NVIDIA Blackwell B200 GPUs Reach General Availability on Amazon SageMaker AI Training
Summary
- • P6-B200 instances with 8 NVIDIA Blackwell B200 GPUs are now GA on Amazon SageMaker AI
- • B200 packs 180 GB HBM with NVLink 5 delivering 1.8 TB/s GPU-to-GPU bandwidth
- • Expanded memory lets teams train 64B-parameter models on a single node, eliminating multi-node overhead
- • AWS Flexible Training Plan offers predictable Blackwell capacity booking with cost management
Details
P6-B200 GA on SageMaker
P6-B200 instances with 8 NVIDIA Blackwell B200 GPUs are now generally available for Amazon SageMaker AI Training jobs
B200 memory: 180 GB HBM
NVIDIA B200 GPU carries 180 GB HBM; the B300 variant carries 268 GB HBM — a major increase over prior-generation GPUs
NVLink 5 at 1.8 TB/s
NVLink 5 provides up to 1.8 TB/s bidirectional GPU-to-GPU bandwidth, reducing inter-GPU communication bottlenecks in distributed training
Single-node viability for large models
Models previously requiring multi-node setups can now run on a single 8-GPU Blackwell node, cutting networking overhead and iteration cycle time
PyTorch FSDP recommended
AWS recommends PyTorch Fully Sharded Data Parallel (FSDP) as the distributed training framework for 1B–64B parameter models on Blackwell
Flexible Training Plan
AWS Flexible Training Plan enables predictable Blackwell capacity booking with automated resource management and cost controls
Throughput and sequence gains
Higher memory reduces gradient sync steps and enables longer sequence lengths, improving throughput for long-range dependency tasks
Technical details sourced from AWS ML Blog official product guidance post.
What This Means
The general availability of NVIDIA Blackwell B200 GPUs on SageMaker marks a meaningful step-change in what AWS customers can train within fully managed infrastructure. The B200's memory jump from prior GPU generations directly addresses the most common practical constraint in large-model training: running out of memory and being forced into expensive multi-node configurations. For AI teams using AWS, this reduces both infrastructure complexity and cost for models in the 1B–64B parameter range. It also signals AWS's ongoing effort to keep pace with Google Cloud and Azure on frontier GPU availability for enterprise AI workloads, with managed access being the key differentiator over raw hardware provisioning.