The idea of general-purpose computing on graphics processing units (GPGPU) continues to capture the imagination of the HPC community. But the three big players — Intel, NVIDIA and AMD — all have their ideas on how this new technology should play out.
When Intel rejected the whole notion of general-purpose computing on graphics processing units (GPGPU) at the spring 2007 IDF meeting with its announcement of its upcoming Larrabee product line, the digerati began to buzz about what the future might hold for the GPU. For those who might not have heard about it, Larrabee is Intel’s answer to the programmable GPU, the technology that is bringing GPGPU to the masses.
The Larrabee architecture could be characterized as the anti-GPU entry. The overall approach is an attempt to evolve the CPU into a terascale data parallel engine. According to Intel, Larrabee will be a manycore (i.e., more than 8 cores) device and will be based on a subset of the IA instruction set with some extra GPU-like instructions thrown in. Intel has not elaborated on how it intends to do this, but one could imagine super-sized SSE units with just enough x86 CPU silicon to enable general-purpose flow control and data access. The first product release will probably come in 2009, but Intel says it may have something to demo as early as next year.
The idea behind Larrabee is to bring both traditional graphics processing and data parallel computing under the IA umbrella. I’m not going to talk about the traditional graphics side of the story here (I’ll let the game weenies argue about the advantages of ray-tracing over rasterization.) What’s interesting about Larrabee and its GPU brethren is the extent to which a graphics engine can become a general-purpose computing engine without compromising its performance.
The combination of a data parallel engine with more of the general-purpose flexibility of a traditional CPU could offer a powerful model for scientific computing applications, which usually consist of an irregular mix of matrix math and other logic. One of the drawbacks of traditional GPUs is that they depend upon an accompanying CPU for virtually all of the non-vector logic. That’s fine if the application divides neatly between a vector computing kernel and the rest of the application logic in such a way as to keep both types of processing engines busy. But if it doesn’t, the software developer has to find a way to tease out enough parallelism for the GPU to make sending the vector data on a round trip from the CPU worthwhile. This will only get worse in the future, since chip-to-chip bus performance is not expected to keep pace with either CPU or GPU performance.
The division of labor problem is at the heart of the GPGPU critique elaborated by Anwar Ghuloum, an engineer at Intel’s Microprocessor Technology Lab. In a blog entry last week, The Problem(s) with GPGPU, he writes about some of the ramifications of the current CPU-GPU dichotomy:
[B]ecause of the underlying constraints of GPU architecture, oftentimes the program relies heavily on the CPU to manage the difficult parts of the control and data flow, as well as all the other (necessary) stuff like I/O, etc. Here’s the problem with this, the CPU-GPU link is relatively lower performance, engendering relatively high latencies for CPU-GPU interactions (like using a CPU to handle an outer level loop that the GPU can’t handle). This can have a devastating effect on performance.
Ghuloum is not explicitly making a pitch for Larrabee here. He’s really questioning the validity of the GPGPU programming approach, which he believes is too narrowly defined to exploit all avenues of data parallelism. In a previous blog post, Ghuloum makes a case for Ct, a language Intel is developing that supports a more general-purpose, deterministic parallel programming model. While Ct assumes no specific architecture, the underlying model he’s describing seems to point to a more generalized parallel processing architecture, like Larrabee.
NVIDIA offers a more traditional approach to GPGPU. Its Tesla product line and CUDA C-programming environment were specifically developed to deliver GPU computing to the HPC market. The current Tesla products, released in June 2007, are based on the G80 architecture but packaged in form factors that are geared toward high performance computing setups, both workstations and servers. Host communication is done via PCI Express (PCIe).
There’s plenty of low-hanging fruit to be had with Tesla. Seismic analysis, medical diagnostics, molecular modeling and other such applications can realize performance increases of one or two orders of magnitude from this type of GPU acceleration. The next generation of Tesla offerings are expected to support double precision floating point. This will expand the GPGPU application domain even more, since 64-bit floating point is the de facto standard for scientific computing.
NVIDIA may eventually move its high performance computing Tesla line, or its descendents, in the same direction as Larrabee. But unlike Intel, NVIDIA starting point is the GPU, and it has no in-house CPU to draw from, so the path is bound to be different. For now, NVIDIA is content to exploit its lead in the GPGPU arena, especially since its nearest competitor, AMD, is still in the process of putting its GPU computing strategy together.
At one time, AMD seemed to be ready to take advantage of the renewed interest in GPGPU. Soon after the company acquired ATI in July 2006, it launched its “Stream Computing” strategy, with the idea of leveraging ATI’s GPUs and AMD’s HyperTransport interconnect technology. The company’s first GPGPU platform consisted of a PCIe-connected ATI R580 GPU bundled with their “Close to the Metal” software development kit. But it’s not clear how many of these platforms have been sold, and AMD hasn’t talked much about stream computing since 2006.
Over the past year, the company has struggled against Intel’s onslaught of new x86 technology and aggressive chip pricing. If that wasn’t enough of a distraction, NVIDIA’s foray into the GPGPU arena seemed to catch AMD off-guard. Even if the company’s initial GPU plans have slipped, AMD’s long-term commitment to marry its two architectures remains. But with Intel and NVIDIA forging ahead, time is no longer on AMD’s side.
The first instance of AMD’s upcoming Fusion processor, which integrates a CPU and GPU on the same die, is at least a year away and is intended for the consumer market (notebooks). If successful, later generations of Fusion will almost certainly target HPC, and are likely to resemble a Cell processor architecture, with multiple CPU and GPU cores. Chip level CPU-GPU integration offers a number of advantages over discrete components, namely increased energy efficiency and better communication bandwidth and latency (HyperTransport versus PCIe). It’s not the Larrabee model, but it offers the same advantage of using an x86 base to create a platform with much greater capabilities for data parallelism. AMD is also likely to offer discrete GPU products for high-end computing, but no roadmap has been publicized.
Like Intel, AMD has hinted at adding GPU-type instructions to the x86 ISA to allow software to work seamlessly with the graphics engines via a standard compiler/runtime. If AMD and Intel were on speaking terms, they could forge a common GPU ISA, which would be much appreciated by the GPGPU ecosystem. It could also serve to blunt NVIDIA’s lead, and probably force the company to adopt what would be an industry-standard GPU interface. In the short term, standards are unlikely. Everyone involved has their own vision of how the GPU should evolve into its new role.
This is one reason why high-level software environments for parallel programming are needed. While the Ct language looks promising, it’s still in the research stage. (I’m guessing we’ll soon be hearing more about this Intel.) Today, RapidMind offers a high-level software platform that allows developers to exploit data parallelism on a variety of hardware architectures, including NVIDIA and AMD GPUs, the Cell BE, and soon, x86 CPUs. The RapidMind platform has been generally available for less than a year, but has already managed to attract over 1,000 developers.
Given the asymmetric capabilities of the different chip vendors and the immaturity of the GPGPU software ecosystem, it’s too early to make predictions on the future of GPUs for general-purpose computing. What seems more certain is that proprietary vector processor-based supercomputers, like the one just announced by NEC this week, will soon be edged out by commodity-based systems that contain the equivalent vector smarts. Whether these machines turn out to be based on double precision GPUs, GPU-CPU hybrids, SIMD-enhanced CPUs, Cell BE processors, FPGAs, or SIMD ASICs, remains to be seen.
As always, comments about HPCwire are welcomed and encouraged. Write to me, Michael Feldman, at firstname.lastname@example.org.