NVIDIA GB300 (Grace Blackwell 300)
Overview
NVIDIA GB300 (Grace Blackwell 300) is NVIDIA's third-generation Grace Blackwell Superchip, launched 2025 H2. It pairs a B300 Ultra GPU (upgraded Blackwell) with an Arm Grace/Vera CPU via a high-speed NV-HBI interface, delivering 1 GPU + 1 CPU per Superchip. 72× GB300 units form the NVL72 rack, interconnected via NVLink 5 + ConnectX-9 1.6T networking, making it NVIDIA's flagship rack-scale data center AI product.
Generational evolution:
- GH200 (2023-Q3): Grace Hopper, H100 + 72-core Arm Grace
- GB200 (2024-Q4): Grace Blackwell, B200 + 72-core Arm Grace
- GB300 (2025 H2): B300 Ultra + Arm Vera 88-core CPU + ConnectX-9 1.6T
Core Specifications
| Item | Spec |
|---|
| Architecture | Grace Blackwell 300 Superchip |
| GPU Chip | 1× B300 Ultra (upgraded Blackwell) |
| CPU Chip | 1× Arm Vera (88 Olympus cores) |
| NV-HBI | 900 GB/s bidirectional CPU-GPU interconnect |
| CPU-GPU Coherent Memory | Unified addressing (480GB LPDDR5X on Grace + 288GB HBM3E on B300) |
| GPU Memory | 288GB HBM3E (B300 Ultra upgrade) |
| GPU Bandwidth | 10 TB/s |
| CPU Memory | 480GB LPDDR5X |
| CPU Bandwidth | 512 GB/s |
| FP4 sparse | 15 PFLOPS (B300 Ultra single GPU) |
| FP8 dense | 3.75 PFLOPS |
| BF16 dense | 1.875 PFLOPS |
| TDP (Single Superchip) | 1200 W (B300 1000W + Vera 200W) |
| Form Factor | Board-integrated (non-removable) |
| Production | 2025 H2 |
| Unit Price | ~$70,000-80,000 (Superchip module) |
GB300 NVL72 Rack
| Item | Configuration |
|---|
| Superchip Count | 72× GB300 |
| GPU Count | 72× B300 Ultra |
| CPU Count | 72× Arm Vera (88 cores × 72 = 6,336 cores) |
| Total HBM | 20.7 TB HBM3E |
| Total LPDDR5X | 34.6 TB |
| NVLink 5 Intra-domain | 130 TB/s full interconnect |
| ConnectX-9 Egress | 72× 1.6T = 115 Tb/s |
| FP4 sparse Total | 1.08 EFLOPS |
| FP8 dense Total | 270 PFLOPS |
| Rack TDP | ~120 kW (including cooling) |
| Rack Count | 8 (standard data center row) |
| Price | ~$3.3M / rack (estimated) |
Generational Comparison
| Metric | GB300 (2025 H2) | GB200 (2024-Q4) | GH200 (2023-Q3) | Note |
|---|
| GPU | B300 Ultra | B200 | H100 | Upgrade |
| CPU | Arm Vera 88-cores | Arm Grace 72-cores | Arm Grace 72-cores | New gen |
| GPU Memory | 288GB HBM3E | 192GB HBM3E | 96GB HBM3 | +50% |
| GPU Bandwidth | 10 TB/s | 8 TB/s | 3.35 TB/s | +25% |
| NV-HBI | 900 GB/s | 900 GB/s | 900 GB/s | Same |
| NVLink Interconnect | 130 TB/s | 130 TB/s | 60 TB/s | Same/2× |
| Networking | ConnectX-9 1.6T | ConnectX-8 800G | ConnectX-7 400G | 2×/4× |
| FP4 sparse | 15 PF (per GPU) | 10 PF | N/A (FP8 2 PF) | 1.5× |
| FP8 dense | 3.75 PF | 2.25 PF | 1 PF | 1.67×/3.75× |
| TDP | 1200W | 1000W | 1000W | +20% |
NVL72 vs NVL576 Rack-Scale Comparison
| Dimension | NVL72 (1 Rack) | NVL576 (8 Racks) |
|---|
| Superchip Count | 72 | 576 |
| GPU Count | 72 B300 Ultra | 576 B300 Ultra |
| CPU Count | 72 Vera | 576 Vera |
| Total HBM | 20.7 TB | 165 TB |
| FP8 dense Total | 270 PF | 2.16 EF |
| FP4 sparse Total | 1.08 EF | 8.64 EF |
| NVLink Domain | Single rack 130 TB/s | Cross-rack 130 TB/s |
| Domain Size | 72 GPUs | 576 GPUs |
| Rack TDP | 120 kW | 960 kW |
| Price | $3.3M | $26M |
GB300 NVL576 advantage: 8 racks form a single NVLink domain, 576 GPUs sharing 130 TB/s interconnect, the largest NVLink domain in the industry, critical for ultra-large LLM (trillion-parameter) training.
