Compute Unified Device Architecture
CUDA is NVIDIA’s brand name for general purpose computing on GPUs and this is absolutely part of NVIDIA’s core vision for the next decade. CUDA has its roots in NVIDIA Gelato. Gelato is NVIDIA’s offline rendering technology that allows Hollywood production level quality renders to be calculated using GPU. Gelato renders have made their way into movies such as Resident Evil: Extinction, and the same visual effects studio is working on John Woo’s Battle for Red Cliff.
CUDA took the idea of using the GPU to perform complicated math at non-real-time speeds further by making it fully programmable. Based upon the Pathscale C compiler, CUDA allowed developers to take advantage of the number crunching capabilities of today’s gaming GPUs. Going beyond Folding @ Home, real-world uses of CUDA include molecular dynamics, MRI processing, weather simulation, as well as seismic databases. Compared to traditional CPU models, GPU based solutions can provide stratospheric increases in performance given the right application.
While CUDA will only work for the right applications, there are still limitations in the current platform including the lack of recursion and the bandwidth/latency limits between the CPU and GPU. This is where the addition of on-chip x86 capabilities will be helpful. The C compiler is already based upon the x86 architecture, so the software development process is “simplified” (as opposed to going with a different instruction set). As important, these x86 processors would simply be performing all of the “housekeeping” operations, relying on the heavy number crunching to be done by the GPU, “simplifying” the hardware side of things (as opposed to developing a full-fledged FPU to go with the CPU). “Simplifying” has been put into quotes because this is no easy task and while we believe that NVIDIA will include some x86 capabilities in the future to extend CUDA, we don’t think it’s ready for NVISION ‘08.
“System on a Chip”
Let’s step back from the scientific computing side of things and look at the second area where NVIDIA is going to be using x86 technology: a system on a chip. While Intel’s GPU and next-generation i7 CPU has caught the attention of most technophiles, the Intel Atom is a CPU that also deserves attention. Even with Intel’s fabrication capabilities, reports suggest that demand for the Intel Atom is exceeding supply. Even though this first generation CPU has primarily been found in Netbooks such as the Asus EEE, Intel’s vision for the Atom includes use in the smartphones and CE devices. That’s because, at 1.6 GHz, the Atom offers almost equal performance to the Pentium III 1GHz at an amazing 2W TDP.
A few caveats though. Intel has been pairing the Atom with the older 945GC chipset and as a result, an 8W AMD Athlon 64 2000+ (1GHz) paired with AMD’s 780G actually offers better performance and economy. This is one area where NVIDIA has the potential to compete, all they need is a CPU that’s competitive to Atom.
Enter Transmeta, a company that has broadly licensed its technology to NVIDIA.
Remember the Crusoe and the Efficeon? Exactly. Either you don’t remember these because they were 2004-era products, or you do, and remember that these CPUs were too slow to compete against the standard AMD and Intel CPUs of the day. However, a look back at Transmeta’s performance brings up a lot of interesting points.
Intel’s own presentation slides had the 1.6GHz Atom reaching 126-130 on the Embedded Microprocessor Benchmark, with hyperthreaded versions reaching 172 points. The Transmeta 1GHz Efficeon hit 137 points. The Efficeon had a TDP of 3W at 1GHz. In 2004. At 90 nm. So not only did the Transmeta CPU offer Intel Atom performance, it offered nearly the same power envelope.
Oh by the way, the Efficeon used a Hypertransport interconnect, was manufactured by TSMC, and used the NVIDIA nForce3 as the reference platform.
It turns out that thermal issues meant that the Efficeon would throttle down, and the integrated memory controller wasn’t that fast. But considering the improvements in manufacturing process, the improvements in NVIDIA’s chipset platforms, and the today’s market for low-power CPUs, the opportunity is ripe for NVIDIA to launch an x86 platform built upon around Transmeta’s original work. They could launch a multi-chip platform today and be competitive with Intel Atom, but we suspect the ultimate goal will be SOC capabilties.
x86 from NVIDIA?
Absolutely. It makes sense in at least two areas of NVIDIA’s active focus, they have the technical resources and the economic incentive.
x86 from NVIDIA next week?
Publically? R21 doubts it. NVIDIA would have to be way ahead of schedule.