Calculate the TPS of the B200 on local AI models

NVIDIA 180 GB HBM3e 8,000 GB/s January 2024

Every model in our catalogue assessed against this card at the context length and minimum quality you choose. Speed is an estimate for a single request, calculated from this card's memory bandwidth and the size of each model once compressed.

Calculated for this card

633 of 679 models it can run

Largest model it holds

GLM-4.7

358B · Q3_K_M · 142 tok/s

Fastest model

Gemma 3 QAT 1B

3,388 tok/s · 1B

What AI models can a B200 run?

Set the inputs, read the answer

More context means more memory for the conversation cache. Speed is for a fresh conversation and does not change with this setting.

Hides models that would only fit by being compressed below this point.

633 models match

Calculating
Quantisation Fit
3,388 tok/s

2,880–4,066

Gemma 3 1B 1B Mar 2025 1.8 GB 33k tokens Q8_0 Comfortable
3,388 tok/s

2,880–4,066

Gemma 3 QAT 1B 1B Apr 2025 1.8 GB 33k tokens Q8_0 Comfortable
3,388 tok/s

2,033–5,421 · low confidence

HGRN 1B (WT 103) 1B Nov 2023 1.8 GB 131k tokens ? Q8_0 Comfortable
3,388 tok/s

2,033–5,421 · low confidence

LLama 3..2 Typhoon 2 1B 1B Dec 2024 1.8 GB 131k tokens ? Q8_0 Comfortable
3,388 tok/s

2,033–5,421 · low confidence

OLMo-1B 1B Feb 2024 1.8 GB 131k tokens ? Q8_0 Comfortable
3,388 tok/s

2,033–5,421 · low confidence

Pythia-1b 1B Apr 2023 1.8 GB 131k tokens ? Q8_0 Comfortable
3,137 tok/s

1,882–5,020 · low confidence

OpenELM-1.1B 1.1B May 2024 1.9 GB 131k tokens ? Q8_0 Comfortable
3,080 tok/s

1,848–4,928 · low confidence

DeciCoder-1B 1.1B Aug 2023 1.9 GB 131k tokens ? Q8_0 Comfortable
3,080 tok/s

1,848–4,928 · low confidence

SantaCoder 1.1B Jan 2023 1.9 GB 131k tokens ? Q8_0 Comfortable
3,080 tok/s

1,848–4,928 · low confidence

TinyLlama-1.1B (1T token checkpoint) 1.1B Oct 2023 1.9 GB 131k tokens ? Q8_0 Comfortable
3,080 tok/s

1,848–4,928 · low confidence

TinyLlama-1.1B (3T token checkpoint) 1.1B Oct 2023 1.9 GB 131k tokens ? Q8_0 Comfortable
2,824 tok/s

1,694–4,518 · low confidence

EXAONE 4.0 (1.2B) 1.2B Jul 2025 2.0 GB 131k tokens ? Q8_0 Comfortable
2,824 tok/s

1,694–4,518 · low confidence

MinerU2.5 1.2B Sep 2025 2.0 GB 131k tokens ? Q8_0 Comfortable
2,824 tok/s

1,694–4,518 · low confidence

Pleias 1.0 1.2B 1.2B Dec 2024 2.0 GB 131k tokens ? Q8_0 Comfortable
2,824 tok/s

1,694–4,518 · low confidence

Pleias-RAG-1B 1.2B Apr 2025 2.0 GB 131k tokens ? Q8_0 Comfortable
2,755 tok/s

2,341–3,306

Llama 3.2 1B 1.2B Sep 2024 2.2 GB 131k tokens Q8_0 Comfortable
2,716 tok/s

1,630–4,346 · low confidence

MiniCPM-1.2B 1.2B Jun 2024 2.0 GB 131k tokens ? Q8_0 Comfortable
2,606 tok/s

1,564–4,170 · low confidence

DeepSeek Coder 1.3B 1.3B Jan 2024 2.1 GB 131k tokens ? Q8_0 Comfortable
2,606 tok/s

1,564–4,170 · low confidence

DeepSeek-VL-1.3B 1.3B Mar 2024 2.1 GB 131k tokens ? Q8_0 Comfortable
2,606 tok/s

1,564–4,170 · low confidence

DigiRL 1.3B Jun 2024 2.1 GB 131k tokens ? Q8_0 Comfortable
2,606 tok/s

1,564–4,170 · low confidence

GLA Transformer 1.3B 1.3B Aug 2024 2.1 GB 131k tokens ? Q8_0 Comfortable
2,606 tok/s

1,564–4,170 · low confidence

Janus 1.3B 1.3B Oct 2024 2.1 GB 131k tokens ? Q8_0 Comfortable
2,606 tok/s

1,564–4,170 · low confidence

Kosmos-2.5 1.3B Aug 2024 2.1 GB 131k tokens ? Q8_0 Comfortable
2,606 tok/s

1,564–4,170 · low confidence

Otter 1.3B May 2023 2.1 GB 131k tokens ? Q8_0 Comfortable
2,606 tok/s

1,564–4,170 · low confidence

Phi-1 1.3B Oct 2023 2.1 GB 131k tokens ? Q8_0 Comfortable

Speeds are estimates for a single request — one conversation at a time — calculated from memory bandwidth, model size and quantisation. Real throughput varies with the inference runtime and its version. Figures published by hardware vendors measure many simultaneous requests and are much higher.

