|
@@ -91,12 +91,12 @@ CHIP_FLOPS = {
|
|
"NVIDIA GEFORCE RTX 2080": DeviceFlops(fp32=10.07*TFLOPS, fp16=20.14*TFLOPS, int8=40.28*TFLOPS),
|
|
"NVIDIA GEFORCE RTX 2080": DeviceFlops(fp32=10.07*TFLOPS, fp16=20.14*TFLOPS, int8=40.28*TFLOPS),
|
|
"NVIDIA GEFORCE RTX 2080 SUPER": DeviceFlops(fp32=11.15*TFLOPS, fp16=22.30*TFLOPS, int8=44.60*TFLOPS),
|
|
"NVIDIA GEFORCE RTX 2080 SUPER": DeviceFlops(fp32=11.15*TFLOPS, fp16=22.30*TFLOPS, int8=44.60*TFLOPS),
|
|
"NVIDIA TITAN RTX": DeviceFlops(fp32=16.31*TFLOPS, fp16=32.62*TFLOPS, int8=65.24*TFLOPS),
|
|
"NVIDIA TITAN RTX": DeviceFlops(fp32=16.31*TFLOPS, fp16=32.62*TFLOPS, int8=65.24*TFLOPS),
|
|
- # QUATRO RTX Ampere series
|
|
|
|
- "NVIDIA QUATRO RTX A2000": DeviceFlops(fp32=7.99*TFLOPS, fp16=7.99*TFLOPS, int8=31.91*TFLOPS),
|
|
|
|
- "NVIDIA QUATRO RTX A4000": DeviceFlops(fp32=19.17*TFLOPS, fp16=19.17*TFLOPS, int8=76.68*TFLOPS),
|
|
|
|
- "NVIDIA QUATRO RTX A4500": DeviceFlops(fp32=23.65*TFLOPS, fp16=23.65*TFLOPS, int8=94.6*TFLOPS),
|
|
|
|
- "NVIDIA QUATRO RTX A5000": DeviceFlops(fp32=27.8*TFLOPS, fp16=27.8*TFLOPS, int8=111.2*TFLOPS),
|
|
|
|
- "NVIDIA QUATRO RTX A6000": DeviceFlops(fp32=38.71*TFLOPS, fp16=38.71*TFLOPS, int8=154.84*TFLOPS),
|
|
|
|
|
|
+ # QUADRO RTX Ampere series
|
|
|
|
+ "NVIDIA RTX A2000": DeviceFlops(fp32=7.99*TFLOPS, fp16=7.99*TFLOPS, int8=31.91*TFLOPS),
|
|
|
|
+ "NVIDIA RTX A4000": DeviceFlops(fp32=19.17*TFLOPS, fp16=19.17*TFLOPS, int8=76.68*TFLOPS),
|
|
|
|
+ "NVIDIA RTX A4500": DeviceFlops(fp32=23.65*TFLOPS, fp16=23.65*TFLOPS, int8=94.6*TFLOPS),
|
|
|
|
+ "NVIDIA RTX A5000": DeviceFlops(fp32=27.8*TFLOPS, fp16=27.8*TFLOPS, int8=111.2*TFLOPS),
|
|
|
|
+ "NVIDIA RTX A6000": DeviceFlops(fp32=38.71*TFLOPS, fp16=38.71*TFLOPS, int8=154.84*TFLOPS),
|
|
# Common Server GPUs
|
|
# Common Server GPUs
|
|
"NVIDIA A40 48GB PCIE": DeviceFlops(fp32=37.4*TFLOPS, fp16=149.7*TFLOPS, int8=299.3*TFLOPS),
|
|
"NVIDIA A40 48GB PCIE": DeviceFlops(fp32=37.4*TFLOPS, fp16=149.7*TFLOPS, int8=299.3*TFLOPS),
|
|
"NVIDIA A100 40GB PCIE": DeviceFlops(fp32=19.5*TFLOPS, fp16=312.0*TFLOPS, int8=624.0*TFLOPS),
|
|
"NVIDIA A100 40GB PCIE": DeviceFlops(fp32=19.5*TFLOPS, fp16=312.0*TFLOPS, int8=624.0*TFLOPS),
|