Chances are that this pharmaceutic company's "embarrassingly parallel" workload could be ported to GPUs to run on anywhere from 1/10th to 1/100th the number of machines.
Porting the application to GPUs may offer a good ROI if the company intends to run this workload often enough.
Chances aren't, actually. GPUs impose many constraints. I engineer embarrassingly parallel workloads for a living, using both CPUs and GPUs, and only a small minority of real codes can be usefully ported to GPUs today. Those that can usually see a ridiculous speedup, but they're the exception, not the rule.
In any case, I know these guys and they're not newbies. They're familiar with GPUs and what they can do. They're running the workload they need to be running.
I seem to be right. The workload was molecular modeling (see blog post); and many of the algorithms are very well suited to GPU, see: http://www.ks.uiuc.edu/Research/gpu/
I am not saying they are newbies for failing to exploit GPUs. There could be other reasons why they did that. For example the pharma needed the results ASAP and may not have ported that particular molecular modeling app to GPU yet. GPGPU is a nascent field after all.
I do too write GPU applications (crypto bruteforcers), just so you know ;)
And I was right. They replied to my comment, saying that workload "still need to be ported [to GPUs]". They just couldn't do it because it was a proprietary app.
Maybe and maybe not. GPUs are good for problems where the appropriate data can be placed near the vector processor, a priori. There are many embarrassingly parallel problems where memory coherency does not exist, or where the input data sizes exceed what is currently available on GPUs, and for these, the bandwidth to system memory is going to be a severe bottleneck on performance.
Depends on how memory/storage intensive the tasks where:
"26.7TB of RAM and 2PB (petabytes) of disk space. "
One advantage of going with Amazon, is their really high-speed and voluminous ephemeral storage available per instance in addition to your EBS backed root volume.
They sounded completely CPU-bound from the blog post. They went with "high-CPU" c1.xlarge instances that had modest RAM and storage specs. They gave no details and expressed no concerns whatsoever about RAM or storage bottlenecks.
The stated application, molecular modeling, often requires double-precision floating point, for which GPUs offer less of an advantage than for single-precision and integer operations.
Double precision used to be a problem for GPUs, but not anymore with the latest GPU microarchitectures.
- A single AMD HD 6990 is capable of 1275 DP GFLOPS
- A single Nvidia Tesla 20xx is capable of 515 DP GFLOPS
- A 6-core CPU at 3GHz is capable with SSE of only 72 DP GFLOPS
Porting the application to GPUs may offer a good ROI if the company intends to run this workload often enough.