The GPU specture on the horizon
So the point of the HPCJCC is to offer customers an opportunity of using an HPC system to benchmark their particular applications. They know they need something like it but may be unsure on particular specification. Once benchmarked, the sales teams go in for the monetary kill. Think of it as a pragmatic advisement tool.Thinking of the broader picture, we know that a number of HPC-oriented applications are suited towards the low-latency, serial processing offered by modern CPUs. Put a bunch of them together, connect them up, and you have an HPC setup that can be hugely expensive. That's precisely why the likes of IBM and HP offer optimisation and benchmark services, to custom-fit to requirements and grab customers who may have gone to other large-installation vendors.
The most pertinent question, though, has to be what happens when companies realise that many of the applications in an HPC space can be coded - with particular attention to NVIDIA's CUDA and ATI's CTM - such that they execute in a parallel fashion, on massively-parallel GPUs, and are run much faster at a lower cost?
We've already seen the undeniable advantages of GPGPU, with potential teraFLOPS-in-a-box performance espoused by ATI and soon-to-be-matched with NVIDIA's next offering, GT200. You'll still need a CPU and bags of memory, of course, but it's more than feasible that the average HPC installation in three years' time will focus on GPUs rather than CPUs: there's simply more throughput.
Intel and AMD sells a significant number of its CPUs to the HPC space and Intel, in particular, won't allow sales and margins to be eroded via a migratory process to an HPC paradigm that has more GPUs than CPUs. We guess that's where the first-run Larrabee fits in. You can't have too many cores when running parallelisable code, right?
Yes, the HPCJCC provides an interesting insight into how the burgeoning cluster-based HPC environment is shaping up in 2008. Got a problem that requires serious compute power? Need something in the $250K range? You could always go the HPCJCC or equivalent centre from other tier-1 vendors. That's not likely to change, but what may, and indeed will, is how much of that compute is done on GPUs rather than CPUs.
