Loading the catalog…
NVIDIA Blackwell Ultra (NVIDIA)
Blurred values are subscriber-only. Sign in to start a free trial — 2 full pages a day — or subscribe for unlimited access.
| Memory Bandwidth Tb S | 8 |
|---|---|
| Interconnect | Fifth-Generation NVLink: 1.8 TB/s; PCIe Gen6: 256 GB/s |
| Tdp W | 1400 |
| Memory Type | HBM3E |
| Gpu Memory Gb | 279 |
| Fp16 Bf16 Tflops Dense | 2500 |
| Fp8 Tflops Dense | 5000 |
| Fp16 Bf16 Tflops Sparse | 5000 |
| Fp32 Tflops | 80 |
| Int8 Tops Dense | 165 |
| Fp4 Tflops Dense | |
| Fp4 Tflops Sparse | |
| Fp64 Tflops | |
| Fp8 Tflops Sparse | |
| Int8 Tops Sparse | |
| Media Engines | |
| Product Name |
| System Total Memory Gb |
|---|
| Tf32 Tflops Dense |
|---|
| Tf32 Tflops Sparse |
|---|
| Product family | NVIDIA Blackwell Ultra |
|---|---|
| SKU | GB300 NVL72 |
| FP4 sparse inference performance (per GPU) | |
|---|---|
| HBM3E memory (per Blackwell Ultra chip) | |
| ConnectX-8 networking | |
| Attention performance vs NVIDIA Hopper | |
| Architecture | |
| NVIDIA Blackwell Ultra GPUs | |
| NVIDIA Grace CPUs | |
| AI factory output performance vs Hopper-based platforms | |
| HBM3E memory per Blackwell Ultra chip | |
| High-speed memory per rack | |
| Compute | |
| NVLink unified domain | |
| Real-time video generation performance vs Hopper | |
| Key Feature: NVIDIA Grace CPUs | |
| Key Feature: NVIDIA Blackwell Ultra GPUs | |
| Key Feature: LPDDR5X memory with error-correction code (ECC) | |
| Key Feature: HBM3E | |
| Key Feature: Fast-access memory | |
| Key Feature: NVLink domain GPU communication bandwidth | |
| Blackwell Ultra GPUs | |
| Grace CPUs | |
| CPU Cores | |
| Total FP4 Tensor Core (Sparse) | |
| Total FP4 Tensor Core (Dense) | |
| Total FP8/FP6 Tensor Core | |
| Total Fast Memory | |
| Total Memory Bandwidth | |
| Total NVLink Switch Bandwidth | |
| FP4 Tensor Core (Sparse) - per GPU | |
| FP4 Tensor Core (Dense) - per GPU | |
| FP8/FP6 Tensor Core - per GPU | |
| INT8 Tensor Core - per GPU | |
| FP16/BF16 Tensor Core - per GPU | |
| TF32 Tensor Core - per GPU | |
| FP32 - per GPU | |
| FP64/FP64 Tensor Core - per GPU | |
| GPU Memory - per GPU | |
| GPU Memory Bandwidth - per GPU | |
| Multi-Instance GPU (MIG) | |
| Decompression Engine | |
| Server Options | |
| ConnectX-8 SuperNIC network connectivity per GPU | |
| Attention-layer acceleration vs NVIDIA Blackwell GPUs | |
| AI compute FLOPS vs NVIDIA Blackwell GPUs | |
| HBM3E memory capacity vs predecessor |
