August 8, 2012

AMD Unveils Teraflop GPU with ECC Support

Michael Feldman

Advanced Micro Devices (AMD) has launched six new FirePro processors for workstation users who want high-end graphics and computation in a single box. One of them promises a teraflop of double precision performance as well as support for error correcting code (ECC) memory. The new offerings also includes two APUs (Accelerated Processing Units) that glue four CPU cores and hundreds of FirePro GPU stream processors onto the same chip.

The straight-up GPU-based cards are the FirePro W9000, W8000, W7000 and W5000. All are meant to provide hefty graphics support as well as respectable number-crunching performance. They are based on AMD’s new Graphics Core Next Architecture, which according to the company is their “first design specifically engineered for general computing.” Application development is supported via C++ AMP (Accelerated Massive Parallelism) and OpenCL, two open standard languages that are meant to offer an alternative to NVIDIA’s CUDA programming framework.

The top-of-the-line W9000 and W8000 are the ones built to chew on heavy-duty numeric codes such as CAD, CAE, medical imaging, and digital content creation, while also providing enough graphics muscle to drive up to six 30-inch displays. Both are double-slot cards that support the newer, faster PCIe Gen3 interface.

Performance-wise, the W9000 and W8000 are rather impressive beasts. The W9000 is the one that will deliver a double precision (DP) teraflop of peak performance. If only 32 bits of precision are required, this same chip will provide a whopping four teraflops in single precision (SP). That outruns NVIDIA’s fastest Tesla GPU (665 DP gigaflops and 1330 SP gigaflops) by a fair margin, although their newer Kepler K10 card edges out the W9000 in the single precision department with 4.58 teraflops, and the upcoming Kepler K20 is likely to be well above a DP teraflop when it’s launched in a few months. The K20 will be NVIDIA’s flagship supercomputing chip, but like these FirePros, will also be tapped for workstation duty.

The most interesting new feature in the FirePro lineup is the addition of ECC support, which is incorporated in the W9000 and W8000 products. ECC is used to insure that errant memory bits flips don’t mess up numeric calculations and is specifically aimed at high performance computing (HPC) codes. NVIDIA introduced this feature into its GPGPU Tesla lineup back in 2009.

The introduction of ECC suggests AMD is taking GPU computing a lot more seriously that it did when it was hawking its non-ECC FireStream GPU cards a few years ago. It also suggest that we may soon see some server-class FirePro offering with this capability in the near future. In fact, ECC actually has limited use in workstations. It’s real value becomes apparent when applications are deployed at scale, across multiple nodes of a cluster, where there is a much more likelihood that flipped memory bits will result in software failures.

ECC aside, the latest FirePro cards have a decent amount of memory and bandwidth for both graphics or computation (and a lot more than the older FireStream offerings). The top-end W9000 sports 6 GB of on-board GDDR5 memory and 264 GB/sec of bandwidth. Those specs are pretty much on par with the latest NVIDIA Tesla modules, but again, the upcoming Kepler parts will probably leapfrog the W9000 in this area. The W9000 card draws 274 watts at peak load and its suggested retail price is $3,999.

For $2,400 less, the W8000 comes with nearly as much number-crunching capability (3.23 SP teraflops and 806 DP gigaflops), but just two-thirds the memory (4 GB) and bandwidth (176 GB/sec). The less muscular W7000 and W5000 represent the mid-range FirePro lineup. Both provide more than one SP teraflop and a token number of DP flops, but they lack ECC support. MSRP is $899 and $599, respectively.

AMD’s new APUs, the FirePro A300 and A320, are a different breed altogether. They represent the chipmaker’s first heterogeneous processors aimed at science and engineering, albeit only for the workstation market. Unlike the company’s previous APUs for desktops and laptops, these latest ones include a lot more GPU heft. That gives users something akin to a mid-range discrete GPU on the same chip as a quad-core CPU. The advantage here is that it’s much easier to share data between the two compute engines on the chip; there’s no need to be sending bytes back and forth across a relatively slow PCIe bus.

These new APUs aren’t computational powerhouses however. At 28nm, there’s just not enough room to lay down a lot of GPU silicon on the same die as a CPU with reasonable-sized memory caches. As a result, AMD is aiming these cards at 2D modeling and entry-level 3D modeling, rather than more demanding applications like CAE and medical imaging.

The 100 watt A300 delivers 736 SP gigaflops and 184 DP gigaflops, with the 65 watt A320 just a tad slower at 693 SP gigaflops and 173 DP gigaflops. From the standpoint of double precision performance, that’s not much better than a top bin Sandy Bridge CPU, but with the APU, of course, you get the added functionality of graphics support, not to mention a lot more SP flops. If you need more compute than the A300 or A320 can provide, AMD offers what they call Discrete Compute Offload, which enables the devices to work with a separate FirePro GPU running in parallel.

For HPC users, perhaps the most interesting news here is that AMD is gearing up its GPU computing portfolio, both for its discrete and heterogeneous lines. What we’re likely seeing are the precursors to server-capable GPUs and APUs that will be aimed at HPC and related types pf applications. In all likelihood, the server GPUs will come first, perhaps as early as this year.

But with NVIDIA’s dominance of the HPC accelerator market and Intel’s imminent entry into that space with its upcoming Knights Corner coprocessor, AMD will have an uphill battle against a couple of formidable competitors. The company’s natural advantage in CPU-GPU integration may eventually give them a leg up when transistor geometries allow teraflop graphics engines to inhabit the same die with multicore CPUs. But until then, AMD will have to play catch up.