5
Tesla Kepler K80/K20m (GK 210/110)
• Sorken has device capability 3.7, Stampede has 3.5
u 11¼ (5) GB device memory (frame buffer)@ 240 (208) GB/s
u 1.5MB (1.25MB) shared L2 Cache (by all SMXs)
u 13 SMXs (2496 cores) on Sorken and Stampede
• Sorken’s K80 (GK210 GPU) has more registers and larger shared memory
per device than Stampede’s K20m (GK110 GPU)
u 192 SP cores, 64 DP cores, 32 SFUs, 32 Load/Store units
u Each scalar core: fused multiply adder, truncates intermediate result
u 112K (64KB) on-chip memory configurable as scratchpad memory + L1 cache
u 128K (64K) x 32-bit registers up to 255/thread
u 1 FMA /cycle = 2 flops/cycle/ DP core*64 DP/SMX*13 SMX = 1664 flops/cyc
@823.5 MHz (705.5 MHz ) = 2.74 TFLOPS per GPU (1.17)
Nvidia
Scott B. Baden / CSE 260, UCSD / Fall '15 5
