The only way to know the performance of this approach would be to measure it on a real system. The chip might become limited by power before it reaches this performance. If the M1 Ultra can run at 3 GHz with this approach, the performance would be a total of 3.84 TFLOPS for FP64 matrix multiply on P-cores. Matrix and tensor multiplies have a higher FLOPS/bandwidth ratio than pretty much anything so L1 cache bandwidth should not be a bottleneck.Īs a ballpark figure, since AMX is 2x the speed of Neon for FP64 matrix multiply, one could hope to get 2.5x the speed of AMX by using both AMX and Neon (2.5 AMX = AMX + (3/2)AMX ). Since each P-core has a 128 KByte L1 data cache, it can do 2 x (128K/8)^2 = 537M FLOPS before there is contention with the other cores in the cluster for the shared L2 cache. To get the most FP64 matrix multiply performance, one of the four P-cores in a cluster could use the Accelerate library and the other three cores could use Neon. In other words, AMX is twice the performance of Neon for FP64 matrix multiply. With the Quadro K5200, you only have the same FP64 performance as with gaming cards, notwithstanding the advanced GPU. The 4 P-core AMX units can do a total of 512 FLOPS per cycle for FP64 matrix multiply. Im unsure on the FP64 ratio of the 1080, any idea EDIT: No, not for 6Sigma, just a range of MPI and CUDA accelerated CFD applications. If each P-core can do 8 FP64 MADDs per cycle, that would be 2 FLOPS/MADD x 8 MADDS/core x 16 cores = 256 FLOPS per cycle not 128 FLOPs per cycle. It looks like with the Titan Black, you could change the FP64:FP32 ratio to 1:3, which is comparable to the Tesla cards (but at 1/4 the price). There are 16 Firestorm cores (P-cores) in M1 Ultra. I'm interested in doing some research using Molecular Dynamics, which very much benefits from having FP64 precision. 'Finally, one of the FPUs provides extended math capability to support high-throughput transcendental math functions and double precision 64-bit floating-point.
Are there any other cards that allow you to do this Id like to get close to 1:3 if possible without buying a Tesla range card (Id get some Titan Blacks, but they are basically impossible to find at this point). In the document GVCS001-The Compute Architecture of Intel Processor Graphics Gen8.pdf. Don’t forget that there is also the actual CPU :) Firestorm SIMD units can do 8 FP64 MADDs per cycle, which adds up It looks like with the Titan Black, you could change the FP64:FP32 ratio to 1:3, which is comparable to the Tesla cards (but at 1/4 the price).
0 kommentar(er)
