288 GB
288 GB · 8,000 GB/s
Verified 2026-08-30
datacenter-gpu · introduced 2025-03
NVIDIA B300 SXM
HBM3e · Dedicated device memory
- Configurations
- 1
- Up to
- 288 GB usable
- Peak
- 8,000 GB/s
Permanent config IDs
Available configurations
Capacity, bandwidth, price, and availability belong to each SKU—not the device family.
| Configuration | Usable memory | Bandwidth | Price | Status |
|---|---|---|---|---|
| NVIDIA B300 SXM 288 GB | 288 GB | 8,000 GB/s | Unknown | available |
Sourced bandwidth · 60 configurations
Bandwidth ladder
Every sourced configuration, ranked by peak memory bandwidth. 100% is NVIDIA B300 SXM's fastest configuration at 8,000 GB/s.
Read before comparing
Notes and constraints
- Blackwell Ultra. Sold only as one GPU of an eight-GPU HGX B300 baseboard, a DGX B300 system, or a GB300 rack; there is no standalone B300 card.
- Capacity and bandwidth are the per-GPU figures NVIDIA publishes; a baseboard holds eight such GPUs (2.30 TB, up to 64 TB/s aggregate) across eight separate memory pools.
- Compute is the eight-GPU HGX B300 tensor-core specification divided by eight, with NVIDIA's sparse figures halved per that table's dense footnote. FP4 is omitted because the schema has no FP4 precision.
- Blackwell Ultra trades INT8 throughput for FP4: NVIDIA's published INT8 rate is far below the B200's, which is not a transcription error.
- NVIDIA publishes no per-module TDP for the HGX/DGX B300 SXM, so power is left unset.
Audit trail
Primary sources
Every required capacity and bandwidth value includes a source and verification date.