On record

B200 full specification

Everything on record for this board, ordered by how much it bears on running a language model rather than by how a spec sheet would list it. Memory comes first because it decides the outcome; the rest is context.

Memory

The two specifications that decide what this card can run and how quickly. Capacity sets which models fit; bandwidth sets how many tokens per second they produce once they do.

Memory size
180 GB
Memory bandwidth
8,000 GB/s
Memory type
HBM3e
Memory bus width
8,192 bit
Memory clock
1.95 GHz

The chip

Which processor is on the board and how it was manufactured. A smaller process size generally means more performance for the same power.

Graphics processor
GB100
Architecture
Blackwell
Generation
Server Blackwell(Bxx)
Foundry
TSMC
Process size
5 nm
Transistors
104 billion
Released
1 January 2024

Clock speeds

How fast the processor runs. Worth far less here than on a gaming benchmark: generating text is limited by memory bandwidth, so a higher clock barely moves the result.

Base clock
700 MHz
Boost clock
1.97 GHz

Processing units

What the chip contains. These drive graphics performance and matter mainly for processing a long prompt rather than for producing the answer.

Shading units
18,944
Texture mapping units
592
Render output units
24
Streaming multiprocessors
148
Tensor cores
592
L1 cache
250 KB
L2 cache
50 MB

Theoretical performance

Peak arithmetic rates published for the board. These are ceilings that no real workload reaches, and generating text reaches a small fraction of them because it is limited by memory rather than arithmetic.

Half precision (FP16)
1,191.2 TFLOPS
Single precision (FP32)
74.5 TFLOPS
Double precision (FP64)
37.2 TFLOPS
Pixel rate
47 GPixel/s
Texture rate
1,163 GTexel/s

The board

What it takes to physically install and power the card — the practical constraints that decide whether it fits the machine you already own.

Power draw (TDP)
1,000 W
Suggested power supply
1,400 W
Bus interface
PCIe 5.0 x16
Slot width
SXM Module

Software support

Which graphics and compute interfaces the card supports. CUDA compute capability is the one that bears on inference: below 7.0 there are no tensor cores, and modern inference software falls back to slower code paths.

CUDA compute capability
10.0
OpenCL
3.0

Listings

Where to buy a B200

No vendor is currently listing this card. Listings come from vendors who publish them here directly — browse the vendor directory to see who is selling what.

What the numbers mean

Memory: the specification that decides everything

Memory

180 GB

Bandwidth

8,000 GB/s

Largest model

GLM-4.7

With 180 GB of HBM3e, the B200 is in the class of hardware that holds the largest open-weight models without splitting them across machines. Roughly 162 GB of that is reachable by an inference runtime once the driver takes its share.

Its 8,000 GB/s across a 8,192-bit bus is at the top of what exists. Since each token means reading the whole model out of memory once, that translates almost directly into generation speed — this card is bandwidth-rich enough that model size stops being the limiting factor long before the bus does.

Bandwidth is clock times bus width, and this card clocks its memory at 1.95 GHz. Both halves matter, and neither is visible in a gaming benchmark.

The practical ceiling is GLM-4.7 at 358B, held at Q3_K_M and running at roughly 142 tokens per second.

The chip and how it was built

The B200 is built on the GB100 graphics processor, using NVIDIA's Blackwell architecture, as part of the Server Blackwell(Bxx) generation.

The chip is manufactured by TSMC, on a 5 nm process, holding 104 billion transistors. A smaller process generally means more performance for the same power, though for language models it matters far less than the memory subsystem.

It was released in January 2024, roughly 2 years ago. Inference software support tends to follow hardware by a year or two, so a card of this age generally has mature, well-optimised code paths available to it.

Compute throughput, and why it matters less than it looks

FP16

1,191.2 TFLOPS

FP64

37.2 TFLOPS

Tensor cores

592

On paper the B200 reaches 1,191.2 TFLOPS at half precision and 74.5 TFLOPS at single precision. These are peak figures no real workload sustains, and generating text reaches only a small fraction of them — decoding is limited by memory rather than arithmetic, which is why a card can look enormously powerful here and still produce tokens at an ordinary rate.

Double-precision throughput is 37.2 TFLOPS. It has no bearing on running a language model — no inference runtime uses it — but it separates datacentre parts from consumer ones, since the latter deliberately restrict it.

