What Tianhe-3 Should/Could of been

[Sommer]

The table is incorrect.

E5-2692 have 12 core 24 threads with hyper threading. Threads goes with core, you must have at least the same number of threads to the core otherwise the circuit will not be complete

And you cannot compare side by side for Processor to Co-processor. Processor is located within the C/B hubs, while CO-Processor are existed in PCI card. Many of the specification between processor and co-processor is different, for example, since Co-Processor using PCIe Bus, the bus width are the same and there will be no thread involved as they are connected to the PCI Bus in parallel

And Finally, according to AMD and Oak Ridge, Opteron 6274 have 141GFlop peak, not 563

http://on-demand.gputechconf.com/su...SB005-Bland-Titan-Oak-Ridge-National-Labs.pdf


The core computing power are based with the matrix of instruction calculated per second per processor, and I have already made my calculation.

You are right about E5-2692 have 12 core 24 threads and also Opteron 6274 have 141GFlop peak. My mistake.
But it's Double-precision floating-point format for Tianhe-2 and also Titan. That's why their total Gflops are both not equal Frequency(GHz) times The number of cores times The number of threads. That means, none of them can get the top speed like you said, when you get the number "four times". Ithink, that it's limited by the CPU and GPU or the system.

 

[jhungary]

You are right about E5-2692 have 12 core 24 threads and also Opteron 6274 have 141GFlop peak. My mistake.
But it's Double-precision floating-point format for Tianhe-2 and also Titan. That's why their total Gflops are both not equal Frequency(GHz) times The number of cores times The number of threads. That means, none of them can get the top speed like you said, when you get the number "four times". It's limited by the CPU and GPU.View attachment 275011

Well, again, the 4 times speed is based on hardware spec. Of course it is not going to reach that, but the problem is that Tianhe-2 have overwhelming computing power (Imagine if Titan uses E5-2692 with similar amount of processor and Similar amount of K-20X, the speed would have been a lot faster than Tianhe-2. The problem lies at how Titan 3-D interconnect with GPU instead of Co-Processor unit.

This is my point, next gen supercomputer will be relying on titan style CPU-GPU set up, and this will provide more processing power with less core because GPU provide batched parallel process used to perform high volume graphic rendering, while China is catching up (Still have a LONG WAY TO GO) on CPU technology, Chinese lag even further behind on GPU technology. Which will be a hurdle to next generation supercomputer development.