Aha! I hadn't seen that new x86 FLOPS column and the new FLOPS FAQ that goes with it.jaak ennuste wrote:Isn't there now apples-to-apples comparisin x86 FLOPS metrics up and running now?
http://fah-web.stanford.edu/cgi-bin/mai ... pe=osstats
To partly answer your question, Yes and No. Yes, that new column is a better representation of the FLOPS, and therefore a better comparison of the amount of calculations being done by each client type. But again NO, FLOPS does not directly equal usable science production, and nowhere on that OSTATS page does Stanford say anything about FLOPS equating to science, or FLOPS equating to points.
The GPU and PS3 clients can do one type of calculation really fast. However, Stanford still needs all types of calculations. So while the GPU and PS3 are really valuable for their speed, they are less valuable because they can not do all of the calculations by themselves. Some of the calculations still need to be done by the CPU clients.
Remember how I said that primary concern for setting the PPD of each client benchmark was the value of the science it produces? That's why a GPU that is 20x faster at science than a CPU client is only benchmarked at 15x the points. Speed isn't everything, so therefore FLOPS is not everything. FLOPS do not equal Points.
On a side note, remember how I said the GPU client had to do some calculations twice because it couldn't store the results in memory? That detail is mentioned in the FLOPS FAQ. However, I only had half the story. It only affects the ATI GPU client, not other clients. But as I mentioned above, that is only one of many reasons why FLOPS do not easily equate to points, and still not the best for comparing clients' performance.
At least Stanford is taking steps in the right direction by giving us more information to work with.