原创 算力百科 J 2025-05-13 06:01 贵州
智算中心推理算力5090 还是有性价比优势的,建议大家关注GB10方案,目前还没有量产,如果GB10方案量产,预计推理算力市场单价会进一步降低。
GraphicsCard | GeForceRTX 5090 |
GPU Codename | GB202 |
GPU Architecture | NVIDIA Blackwell |
GPCs | 11 |
TPCs | 85 |
SMs | 170 |
CUDA Cores/ SM | 128 |
CUDA Cores/ GPU | 21760 |
TensorCores/ SM | 4(5thGen) |
TensorCores/ GPU | 680(5thGen) |
GPU BoostClock(MHz) | 2407 |
RT Cores | 170(4thGen) |
RT TFLOPS | 317.5 |
PeakFP32 TFLOPs(non-Tensor). | 104.8 |
PeakFP16 TFLOPs(non-Tensor). | 104.8 |
PeakBF16 TFLOPs(non-Tensor). | 104.8 |
PeakINT32 TOPs(non-Tensor).. | 104.8 |
RT TFLOPS | 317.5 |
PeakFP4 TensorTFLOPS withFP32 Accumultate(FP4 AI TOPS) | 1676 / 3352 . |
PeakFP8 TensorTFLOPS withFP16 Accumulate. | 838/ 1676. |
PeakFP8 TensorTFLOPS withFP32 Accumulate. | 419/ 838. |
PeakFP16 TensorTFLOPS withFP16Accumulate. | 419/ 838. |
PeakFP16 TensorTFLOPS withFP32Accumulate. | 209.5/ 419. |
PeakBF16 TensorTFLOPS withFP32Accumulate. | 209.5/ 419. |
PeakTF32 TensorTFLOPS. | 104.8/ 209.5. |
PeakINT8 TensorTOPS. | 838/1676. |
FrameBufferMemorySizeandType | 32GB GDDR7 |
MemoryInterface | 512Bit |
MemoryClock(DateRate) | 28Gbps |
MemoryBandwidth | 1792GB / sec |
ROPs | 176 |
PixelFill-rate(Gigapixels/sec) | 423.6 |
TextureUnits | 680 |