HPUs will transform the world

Posted: 04.21.11

Just got back from COFES in sunny warm Scottsdale Arizona. Ensconced in the conference at a resort, and venturing out only at night to nearby restaurants, we were immune to the tremendous political stress this old west state is suffering through. Battling on multiple fronts with immigration pressure and unconstitutional remedies, minimal gun laws, tea parties, budget cuts, birther bills, an infamous country sheriff who's being sued by the feds, collapsed real estate values and tax base, it's amazing how cheerful everyone we met seems to be. Either they're living in some kind of a denial dome, or it's true what they say about themselves that they can handle anything—with a smile. Maybe it's the sunshine.

The sun is shining on the opportunities for the HPUs and I gave a little talk on the power of these devices at the COFES Conference. COFES deals with engineering software, CAD, CAE, PLM, BIM, and FEA.

The HPU has applicability in most of those industries, other than PLM. And one area, although not a huge market, may become the poster child for what a HPU can do – I'm talking about FEA and CFD—finite element analysis and computational fluid dynamics—things with zillions of data points all interrelated with algorithmic functions.

FEA/CFD programs were early experiments in GPU-compute and they showed great promise, but the final results were disappointing. Due to the latency and transfer of data, and speed of the PCI Express bus, moving those big datasets ate up the gains from the GPU's SIMD construction. But an HPU, with on-chip data transfers via high-speed caches changes the whole dynamic.

ANSYS, DS Abaqus, Autodesk, MSC, Siemens, and a dozen other companies offer FEA/CFD programs. The market is worth about $1.6 billion including software and services. Not everyone uses (or can use) FEA/CFD software—it's complicated. As mentioned, it involves huge data sets, the software is expensive, and usually comes with services from the software supplier to get things working. The computational run-times can be days to weeks. As a result, there are only about 300,000 packages sold a year. But, without such software no airplane would fly, no building would be built, and no sports shoe would get made. Gradually, fairly simple types of analysis are being included in software so that more people can use it. The trend is clearly towards more analysis capabilities becoming available to more designers.

HPUs and discrete GPUs are also changing the way analysis is being done as computation time is brought down, datasets are allowed to get bigger, and/or more accurate. But as good as they are, HPUs and GPUs still represent a programming challenge to the software vendors and the users. You don't just drop a spreadsheet on a FEA program and push a button, there's lot of user (read multiple engineers) involvement. And there should be—this is deadly serious stuff being evaluated. The wrong answer can put hundreds, if not millions of lives at risk.

HPUs and GPU-compute is at the threshold of their growth cycle. As programmers learn to tame them, and users regale us with the joy of using them (usually without knowing why things are faster and better), the deployment of SIMD architectures as stand-alone or closely coupled with a resident CPU will explode. HPU/GPU-compute will grow much faster than the PC market or the vaunted mobile market.

In addition to the software suppliers of FEA/CFD and other computational engineering software there is a cadre of consultants who design tests, datasets, and training programs.

We will, in my opinion, see a renaissance in FEA/CFD due to the development and deployment of HPUs and massively powerful GPUs. And as we do, life in general will get better, products will last longer and be safer, and a new age of product development will be spawned.

HPU/GPU-COMPUTE MARKET potential in FEA/CFD (source JpR)