NVIDIA Unveils 1.3 Teraflop GPU for Supercomputing
The battle of teraflop accelerators began today as NVIDIA launched a new family of supercomputing GPUs based on the Kepler architecture. The Tesla K20 and the K20X represent the company’s latest and greatest and are intended to keep NVIDIA’s successful HPC accelerator franchise out in front of the competition. The chipmaker announced the new hardware as the 2012 Supercomputing Conference, in Salt Lake City, got underway.
NVIDIA has been the dominant provider of HPC accelerators for the world’s supercomputers. But with the imminent release of Intel’s Xeon Phi, NVIDIA will have its first serious competitor since it started shipping Tesla products in 2006. AMD is also announcing it’s new teraflop-plus FirePro offering for servers today. For now though, the new K20 and K20X look like unbeatable, at least from a pure FLOPS per chip perspective.
The top-of-the-line K20X offers 1.31 peak teraflops of double precision floating point performance and 3.95 teraflops single precision. That is twice the double precision performance and three times the single precision performance of the fastest Fermi-generation Tesla, the M2090. The chip itself encapsulates 2,688 cores and runs and runs at 732 MHz. Maximum power draw is 235 watts, which is about 10 watts above a standard Fermi part. Memory bandwidth on the K20X has been kicked up by about 40 percent compared to the M2090 — from 177 to 250 GB/second, although memory capacity remains steady at 6 GB.
It is the K20X that powers the new 27-petaflop (peak) Titan machine at Oak Ridge National Laboratory. And thanks to much improved Linpack yield on the new GPUs, Titan has also become the number one system on the TOP500, with a mark of 17.6 petaflops, which knocks Lawrence Livermore Lab’s 16.3-petaflop Sequoia into second place.
While the K20X is the go-to chip for top performance, the base model Tesla K20 is only slightly less powerful at 1.17 teraflops double precision and 3.52 teraflop single precision. The lesser performance is due to a smaller core count (2,496) and slightly slower GPU clock (706 MHz). The power draw is correspondingly lower, as well, at 225 watts. Memory capacity has been shaved to 5 GB, along with a somewhat reduced bandwidth — 208 GB/second. The K20 is not aimed at any particular demographic; it’s just a less performant K20X for situations where max compute and memory performance are not needed.
Note that single precision FP performance with both K20s is three times that of its double precision values, rather than two times as it was on the Fermi devices. This was an explicit design choice by NVIDIA. The rationale is that using single precision for much of the computation and double precision, only when that extra level of accuracy is required, will boost application performance, as well as energy efficiency, on codes that can take advantage of mixed precision computing.
Although the increase in raw memory performance is modest relative to the compute performance, memory bandwidth with error correction (ECC) turned on has been improved. Usually ECC incurs a significant overhead, but the NVIDIA engineers have managed to reduced it substantially. According to Sumit Gupta, NVIDIA’s GM of the Tesla business, typical applications will now see just a 6 to 8 percent penalty with ECC enabled, which is about half of what it was on the Fermi-generation devices.
Gupta says it’s not just raw speed that has been kicked up. Thanks to a variety of microarchitectural optimizations, execution performance of real software will be much improved as well. For example, the K20X can deliver 1.22 teraflops on DGEMM, a double precision matrix multiplication routine used across a number of science applications. That’s three times faster than the previous generation Fermi chip — remember for peak double precision, the K20X was only twice that of its predecessor.
Likewise, Linpack efficiency has been improved, from 61 percent of peak in Fermi, to 76 percent in the K20X. A single dual-K20X server with a couple of host Xeons for company deliver 2.25 Linpack teraflops. In the past, GPUs-accelerated machines have suffered from poor Linpack yield — more in the range of 50 to 60 percent..
“If Fermi was a big leap forward, Kepler is going to be twice as big in terms of revolutionizing high performance computing,” Gupta told HPCwire.
More to the point is real-world application performance, which Gupta says is going to get a big boost for users who upgrade to the new Kepler hardware. Compared against a standard dual-socket Xeon server, the same box equipped with a couple of matching K20X devices will enjoy a significant speed-ups on a variety of science apps, including MATLAB FFT and Chroma (18 times faster), geodynamics code SPECFEM3D (10 time faster) and molecular dynamics code AMBER (8 times faster).
Best all of is WL-LSMS, a material science code that gets more than a 32X boost with K20X acceleration. WL-LSMS captured the Gordon Bell prize a year ago at SC11, running at 3 petaflops on the Fujitsu’s K computer. Reworked for GPUs on the new Titan machine at ORNL, the same code hits 10-plus petaflops.
On the performance per watt front, the new Kepler hardware looks to be just as impressive. In fact, since the new chips now yield two or three times the performance in the same power envelope as the previous generation GPUs, the systems that house them will likely gravitate to the top of the Green500 list. After some preliminary tests, a small K20X-equipped supercomputer was able to deliver 2,142 megaflops per watt on a Linpack run. That would beat out the Green500’s top-ranked IBM Blue Gene/Q machine, which delivers just north of 2100 megaflops/watt.
The K20s are shipping now and sales are apparently off to a fast start. In the last 30 days, NVIDIA says its has shipped 30 petaflops worth of the K20 gear (24 petaflops of which are installed in Titan). That’s more than the aggregate capacity of the entire TOP500 list a year ago.
With Intel’s Knights Corner products and AMD’s FirePro S10000 cards also being launched today, the new K20 offerings will have some company in the teraflop-plus HPC accelerator category. To get an idea those products that stack up against NVIDIA’s finest, check out the accompanying stories today in HPCwire.