B200 SXM 192GB
The Blackwell training part. 192 GB HBM3e, 8 TB/s, NVLink 5 at 1.8 TB/s per GPU.
Allocation. Lead time 4 to 8 weeks.
Volume pricing
| Units | Discount | Unit price |
|---|---|---|
| 1 unit | List | $45,000 |
| 2 to 7 units | 3% off | $43,650 |
| 8 to 31 units | 6% off | $42,300 |
| 32+ units | 10% off | $40,500 |
Unit price at 1: $45,000
Quote only. Price shown is indicative; we confirm price and date in writing before you pay.
Request quote
The full sheet.
Figures from the manufacturer datasheet. Price is an indicative street price, shown as from.
- VRAM
- 192 GB
- Memory
- HBM3e
- Bandwidth
- 8 TB/s
- FP8
- 9 PFLOPS
- FP4
- 18 PFLOPS
- Board power
- 1,000 W
- Form factor
- SXM
- NVLink
- Yes
- Architecture
- Blackwell
- GPUs
- 1
- Per GB of VRAM
- $234.38
- Availability
- Allocation, 4 to 8 weeks
- Sale
- Quote only
- SKU
- B200-SXM
What it is for.
Workloads, sizing and what ships in the box.
192 GB of HBM3e at 8 TB/s per GPU.
The Blackwell training part. 192 GB of HBM3e at 8 TB/s, 9 PFLOPS FP8 and 18 PFLOPS FP4, NVLink 5 at 1.8 TB/s per GPU. Sold in sets of eight for HGX baseboards, or as single spares matched to your serials. Allocation, 4 to 8 weeks, dated in writing. From $45,000, about $234 per GB of VRAM.
Good for
- Llama 3.1 70B at FP16 (168 GB with headroom) on a single GPU.
- Mixtral 8x22B (141B) at FP8 (169 GB) on one GPU.
- Llama 3.1 405B at FP8 (486 GB) across a set of 4, or FP16 across a set of 8.
Fit rule: FP16 weights take 2 bytes per parameter, FP8 1 byte, 4-bit 0.5 bytes, plus 20% headroom for KV cache at short context.
Pairs with
- HGX B300 8-GPU, the next step up in the same family.
- H200 SXM 141GB, for inference nodes that ship sooner.
Sold in sets of 8 for HGX baseboards or as spares with serial match.
How much VRAM per box.
Card count is capped by slots and by the chassis power budget.
Parts bought alongside.
Same workload, a different trade between VRAM, bandwidth and price.