Arm Vera CPU 88-Core
| Dimension | Spec |
|---|
| Architecture | Arm Olympus (Armv9.4) |
| Cores | 88 cores (vs Grace 72 cores) |
| L3 Cache | Shared 256 MB |
| LPDDR5X | 480GB |
| Bandwidth | 512 GB/s |
| TDP | 200W |
| PCIe | Gen5 ×32 |
| Features | Enhanced SVE2 + hardware confidential computing |
| Generation | Armv9.4 successor to Grace v9.0 |
Vera vs Grace upgrade: Cores +22% (72→88), L3 Cache +33% (192→256MB), Memory +20% (384→480GB). The key improvement is the memory subsystem (critical for LLM inference CPU decode steps).
ConnectX-9 1.6T Networking
| Dimension | Spec |
|---|
| Speed | 1.6 Tb/s single port |
| Ports | 2-4 per Superchip |
| Protocol | InfiniBand NDR / NDR400 + RoCE v2 |
| Latency | < 0.5 μs |
| GPUDirect | GPU-NIC direct DMA |
| Congestion Control | SHARP v4 |
| PCIe | Gen6 ×16 (GB300 upgraded to Gen6) |
| 2025 Deployment | Mainstream supercomputers (ORNL Frontier successor) |
vs ConnectX-8 800G:
- Bandwidth 2×
- Latency -50%
- GPUDirect RDMA 3.0
- Supports NVLink over IB (cross-rack NVLink)
Key Features
- NVL576 domain: 8 racks, 576 GPUs sharing 130 TB/s interconnect (largest in industry)
- ConnectX-9 1.6T: 1.6 Tb/s single-port cross-rack networking
- Arm Vera 88-core: 1 per Superchip, +30% CPU performance
- FP4 15 PFLOPS: Single GPU compute (sparse), optimized for inference
- Unified memory: GPU 288GB HBM + CPU 480GB LPDDR5X coherent addressing
- Drawbacks: TDP 1200W (single Superchip), software stack CUDA-only compatible
NVL72 Use Cases
- ✅ Trillion-parameter LLM training (576 GPU domain, 130 TB/s NVLink)
- ✅ MoE model training (expert parallel + tensor parallel hybrid)
- ✅ Ultra-large-scale RLHF (576 GPU synchronous)
- ✅ Multimodal large models (video + text + image)
- ✅ AI for Science (climate, materials, life sciences)
- ❌ Small-scale inference (prohibitively expensive)
- ❌ China market (export controls)
GB300 vs AMD MI400 (2026)
| Metric | NVIDIA GB300 (2025 H2) | AMD MI400 (2026) | Difference |
|---|
| GPU | B300 Ultra (single) | MI400 (single) | - |
| Memory | 288GB HBM3E | 432GB HBM4 | MI400 +50% |
| Bandwidth | 10 TB/s | 19.6 TB/s | MI400 2× |
| FP4 dense | 7.5 PF | 40 PF | MI400 5× |
| Interconnect | NVLink 5 1.8 TB/s | UALoF 1.3 TB/s | GB300 1.4× |
| Networking | ConnectX-9 1.6T | Pensando 800G | GB300 2× |
| Software | CUDA | ROCm | NVIDIA advantage |
| TDP | 1200W (Superchip) | 1000W | MI400 -17% |
Note: MI400 is single GPU vs GB300 Superchip (including CPU). In pure GPU comparison, MI400 has clear advantages (open HBM4 + open UALoF), but NVL72 rack + ConnectX-9 remain NVIDIA's strengths.