NVIDIA has had its share of setbacks over the past year. Renewed competition from AMD and a tough economic climate have conspired to knock the company’s revenues (not to mention its stock) down a peg or two. And in the long term, as more client computing migrates to mobile platforms, NVIDIA’s strength in high-end discrete graphics chips will become less sustaining.
But its GPU computing product set — CUDA, Tesla, et al. — continue to be ahead of the curve. Being the industry leader in this area, NVIDIA certainly recognizes the importance of this technology. In a recent article in TechRadar, NVIDIA PR Director Derek Perez makes the case that GPU computing and its offspring, visual computing, are part of a “sea change” in the IT landscape and his company is in a great position to cash in on this.
It’s hard to disagree with that assessment. In the HPC space, GPU computing is the most compelling technology to come on the scene in recent memory, and NVIDIA has jumped out to an early lead. Because of CUDA and some hardware innovations, NVIDIA is probably at least a year ahead of AMD and two years ahead of Intel (Larrabee) on this front. Soon, for example, you’ll be able to buy a Tesla-equipped Lenovo workstation for between $2,000 to $3,000. That machine will deliver in the neighborhood of one teraflop of single-precision floating point performance, and be able to run real live HPC workloads.
All the early adopters of GPU computing that I’ve spoken with are using NVIDIA technology today. Even though production deployment is still pretty rare, I expect a lot of GPU computing experiments to migrate into commercial use over the next year or two. Now with Intel’s Larrabee processor expecting to show up in 2010, and OpenCL ready to make GPU computing platform agnostic, NVIDIA needs to find a way to keep its edge.
At least on the client side of the GPU computing business, one visible threat to NVIDIA-style GPU computing is the CPU-GPU hybrid processor model envisioned by both Intel and AMD. Although the first products won’t have an HPC play, all signs are pointing toward an eventual convergence of CPUs and GPUs below the chipset level. If this turns out to be the case, future general-purpose microprocessors for high-end computing may have GPU-like smarts already built in, making discrete parts somewhat redundant.
That’s why it was not surprising to hear NVIDIA SVP of investor relations Michael Hara talk about the company’s plans for an x86 play in the next two or three years — a topic that I wrote about after Hara made these comments at Morgan Stanley’s Technology Conference earlier this month. But in the TechRadar article this week, NVIDIA’s Perez contradicted Hara’s earlier statement saying “We didn’t need x86 for our first 15 years and we won’t need it for our next 15.” Considering that 15 years is a lifetime for a company in the chip making business, that statement is probably not worth parsing too closely. CPU-GPU convergence will force NVIDIA to make some sort of accommodation for CPUs, even if they’re not strictly x86 cores.
The real danger is to believe that discrete GPU acceleration is the only way forward. It would be especially dangerous for NVIDIA to depend upon its competitors — Intel and AMD — to continue to deliver CPUs that fit neatly into this model. NVIDIA may be in a great position today to deliver GPU computing to the market, but it’s way too early to hang up the “Mission Accomplished” sign.