B200 Specifications
Blackwell, dual-reticle design (two dies acting as one GPU). Adds FP4 for inference. 1 kW means liquid cooling in most deployments. Figures here are per package: B200 is two dies acting as one GPU, so per-die sources will show roughly half the SMs, transistors and bandwidth.
Buy it for
Frontier-scale training and FP4 inference. The only part here with hardware FP4.
Look elsewhere if
You cannot supply liquid cooling or 1 kW per socket. Air-cooled racks generally cannot take this.
Will it fit your system?
This is not a card you can install
SXM is a socketed module that mounts to an OEM baseboard, not a PCIe slot. You buy it inside a complete system — it cannot go into an existing desktop or a standard server under any circumstances.
- Form factor
- SXM
- Cooling
- passive
- Power connectors
- SXM socket — liquid-cooled baseboard required
- Board power
- 1000 W
- Interface
- PCIe Gen 6
Supported formats & features
What the B200 does and does not support in hardware.
Hardware FP4 — roughly doubles inference throughput over FP8
E4M3 and E5M2 — the standard format for modern inference
Wider exponent than FP16 — fewer overflow issues when training
Drop-in FP32 speedup with no code changes
Quantised inference supported
Not yet confirmed against vendor documentation.
Per-layer automatic FP8 precision management
1800 GB/s GPU-to-GPU
Compute-only die — no ray tracing hardware
Corrects single-bit errors — expected for long training runs
What fits in memory
Largest model that fits on one B200, assuming ~20% of the 192 GB is reserved for KV cache and activations.
| Model | FP16 / BF16 | FP8 | INT4 / FP4 |
|---|---|---|---|
| Llama 8B | Fits | Fits | Fits |
| Llama 70B | Fits | Fits | Fits |
| Mixtral 8x22B | 2× GPUs | Fits | Fits |
| Llama 405B | 6× GPUs | 3× GPUs | 2× GPUs |
Weights only. Long contexts, large batch sizes and optimiser state during training all need substantially more — full fine-tuning typically needs 3–4× the weight memory.
Workload character
Bandwidth and compute are matched closely enough that neither dominates across mixed training and inference.
Cost to run
Electricity for one card at 100% duty, including cooling and PSU overhead (12,264 kWh/year).
At the US average rate, three years of continuous operation costs $4,415 — roughly 11% of the card's $40,000 launch price, before any cooling or hosting fees.
Where it sits in the line
A newer part existing does not make this one a bad buy — it makes it cheaper. What matters is price per unit of the throughput you actually need.
Availability & market
Allocation-constrained and sold through system integrators. Expect lead times rather than shelf stock.
Memory
- Capacity
- 192 GB
- Type
- HBM3e
- Bandwidth
- 8,000 GB/s
- Bus width
- 8,192-bit
- GB per TFLOPMemory relative to compute — higher favours large models
- 0.085
Compute
- FP64
- 40 TFLOPS
- FP32
- 80 TFLOPS
- TF32 dense
- 1,125 TFLOPS
- FP16 dense
- 2,250 TFLOPS
- FP16 sparseRequires structured sparsity; rarely reached in practice
- 4,500 TFLOPS
- FP8 dense
- 4,500 TFLOPS
- INT8 dense
- 4,500 TOPS
Silicon
- Codename
- GB100
- Process node
- TSMC 4NP
- Transistors
- 208B
- Die size
- 1,600 mm²
- CUDA cores
- —
- Tensor cores
- —
- SM count
- 264
- Boost clock
- —
Platform & power
- TDP
- 1,000 W
- TFLOPS per wattDense FP16 ÷ TDP — drives your power bill at scale
- 2.25
- Form factor
- SXM
- Cooling
- passive
- Interconnect
- NVLink 5.0
- Interconnect speed
- 1,800 GB/s
- PCIe generation
- Gen 6
- MIG
- Yes — up to 7 instances
- Launch
- 2025
- MSRP at launch
- $40,000
B200 — common questions
Answered from the specifications above. Questions this card's data cannot answer with confidence are left out rather than guessed.
- How much VRAM does the B200 have?
- The B200 has 192 GB of HBM3e memory with 8,000 GB/s of bandwidth. In practice plan on roughly 153 GB for weights: KV cache, activations and fragmentation take the rest, and long contexts or large batches take considerably more.
- What size model can the B200 run?
- On a single B200, about 76B parameters at FP16, or roughly 307B quantised to FP4. Anything larger has to be split across several GPUs. These are weight-only figures — a long context window can consume tens of gigabytes on its own.
- Can you put the B200 in a desktop PC?
- No. The B200 is a SXM module, not a PCIe card. It bolts to an OEM baseboard and is only sold inside a complete system, so there is no way to install one in a desktop or a standard server.
- How much power does the B200 use?
- The B200 is rated at 1000 W board power. It draws that from the baseboard rather than from PSU cables, so there is nothing to wire up yourself.
- How much does it cost to run the B200?
- At the US commercial average of $0.12/kWh, running one B200 continuously costs about $1,472 a year in electricity. That figure includes cooling and power-supply overhead at a factor of 1.4 over board power, and over three years comes to roughly 11% of what the card cost at launch.
- Does the B200 support MIG?
- Yes. The B200 can be partitioned into up to 7 Multi-Instance GPU slices, each with isolated memory and compute. This is what makes it practical to share one card across several users or tenants without them affecting each other's latency.
- Does the B200 support FP8?
- Yes. The B200 has hardware FP8 tensor support, so FP8 weights and activations run at full tensor-core rate — 4,500 dense TFLOPS. This is the format most current inference servers default to for large models.
- Does the B200 support FP4?
- Yes. The B200 supports 4-bit floating point in hardware, which roughly doubles throughput again over FP8 and halves the memory a model occupies.
- Does the B200 support INT8?
- Yes. The B200 supports INT8 acceleration at 4,500 dense TOPS. INT8 remains the most widely supported quantised inference format across frameworks.
- Does the B200 support BF16?
- Yes. The B200 supports bfloat16. Its wider exponent range makes training far less prone to the overflow that plagues FP16, which is why most current training recipes assume it.
- Does the B200 have NVLink?
- Yes. The B200 supports NVLink (NVLink 5.0) at 1,800 GB/s, so GPU-to-GPU traffic bypasses PCIe. This matters for tensor-parallel serving, where every token crosses the link.
- Does the B200 have RT cores?
- No. The B200 is a compute-only part with no ray-tracing hardware. For rendering or virtual workstations, look at the L40S or an RTX workstation card instead.
- Does the B200 have ECC memory?
- Yes. The B200 has ECC, so single-bit memory errors are corrected rather than silently corrupting a long run.
- What replaced the B200?
- The B300 (2026) is its successor, with 288 GB of HBM3e at 8,000 GB/s. At the same dense FP16 throughput, what changed is memory rather than compute. On the used market that gap is what sets a fair price for the B200.
- Is the B200 worth buying?
- Buy it for: Frontier-scale training and FP4 inference. The only part here with hardware FP4. Look elsewhere if: You cannot supply liquid cooling or 1 kW per socket. Air-cooled racks generally cannot take this.