The Sandy Bridge Preview (Anand)

[kl0012]

you missed the part where the 5450 uses 10 watts not including board power. the 5450 is 5gflops/W and the 5970 is 15gflops/W. big difference.

In Anands test the diff between idle to load for SB is ~60W means that we have 3.3gflops/W which is not far from 5450 and also OOO/branch prediction/prefetchers/ which uses most of the power allows to utilize these flop/s in more effective way. BTW nvidia gt310/320 has even lower gflops/w ratio.

but comparing flops to flops as a performance metric is probably about as inaccurate as you can get. we have already seen the 5870 achieve 2.2tflops in sgemm. an ideal case for sandy bridge is 20x slower.

~12x slower you mean?
http://software.intel.com/en-us/arti...mkl-v103-beta/
* AVX DGEMM (M, N, K=8Kx4Kx128) performs 1.8x over NHM. AVX DGEMM/SGEMM achieves 88-90% machine peak.
* The AVX/NHM speedup is 1.8x for radix-2 1D CFFTs with N=1024
* The Intel® Optimized LINPACK benchmark, using Intel AVX optimizations, performs over 1.86x (or over 80% overall efficiency) on 4 cores with N=20000.