The card carries 592 tensor cores across 148 streaming multiprocessors. These accelerate the matrix arithmetic at the heart of a transformer, and they are what make prompt processing — reading a long document before answering — dramatically faster than it would otherwise be.

Clocks run from 700 MHz at base to 1.97 GHz boosted. Worth far less here than on a gaming benchmark: raising the clock speeds up the arithmetic, and the arithmetic is not what generation is waiting on.

Cache and processing units

The B200 has 250 KB of L1 cache, backed by 50 MB of L2. Cache absorbs a share of the memory traffic that would otherwise hit the main bus, which is the one place on this page where a number other than bandwidth quietly affects generation speed — a large L2 lets more of the working set stay close to the cores.

There are 18,944 shading units, 592 texture mapping units, and 24 render output units. These drive graphics workloads and contribute to prompt processing, but they sit idle for much of the time a model spends generating a reply.

Power, size and installation

Power draw

1,000 W

The B200 is rated at 1,000 W, with a 1,400 W power supply suggested for the whole system. Running a language model keeps a card busy in bursts rather than continuously — it draws hard while generating and idles between requests — so sustained draw over a working day is usually well below the rated figure.

The board occupies a sxm module. Worth checking against the case and power supply already in the machine, since the largest cards need considerably more of both than a typical desktop provides.

It connects over PCIe 5.0 x16. The interface governs how quickly a model is loaded from disk into the card, not how fast it runs once there, so a narrower link costs a few seconds at startup and nothing thereafter.

The extremes

The largest AI models a B200 can run

The biggest open-weight models that fit on this card, newest first. Each is shown at the best compression the card can hold.

  1. 01 Solar Open2 250B 250.3B · Q4_K_M · Jun 2026 174 tok/s
  2. 02 DeepSeek-V4-Flash 284B · Q4_K_M · Apr 2026 153 tok/s
  3. 03 Tencent Hy3 preview 295B · Q4_K_M · Apr 2026 147 tok/s
  4. 04 GLM-4.7 358B · Q3_K_M · Dec 2025 142 tok/s
  5. 05 MiMo-V2-Flash 309B · Q3_K_M · Dec 2025 29.6 tok/s
  6. 06 ERNIE-4.5-300B-A47B 300B · IQ4_XS · Jun 2025 154 tok/s
  7. 07 Llama Nemotron Ultra 253B 253B · Q4_K_M · Mar 2025 30.9 tok/s
  8. 08 Ling-Plus ("Bailing") 290B · IQ4_XS · Mar 2025 28.7 tok/s
  9. 09 DeepSeek-V2.5 236B · Q5_K_M · Sep 2024 142 tok/s
  10. 10 Grok-1 314B · Q3_K_M · Nov 2023 29.1 tok/s

The fastest AI models on a B200

Where this card produces tokens quickest. Smaller models dominate here, because generating each token means reading the whole model out of memory once.

  1. 01 Gemma 3 QAT 1B 1B · Q8_0 · 1.8 GB 3,388 tok/s
  2. 02 Gemma 3 1B 1B · Q8_0 · 1.8 GB 3,388 tok/s
  3. 03 LLama 3..2 Typhoon 2 1B 1B · Q8_0 · 1.8 GB 3,388 tok/s
  4. 04 OLMo-1B 1B · Q8_0 · 1.8 GB 3,388 tok/s
  5. 05 HGRN 1B (WT 103) 1B · Q8_0 · 1.8 GB 3,388 tok/s
  6. 06 Pythia-1b 1B · Q8_0 · 1.8 GB 3,388 tok/s
  7. 07 OpenELM-1.1B 1.1B · Q8_0 · 1.9 GB 3,137 tok/s
  8. 08 TinyLlama-1.1B (1T token checkpoint) 1.1B · Q8_0 · 1.9 GB 3,080 tok/s
  9. 09 TinyLlama-1.1B (3T token checkpoint) 1.1B · Q8_0 · 1.9 GB 3,080 tok/s
  10. 10 DeciCoder-1B 1.1B · Q8_0 · 1.9 GB 3,080 tok/s

Step by step

How to work out the tokens per second of a B200

You do not have to calculate anything by hand — the gputps.com calculator on this page has already worked it out for every model this card can hold. Reading off the answer takes six steps.

  1. 01

    Start with the model, not the specification

    Every one of the 633 models this B200 runs is in the table above. Search narrows it by name or by size.

  2. 02

    Match the context to your work

    Longer conversations cost memory on top of the weights. With 180 GB to work in, that is frequently the difference between a model fitting and not.

  3. 03

    Pin the comparison to one quality level

    By default the table picks the least-compressed copy that fits. Setting a floor removes models that only qualify through heavy compression.

