NVIDIA Blackwell, 192 GB HBM3e, SXM
The NVIDIA B200 is the Blackwell-generation flagship with 192 GB HBM3e memory and 8 TB/s of bandwidth, the largest and fastest GPU available. It runs 405B models at FP16 natively and delivers 2.25 PetaFLOPS of FP4 compute. Available on packet.ai from $3.75/GPU-hour.
Dynamic from $3.75/hr · Dedicated from $5.90/hr
The B200 pairs 5th-gen Tensor Cores with 192 GB HBM3e and 1.8 TB/s NVLink for the largest training and inference jobs.
FP4 and FP8 compute. 2.25 PFLOPS FP4 for the largest transformer models at maximum throughput.
192 GB handles 405B at FP16 natively. 8 TB/s bandwidth, the fastest memory access of any GPU.
1.8 TB/s NVLink for tight multi-GPU coupling. Scale to 1,024+ GPUs with full bandwidth coherence.
Native FP4 mixed precision with micro-scaling. 2× the throughput of H100 on transformer inference.
Dynamic, dedicated, or multi-node clusters.
Full B200 on shared infrastructure. Same 192 GB HBM3e — scheduler-isolated, no commitment.
Deploy now →Full B200 reserved exclusively for you. 99.99% SLA, zero noisy-neighbour risk.
Deploy now →Scale frontier training and inference across multiple B200 nodes with NVLink 5.0 and InfiniBand interconnect.
2.25 PFLOPS FP4 and NVLink 5.0. The fastest training GPU available.
192 GB runs 405B at FP16 natively. No sharding, no quantisation.
InfiniBand clusters with 1,024+ B200s for frontier pre-training runs.
Blackwell flagship: 192 GB HBM3e, 8 TB/s, 2.25 PFLOPS FP4. The largest and fastest GPU available.
From $3.75/GPU-hour dynamic, $5.90/hr dedicated.
Yes. 192 GB HBM3e is enough for Llama 3.1 405B at FP16 on a single card.
B200 has 2.4× the memory (192 vs 80 GB), 2.4× the bandwidth (8 vs 3.35 TB/s), and 3× the FP8 compute. It is the current frontier GPU.
Yes. Join the waitlist on packet.ai to secure allocation. Capacity is limited.
The frontier GPU — $3.75/hr dynamic, $5.90/hr dedicated.
On-demand · hourly billing · US & EU regions