Intel Makes Major Moves in Graphics and Learning
THERE HAS BEEN A SERIOUS SHAKEUP in the world of graphics technology this past month, with three major announcements. First, Intel announced a partnership with AMD, where it will integrate a custom Radeon graphics solution on a future Intel Core CPU. It w
And this isn’t a low-end graphics solution. Current indications are that the top configuration will have 1,536 AMD GCN cores, clocked at 1,000–1,150MHz. That’s somewhat slower than an RX 570, but for a laptop, it’s a huge step up from Intel’s previous HD Graphics and Iris solutions. More interesting is that the solution will also utilize HBM2 memory, potentially 2–4GB, which will provide significant VRAM bandwidth advantages compared to previous integrated solutions. The only real question remaining is what sort of power requirements the chips will have, and how much the laptops will cost.
That’s a huge announcement, but it’s only the beginning. Days later, it was revealed that AMD’s Raja Koduri, former head of the Radeon Technology Group, was leaving the company. He will now be the chief architect and senior VP of Intel’s new Core and Visual Computing Group. Raja says he’s “incredibly excited to join the Intel team and have the opportunity to drive a unified architecture vision across its world-leading IP portfolio that helps accelerate the data revolution.” That sounds as though Intel wants to use his talent to do more than just faster graphics, which brings us to the third piece of news.
Intel already has products that compete in the machine learning and supercomputing arena, specifically its Xeon Phi processors. These are found in many of the world’s fastest supercomputers, including China’s Tianhe-2, which currently occupies the number two position on the Top 500. During the past year, Intel also acquired Nervana Systems, a company working on a custom ASIC for machine learning applications. Intel’s upcoming Nervana NNP (Neural Network Processor) will be focused on driving performance even higher, though Intel notes that standard performance metrics for deep learning don’t tell the whole story, and has thus far not revealed any specific performance figures.
The performance expectations for the Nervana NNP are high, and the retail products were slated to begin shipping by the end of 2017. Early hardware already exists, with Intel working to tune performance, and the results must be promising, because Intel also announced that Knights Hill, the 10nm successor to the current 14nm Xeon Phi line (aka Knights Landing), has been canceled. Knights Hill was slated for the Department of Energy’s Aurora supercomputer, with a performance target of 180 PetaFLOPS, but the contract has been rewritten with a goal of being one of the first Exascale systems (capable of at least one ExaFLOPS) in 2021. Whether that will come via Nervana or some other product isn’t clear, but it seems likely.
Given the growth in machine learning and supercomputing applications, where graphics technologies such as Nvidia’s Tesla solutions are major competitors, part of Raja Kudori’s task may be to unify the various solutions that Intel already offers, helping to come up with fewer products that deliver higher performance. Or perhaps Raja won’t be doing much with the Nervana or Xeon Phi products, and will instead work solely to improve Intel’s graphics IP. Regardless, Intel has made a lot of waves, and we look forward to seeing where this leads.
Raja Kudori’s task may be to come up with fewer products delivering higher performance.