  4. 04

    Take the range as the answer

    Speeds come with error bars for a reason. The best case here is 3,388 tok/s on Gemma 3 QAT 1B, and which inference software you use moves that by thirty to fifty per cent.

  5. 05

    Check the headroom before you decide

    A tight fit runs but leaves no room to raise the context later; comfortable has headroom. The memory column shows what each model needs against the 180 GB available.

  6. 06

    Check the same model from the other side

    Every model name in the table links to its own page, which runs the same calculation across every card we hold. That is where you see whether the B200 is the right buy for it or merely a card that fits.

Answers

B200 — common questions

01

What are the TFLOPS of a B200?

The B200 is rated at 1,191.2 TFLOPS at half precision and 74.5 TFLOPS at single precision. These are peak arithmetic ceilings rather than achievable rates, and text generation reaches only a small fraction of them because it is limited by memory bandwidth instead.

02

How many tensor cores does a B200 have?

The B200 has 592 tensor cores across 148 streaming multiprocessors. They accelerate the matrix arithmetic a transformer is built from, which mainly speeds up processing a long prompt rather than producing the reply.

03

Does the B200 support CUDA?

Yes. The B200 reports CUDA compute capability 10.0. Capability 7.0 and above has tensor cores, which modern inference software uses; below that it falls back to slower code paths for quantised models.

04

What bus interface does the B200 use?

It uses PCIe 5.0 x16. This governs how fast a model is loaded onto the card rather than how fast it runs once loaded, so it costs a few seconds at startup and nothing during generation.

05

Is the B200 good for running local AI models?

Its memory is large enough for models most desktop hardware cannot touch and its bandwidth puts it among the fastest hardware available for generation. In total it runs 633 of the models we track. Whether that is enough depends entirely on which model you want — the table above answers that directly.

06

Can a B200 run a model that does not fit in its memory?

Only partly. Layers beyond the 180 GB sit in system memory and run at a fraction of the speed, so a mostly-offloaded model is rarely worth using. Every figure here assumes it is fully resident on the card.

07

Would two B200 cards be twice as fast?

Pairing B200 cards buys headroom rather than pace: 360 GB of combined memory, at roughly the same generation speed as one.

08

What AI models can a B200 run?

633 of the 679 open-weight language models we track fit on a B200 and can be run locally on it. The table on this page lists every one, with the memory it needs, the quantisation it runs at and an estimated generation speed.

09

What is the largest AI model a B200 can run?

The largest model in our catalogue that fits on a B200 is GLM-4.7 at 358B parameters, compressed to Q3_K_M. It generates roughly 142 tokens per second and needs about 155.2 GB of the card's memory.

10

How many tokens per second does a B200 produce?

It depends on the model. On a B200 the fastest model we track is Gemma 3 QAT 1B at about 3,388 tokens per second, while larger models run proportionally slower because each token requires reading the whole model out of memory once. Speeds are estimates for a single conversation at a time.

11

Can a B200 run a 7B model?

Yes. For example a B200 runs Multi-Token Prediction 7B at Q8_0, using about 7.9 GB of memory and generating around 506 tokens per second.

12

Can a B200 run a 13B model?

Yes. For example a B200 runs DeepSeekMoE-16B at Q8_0, using about 17.5 GB of memory and generating around 1,176 tokens per second.

13

Can a B200 run a 30B model?

Yes. For example a B200 runs ERNIE-4.5-VL-28B-A3B at Q8_0, using about 29.2 GB of memory and generating around 672 tokens per second.

14

Can a B200 run a 70B model?

Yes. For example a B200 runs Qwen3-Coder-Next at Q8_0, using about 80.7 GB of memory and generating around 235 tokens per second.

15

How much memory does a B200 have?

A B200 has 180 GB of HBM3e memory. Around a tenth of that is reserved by the inference runtime and the driver, leaving roughly 162 GB available for a model and its conversation.

16

What is the memory bandwidth of a B200?

The B200 has 8,000 GB/s of memory bandwidth, across a 8,192-bit memory bus. This is the single best predictor of how fast it generates text, because producing each token means reading the entire model out of memory once.

17

What type of memory does a B200 use?

It uses HBM3e clocked at 1.95 GHz. HBM types are found on datacentre accelerators and carry far more bandwidth than the GDDR used on desktop cards, which is why they generate tokens considerably faster at the same capacity.

18

Who makes the B200?

The B200 is a NVIDIA product, with the chip manufactured by TSMC, on a 5 nm process.

19

When was the B200 released?

The B200 was released in January 2024.

20

How much power does a B200 use?

The B200 has a rated board power of 1,000 W, and a 1,400 W system power supply is suggested. Generating text draws hard in bursts and idles between requests, so average consumption over a working session is normally well below the rated figure.

21

How much cache does a B200 have?

The B200 has 250 KB of L1 cache, and 50 MB of L2 cache. