NVIDIA GB200 (Grace Blackwell 200, 2024-Q4)
Product Overview
NVIDIA GB200 (Grace Blackwell 200) is NVIDIA's second-generation Grace Blackwell Superchip, entering mass production Q4 2024. It pairs a B200 GPU with an Arm Grace 72-core CPU via NV-HBI 900 GB/s high-speed interconnect — 1 Superchip = 1 GPU + 1 CPU. 72 GB200 units form an NVL72 rack delivering 1 EFLOPS FP4 sparse, NVLink 5 130 TB/s interconnect, and ConnectX-8 800G networking — NVIDIA's 2024–2025 data center AI flagship.
Generational comparison:
- GH200 (2023-Q3): Grace + H100, FP8 1 PF sparse, NVLink 4 60 TB/s
- GB200 (2024-Q4): Grace + B200, FP4 10 PF sparse, NVLink 5 130 TB/s, 1 EFLOPS / NVL72
- GB300 (2025 H2): Vera 88-core + B300 Ultra + ConnectX-9 1.6T, 1.08 EFLOPS / NVL72 (page available)
Core Specifications
| Parameter | Value |
|---|
| Architecture | Grace Blackwell 200 Superchip |
| GPU Chip | 1× B200 (Blackwell) |
| CPU Chip | 1× Arm Grace (72-core Neoverse V2) |
| NV-HBI | 900 GB/s bidirectional CPU-GPU interconnect |
| CPU-GPU Coherent Memory | Unified addressing |
| GPU Memory | 192GB HBM3E |
| GPU Bandwidth | 8 TB/s |
| CPU Memory | 480GB LPDDR5X (on Grace) |
| CPU Bandwidth | 512 GB/s |
| FP4 Sparse | 20 PFLOPS (B200 single GPU ×2 Superchip) |
| FP8 Dense | 4.5 PFLOPS |
| BF16 Dense | 2.25 PFLOPS |
| TDP (per Superchip) | 1000W |
| Form Factor | Board-integrated (non-removable) |
| Mass Production | 2024-Q4 |
| Unit Price | ~$60,000–70,000 (Superchip module) |
GB200 NVL72 Rack
| Parameter | Configuration |
|---|
| Superchip Count | 72× GB200 |
| GPU Count | 72× B200 |
| CPU Count | 72× Arm Grace (72 cores × 72 = 5,184 cores) |
| Total HBM | 13.8 TB HBM3E |
| Total LPDDR5X | 34.6 TB |
| NVLink 5 Domain | 130 TB/s fully connected |
| ConnectX-8 Egress | 72× 800G = 57.6 Tb/s |
| FP4 Sparse Total | 1,440 PFLOPS (72× 20 PFLOPS) |
| FP8 Dense Total | 324 PFLOPS (72× 4.5 PFLOPS) |
| Rack TDP | ~120 kW |
| Rack Count | 8 (standard data center row) |
| Price | ~$3M / rack (estimated) |
GB200 NVL576 (8-Rack)
| Parameter | Configuration |
|---|
| Superchip Count | 576× GB200 |
| GPU Count | 576× B200 |
| NVLink 5 Domain | Cross-rack 130 TB/s |
| Total HBM | 110 TB |
| FP4 Sparse Total | 11.52 EFLOPS (576× 20 PFLOPS) |
| FP8 Dense Total | 2.6 EFLOPS (576× 4.5 PFLOPS) |
| Rack TDP | 960 kW |
| Price | ~$24M |
GB200 NVL576 advantage: 8 racks, 576 GPUs sharing a single 130 TB/s NVLink domain — the largest single AI compute domain in the industry as of 2024, critical for trillion-parameter LLM training.
