Skip to main content

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

ItemSpec
ArchitectureGrace Blackwell 300 Superchip
GPU Chip1× B300 Ultra (upgraded Blackwell)
CPU Chip1× Arm Vera (88 Olympus cores)
NV-HBI900 GB/s bidirectional CPU-GPU interconnect
CPU-GPU Coherent MemoryUnified addressing (480GB LPDDR5X on Grace + 288GB HBM3E on B300)
GPU Memory288GB HBM3E (B300 Ultra upgrade)
GPU Bandwidth10 TB/s
CPU Memory480GB LPDDR5X
CPU Bandwidth512 GB/s
FP4 sparse15 PFLOPS (B300 Ultra single GPU)
FP8 dense3.75 PFLOPS
BF16 dense1.875 PFLOPS
TDP (Single Superchip)1200 W (B300 1000W + Vera 200W)
Form FactorBoard-integrated (non-removable)
Production2025 H2
Unit Price~$70,000-80,000 (Superchip module)

GB300 NVL72 Rack

ItemConfiguration
Superchip Count72× GB300
GPU Count72× B300 Ultra
CPU Count72× Arm Vera (88 cores × 72 = 6,336 cores)
Total HBM20.7 TB HBM3E
Total LPDDR5X34.6 TB
NVLink 5 Intra-domain130 TB/s full interconnect
ConnectX-9 Egress72× 1.6T = 115 Tb/s
FP4 sparse Total1.08 EFLOPS
FP8 dense Total270 PFLOPS
Rack TDP~120 kW (including cooling)
Rack Count8 (standard data center row)
Price~$3.3M / rack (estimated)

Generational Comparison

MetricGB300 (2025 H2)GB200 (2024-Q4)GH200 (2023-Q3)Note
GPUB300 UltraB200H100Upgrade
CPUArm Vera 88-coresArm Grace 72-coresArm Grace 72-coresNew gen
GPU Memory288GB HBM3E192GB HBM3E96GB HBM3+50%
GPU Bandwidth10 TB/s8 TB/s3.35 TB/s+25%
NV-HBI900 GB/s900 GB/s900 GB/sSame
NVLink Interconnect130 TB/s130 TB/s60 TB/sSame/2×
NetworkingConnectX-9 1.6TConnectX-8 800GConnectX-7 400G2×/4×
FP4 sparse15 PF (per GPU)10 PFN/A (FP8 2 PF)1.5×
FP8 dense3.75 PF2.25 PF1 PF1.67×/3.75×
TDP1200W1000W1000W+20%

NVL72 vs NVL576 Rack-Scale Comparison

DimensionNVL72 (1 Rack)NVL576 (8 Racks)
Superchip Count72576
GPU Count72 B300 Ultra576 B300 Ultra
CPU Count72 Vera576 Vera
Total HBM20.7 TB165 TB
FP8 dense Total270 PF2.16 EF
FP4 sparse Total1.08 EF8.64 EF
NVLink DomainSingle rack 130 TB/sCross-rack 130 TB/s
Domain Size72 GPUs576 GPUs
Rack TDP120 kW960 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

DimensionSpec
ArchitectureArm Olympus (Armv9.4)
Cores88 cores (vs Grace 72 cores)
L3 CacheShared 256 MB
LPDDR5X480GB
Bandwidth512 GB/s
TDP200W
PCIeGen5 ×32
FeaturesEnhanced SVE2 + hardware confidential computing
GenerationArmv9.4 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

DimensionSpec
Speed1.6 Tb/s single port
Ports2-4 per Superchip
ProtocolInfiniBand NDR / NDR400 + RoCE v2
Latency< 0.5 μs
GPUDirectGPU-NIC direct DMA
Congestion ControlSHARP v4
PCIeGen6 ×16 (GB300 upgraded to Gen6)
2025 DeploymentMainstream supercomputers (ORNL Frontier successor)

vs ConnectX-8 800G:

  • Bandwidth 2×
  • Latency -50%
  • GPUDirect RDMA 3.0
  • Supports NVLink over IB (cross-rack NVLink)

Vendor Information

ItemDetail
CompanyNVIDIA Corporation
Product Pagehttps://www.nvidia.com/en-us/data-center/grace-blackwell/
CEOJensen Huang
FoundryTSMC 4NP (B300 Ultra) + TSMC N3 (Vera)
2025 H2 ProductionYes
2026 RoadmapRubin + Vera Rubin (RV200)
PriceSuperchip ~$70-80K, NVL72 rack ~$3.3M

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)

MetricNVIDIA GB300 (2025 H2)AMD MI400 (2026)Difference
GPUB300 Ultra (single)MI400 (single)-
Memory288GB HBM3E432GB HBM4MI400 +50%
Bandwidth10 TB/s19.6 TB/sMI400 2×
FP4 dense7.5 PF40 PFMI400 5×
InterconnectNVLink 5 1.8 TB/sUALoF 1.3 TB/sGB300 1.4×
NetworkingConnectX-9 1.6TPensando 800GGB300 2×
SoftwareCUDAROCmNVIDIA advantage
TDP1200W (Superchip)1000WMI400 -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.