Must …. not …. post …. before …. coffee ….

I disagree with the comment “the general accelerator rush is coming to an end”. On the contrary, speaking with customers, it seems to be picking up. Clearspeed had issues as a company and as a product, but they are/were not the only accelerator provider.

The argument for accelerators is compelling for the smaller users in terms of leveraging power to their applications. It is compelling for the software ISV’s looking to help their customers decrease the cost of their hardware in order to have more budget for the ISV’s software.

Accelerators offer some combination of many more processor cycles per wallclock tick, or more efficient cycles per wallclock tick. As we demonstrated recently with GPU-HMMer, a single machine with 3 GPUs can outperform a more power-hungry/harder to maintain cluster on the same code. That is, there is a financial, ease of use, and believe it or not, a green argument to be made for using accelerators. As you note in the subsequent article, the memory wall is a problem, and curiously enough the GPUs suffer from their own version of it, though it is ~10x further out than the memory wall on the host (what we call computing substrate).

Clearspeed failed due to them having a business model that required ~$5k/accelerator for ~10x wallclock on ordinary applications, where your code needed to be ported, and Clearspeed accelerators were not ubiquitous. Programmable GPUs are in 1E+7 devices, cost from $150-$1800/unit, giving 5-30x per application (not per kernel). Couple this ubiquity with the low cost (I can/do develop Cuda code on my laptop) of the platform and the zero cost of the tools, and you have something of interest for people in need of ever more computing capability with an ever decreasing budget for the same.

We might have to agree to disagree, but accelerators appear to have a very bright, and long future ahead of them in HPC.

Joe