B200 (HGX) specs
NVIDIA · Shipping now · first available Dec 2024 (estimated). Dense compute, memory, bandwidth and power from the vendor's documents; each precision stays separate.
All accelerators · Compare with GB200 NVL72 (per GPU)
Specifications
| Dense BF16 · FP32 accumulate | 2,250 TFLOPS |
|---|---|
| Dense FP8 · FP32 accumulate | 4,500 TFLOPS |
| Dense FP4 · FP32 accumulate | 9,000 TFLOPS |
| Memory | 180 GB · HBM |
| Memory bandwidth | 7.7 TB/s |
| Maximum chip power | 1,000 W |
| Interconnect | NVLink 5 · 1.8 TB/s |
| Process node | TSMC 4NP |
| First availability | Dec 2024est — estimated / secondary-source figure |
| Availability | Shipping |
| Est. purchase price | ~$35,000est — estimated / secondary-source figureMidpoint of the range given for Blackwell |
Dual-die Blackwell; first NVIDIA part with FP4 tensor formats.
Dense tensor rates with FP32 accumulation, never sparse or FP16-accumulate headline figures. Memory is in decimal GB. The est marker flags a value its vendor-document review did not confirm.
Rent it
$6.00/GPU-hrCheapest on-demand · Hyperstack · observed
7 providers: AWS, Hyperstack, Lambda, Nebius, Oracle OCI, Together AI, Verda. Quotes checked through .
B200 rental prices from every provider →Value ratios
| $ per dense BF16 PFLOP-hour | $2.67 |
|---|---|
| Dense BF16 TFLOPS per watt | 2.25 |
| Bandwidth bytes per 1,000 BF16 FLOP | 3.42 |
| $ per dense FP8 PFLOP-hour | $1.33 |
| Dense FP8 TFLOPS per watt | 4.50 |
| Bandwidth bytes per 1,000 FP8 FLOP | 1.71 |
| $ per GB of memory per hour | $0.033 |
Prices divide the cheapest on-demand list price; theoretical ratios do not measure application speed.
Sources
Specifications reviewed . HGX B200 column: 7.7 TB/s per GPU (62 TB/s per board) replaces 8 TB/s, which DGX B200 pages still list. Per-GPU rates are derived from the eight-GPU table.
Purchase price: CNBC, quoting NVIDIA's CEO ↗ (2024-03-19): “will cost between $30,000 and $40,000 per unit”
Rental price: Hyperstack published pricing ↗ (NVIDIA B200 192GB; maximum per GPU: 31 pCPUs and 256 GB RAM; actual allocation not specified; published per GPU-hour).