GH200 → GB200 → GB300 Comparison
| Metric | GH200 (2023-Q3) | GB200 (2024-Q4) | GB300 (2025 H2) |
|---|
| GPU | H100 | B200 | B300 Ultra |
| CPU | Grace 72-core | Grace 72-core | Vera 88-core |
| GPU Memory | 96GB HBM3 | 192GB HBM3E | 288GB HBM3E |
| GPU Bandwidth | 3.35 TB/s | 8 TB/s | 10 TB/s |
| NVLink Domain | 60 TB/s | 130 TB/s | 130 TB/s |
| Networking | ConnectX-7 400G | ConnectX-8 800G | ConnectX-9 1.6T |
| FP4 Sparse | N/A (FP8 2 PF) | 10 PF | 15 PF |
| FP8 Dense | 1 PF | 2.25 PF | 3.75 PF |
| TDP (Superchip) | 1000W | 1000W | 1200W |
ConnectX-8 800G Networking
| Dimension | Specification |
|---|
| Speed | 800 Gb/s single port (2× ConnectX-7) |
| Port Count | 2–4 per Superchip |
| Protocol | InfiniBand NDR / RoCE v2 |
| Latency | < 0.5 μs |
| GPUDirect | GPU-NIC direct DMA |
| Congestion Control | SHARP v3 |
| 2024 Deployments | ORNL Aurora successor, CSCS Alps, EuroHPC |
ConnectX-8 upgrade: 2× speed (400G → 800G), GPUDirect RDMA 3.0, supports NVLink over IB (cross-rack NVLink).
Arm Grace 72-Core
| Dimension | Specification |
|---|
| Architecture | Arm Neoverse V2 |
| Core Count | 72 cores |
| L3 Cache | Shared 192MB |
| LPDDR5X | 480GB |
| Bandwidth | 512 GB/s |
| TDP | 200W (CPU only) |
| PCIe | Gen5 ×32 |
| Features | SVE2 enhanced |
Grace vs Vera upgrade: Vera is Grace's successor (88-core + 256MB L3 + 480GB LPDDR5X). GB200 still uses Grace 72-core; only GB300 upgrades to Vera.
GB200 Use Cases
- ✅ Trillion-parameter LLM training (NVL576 domain, 130 TB/s NVLink)
- ✅ MoE model training (expert parallel + tensor parallel)
- ✅ Ultra-large-scale RLHF (576 GPU synchronized)
- ✅ Multimodal large models (video + text + image)
- ✅ AI for Science (climate, materials, life sciences)
- ✅ Cloud AI services (CoreWeave, Lambda, OVHcloud)
- ❌ Small-scale inference (cost-prohibitive)
- ❌ China market (export controls)
GB200 Customers
- Meta: Llama 4 / 5 training (>$10B order)
- Microsoft Azure: OpenAI GPT-5 + Copilot
- Google Cloud: Gemini 1.5 / 2.0
- AWS: Anthropic Claude 4 + Bedrock
- CoreWeave: 30K+ GB200 deployed (H1 2025)
- xAI Grok 3: Colossus cluster 100K+ GB200
- Oracle Cloud: OCI deployment
- Lambda Labs: Lambda 1-Click Cluster
GB200 vs GB300
| Metric | GB200 (2024-Q4) | GB300 (2025 H2) |
|---|
| GPU | B200 | B300 Ultra |
| CPU | Grace 72-core | Vera 88-core |
| GPU Memory | 192GB HBM3E | 288GB HBM3E |
| GPU Bandwidth | 8 TB/s | 10 TB/s |
| Networking | ConnectX-8 800G | ConnectX-9 1.6T |
| FP4 Sparse | 10 PF | 15 PF |
| FP8 Dense | 2.25 PF | 3.75 PF |
| TDP (Superchip) | 1000W | 1200W |
GB300 upgrade: GPU memory +50% (192→288GB), compute +50% (FP8 2.25→3.75 PF), networking 2× (800G→1.6T), comparable price.
Key Features
- NVL72 domain: 72 GPUs sharing 130 TB/s NVLink
- NVL576 domain: 576 GPUs across 8 racks NVLink
- ConnectX-8 800G: 800G single port
- Arm Grace 72-core: CPU + GPU unified memory addressing
- FP4 10 PFLOPS: Inference optimized
- Drawbacks: 1000W TDP, export controls, CUDA-only software