NVIDIA debuted its much-talked-about Kepler GPU this week, promising much better performance and energy efficiency than its previous generation Fermi-based products. The first offerings are mid-range graphics cards targeted at the heart of the desktop and notebook market, but the more powerful second-generation Kepler GPU for high performance computing is already in the pipeline.
The two new products introduced this week, the GeForce GTX 680 for desktop systems and the GeForce 600M GPUs for notebooks, are twice as energy efficient as their Fermi-based counterparts, according to NVIDIA. And although they represent more powerful graphics processors than the previous generation, the overriding design theme of the new architecture is performance per watt, rather than performance per square millimeter. According to Sumit Gupta, NVIDIA’s senior director of the Tesla GPU Computing business unit, that’s a fundamental change in the company’s architectural strategy. “This is the first time that power is a higher order concern than area,” he says.
That’s because, like nearly every chipmaker on the planet, NVIDIA’s fastest growing market segment is the mobile and notebook/ultrabook space. This architectural emphasis on energy efficiency also dovetails rather nicely with the GPU computing market, where power consumption is also a huge factor. That’s especially true for the Tesla GPU parts that end up in energy-sucking HPC servers. “Every market we’re in has become power sensitive,” says Gupta.
Upping the power efficiency in Kepler relied heavily on a tried-and-true technique, namely increasing the core count while lowering the clock speed. But the architecture is somewhat different. Underneath the covers, the cores are collected into what NVIDIA calls their Streaming Multiprocessors (SMs). In the Fermi version there were only 32 cores per SM. In the Kepler implementation, they reduced the control logic disproportionally and were able to squeeze in 192.
Boosting the core numbers was a no-brainer, given they were moving from the 40nm process technology with Fermi, to the 28nm node for Kepler. In the case of the GeForce GTX 680, for example, there are 1536 cores — three times as many as in the high-end Fermi GPUs, which topped out at 512 cores. At the same time they reduced the clock frequency from 1.5 GHz on the Fermi chip to just a shade over 1 GHz. Although each core is now doing less work, because there are more of them, throughput increases and does so with lower energy consumption.
CPU chipmakers have employed this strategy as well. But because of the greater complexity of the individual CPU cores and their reliance on limited memory bandwidth, core count increases are starting to stagnate (no CPU make ever tripled core count in one generation). Also, since a lot of applications are dependent on single-threaded performance, CPU chip makers try to hold the line on clock speed as much as possible. Ratcheting down the clock speed by a third, as NVIDIA has done here, is unheard for a CPU product.
For Kepler, NVIDIA is claiming a doubling of performance per watt compared to the Fermi-generation GeForce GTX 580. For real gaming applications, the new Kepler products are getting between 1.1 to and 2 times better the performance per watt. In some cases though, it can do even better.
For example, NVIDIA used their Samaritan demo, which illustrates photorealistic gaming, to show a 3X performance boost. Up until this week, that demo required three GeForce GTX 580 cards, drawing a total of 732 watts. It can now be run with a single 195-watt GeForce GTX 680.
To support all the extra throughput, memory bandwidth has been kicked up significantly. The interface on the GTX 680 supports 6.0 Gbps, which is 50 percent more than the 4.0 Gbps available on the GTX 580. According to Gupta, that’s the highest memory bandwidth for any commodity-based chip, NVIDIA or otherwise.
All of these architectural changes — more cores, slower clocks, and more memory bandwidth — will carry over into the second version of the Kepler GPU, a higher-end design which will be aimed primarily at GPU computing applications. This is the one the next-generation Tesla products will be based upon, and the one that will initially end up in two of the most powerful supercomputers in the world: Blue Waters at NCSA and Titan at ORNL.
According to Gupta, the second Kepler implementation will include a lot of capability not present in these first gaming-oriented products. In particular, it will have a lot more double-precision capability (which is not required for most graphics applications) and include new compute-specific features. And of course the raw power of these chips will be quite a bit higher than the mid-range graphics version introduced this week.
Although the company is not yet giving any of the speeds and feeds on the second Kepler, one would expect the core count and peak double precision performance to be two to three times higher, and memory bandwidth to get at least a 50 percent bump. Clock speed will almost certainly be whittled down from the current 1.3 GHz on the Tesla M2090, but perhaps not so aggressively as in these first Kepler gaming parts.
Presumably, the NVIDIA will stick with its 225 watt power envelope for the Tesla lineup, so the engineers just have to balance the core count and clock to land on that thermal design point. Given that power ceiling and the core count increase, NVIDIA should be able to deliver a Tesla GPU with between 1.3 and 1.5 teraflops of double precision performance. On the other hand, there is probably a case to be made to also offer less performant parts that consume less power.
In any case we’ll know soon enough. NVIDIA will probably do their paper launch of the HPC Kepler at the company’s GPU Technology Conference in May. And according to Gupta, the company is on track to put this version into production in Q4. If that goes according to plan, the new Kepler GPUs will be up and running on supercomputers before the end of the year.