be a valuable asset to leverage going forward. Nvidia fundamentally does one thing: build GPUs, and a laser focused, well-capitalized, and well-run company like Nvidia will likely be able to keep pace, although Intel is and will continue to be a very formidable foe.
At any rate, it should be very interesting to watch.
Additional disclosure: I am short ARMH