MKL Itanium 2 Performance

[tcrony70]

I was looking at a variety of information and was trying to determine what the theoretical peak performance for an Itanium 2 (1.5 GHz) is.

From some information I found on various web sites, it would appear that for double precision floating point data, the peak would be 6 GFlops and 12 GFlops for single precision data (based on the 4 floating point units available for single precision). Is this correct?

However, there is a performance graph for DGEMM on one of the MKL pages which shows a performance of more than 6 GFlops (single CPU).

Can someone help clarify this?

Thanks,

Tim

[Todd_R_Intel]

I'll investigate the performance on our website. It may be an error (perhaps with the clock speed reported).

For both single and double precision, the theoretical peak of an Itanium 2 processor is 6 GFLOPS.

Todd

[tcrony70]

Thanks for the response.

Does this mean that there are not 4 floating point units capable of doing single precision arithmetic as I saw in several places on the web or that only 2 can act in a any given cycle (since I would have to believe that they can do a fused multiply and add).

[TimP]

Yes, that's based on retiring 2 fused multiply-add instructions per clock cycle.

[tcrony70]

So what would be the point of having 4 floating point units that can handle short precision if one can still only retire 2 floating point instructions per clock cycle?

[TimP]

You put Itanium 2 in the title. With Itanium 1, in principle, it was possible to execute SSE instructions. If your goal is to execute SSE code efficiently, surely Itanium 2 is not your vehicle.

[Intel_C_Intel]

The data you saw was erroneously labeled as being 1.5 GHz, but was, in fact 1.6 GHz, which explains why the performance exceeded 6 GFLOPS on DGEMM. The maximum performance in either single precision or double precision, as mentioned elsewhere, is 4 floating point operations per clock, or 2 FMAs per clock.