15X
Real-time inference vs Hopper
GPU Compute · Blackwell
NVIDIA B200 — redefining AI and HPC with one of the most advanced GPUs yet
192GB of HBM3e and 8 TB/s of memory bandwidth deliver up to 15X faster real-time inference than NVIDIA Hopper — reserve dedicated B200 power, in limited supply.
Pricing starts at $2.55 p/hr
Specifications
Built for the largest models
Massive memory and bandwidth, with a dedicated decompression engine for data-heavy pipelines.
GPU memory 192GB HBM3e
Memory interface 4096-bit
Memory bandwidth 8 TB/s
Decompression LZ4 · Snappy · Deflate
Additional specs to confirm: FP4 / FP8 dense throughput, fifth-generation NVLink bandwidth (~1.8 TB/s per GPU) and board power (TDP).
Server configuration
NVIDIA HGX B200 GPU Server (Intel)
A fully integrated 8-GPU node — CPU, memory, storage and networking provisioned as one dedicated server.
- GPUs
- 8× NVIDIA B200 SXM (HGX B200), 1,440 GB total HBM3e, NVLink-connected
- CPU
- 2× Intel Xeon 6972P — 192 cores total
- Memory
- 1.5 TB DDR5
- Local storage
- 8× 3.84 TB NVMe SSD (~30 TB)
- GPU fabric
- 8× NVIDIA ConnectX-7 InfiniBand, up to 400 Gb/s each (3.2 Tb/s aggregate) for multi-node scaling
- Data / storage network
- 2× 100 GbE
- Management
- Dedicated 10 GbE + IPMI/BMC
Performance
A generational leap over Hopper
Compared against NVIDIA Hopper and H100, the B200 accelerates both training and inference.
3X
Training speed vs Hopper
2X
Data analytics vs H100
Also up to 6X faster query performance for data analytics versus traditional CPUs.
Use cases
What the B200 powers
From frontier-model training to real-time production inference and HPC.
Generative AI
Train and serve large language and multi-modal models with headroom for the biggest context windows.
Training & inference
End-to-end machine learning — from large-scale training runs to low-latency production inference.
Real-time inference
Computer vision, NLP and recommendation systems served in real time at scale.
Data analytics
Up to 6X faster query performance versus traditional CPUs for analytics and predictive modeling.
High-performance computing
Simulation, modeling and scientific workloads that demand maximum throughput.
Large-scale data processing
A dedicated decompression engine accelerates data pipelines feeding your GPUs.
Reserve your NVIDIA B200 access
Supply is limited. Talk to our team to secure dedicated B200 capacity for your training and inference workloads.