Nvidia CEO outs GPU roadmap
'Perf per watt equals perf'
GTC Nvidia president and CEO Jen-Hsun Huang has outed the company's GPU roadmap.
"For the very first time in the history of our company," Huang during his Tuesday keynote at the company's GPU Technical Conference in San Jose, California, "we are going to tell you the code names and the progression of our next several generations of processors."
Although Nvidia's senior Tesla product manager Sumit Gupta told The Reg earlier that his company wouldn't be outlining its product roadmap at the conference, he was merely trying not to one-up his boss.
Code names and dates, however, are pretty much all that Huang revealed, aside from relative performance plans. Deep details of the next two generations — "Kepler", due next year; and "Maxwell", due in 2013 — will have to wait until a later date.
Still, what Huang did announce will be of interest to the growing CUDA cadre or, for that matter, to anyone interested in the future of what Huang described as "parallel computing, GPU computing, accelerated computing, heterogeneous computing — however you guys want to describe it."
In describing the performance of Kepler and Maxwell, Huang didn't talk mere gigaflops. Instead, he used the metric of dual-precision flops per watt, with Nvidia's Tesla GPU as the baseline. His reasoning basically boiled down to the fact that: "In a constant power-delivery environment ... 'perf per watt' equals 'perf'."
"In the case of performance, Fermi is about one and a half double-precision gigaflops per watt," Huang said. "We expect Kepler to be somewhere between three to four times the performance per watt of Fermi."
Kepler's performance-per-watt advance is well underway, according to Huang. "Kepler is based on 28 nanometers. It's scheduled for production later next year. The design is progressing very rapidly. There are hundreds of engineers working on it. By the time we are done with the Kepler family, we will probably have invested a couple of billion dollars on R&D for it."
In 2013, "We'll have Maxwell," he said. "Maxwell is going to bring with it yet again a big step up in performance."
If Tesla is regraded as a baseline, Huang said, "Maxwell is going to be 16 times relative to [Tesla] in just a couple more years. "In just a few more years, we're going to see a sixteen-times improvement in performance for your parallel-computing applications."
Huang also emphasized that Nvidia's future is, in his view, more than a mere matter of gigaflops per watt. "Now performance isn't the only thing that we will bring. Each generation will bring architectural ideas and features — like we did this time," he said. "This time we brought you ECC so that we can deploy, in a large-scale way, GPUs in server groups."
His reasoning behind the importance of ECC was simplicity itself: "If you need to run a simulation for a week, the last thing that you want to know is that somehow one of the GPUs created the wrong answer along the way."
His company has more non-gigaflopian improvements on the way, Huang said: "So all the way between now and the Maxwell time frame, we're introducing new features that you've been asking for — things like preemption, things like virtual memory. We're also going to continuously enhance the GPU's ability to autonomously process, so that it's non-blocking of the CPU, not waiting for the CPU, relies less on the transfer overheads that we see today."
Huang also promised that Nvidia's roadmap will bring "a very large speedup in performance." Remember, though, that performance in Huang-speak isn't defined by gigaflops alone. "'Perf per watt'," after all, "equals 'perf'." ®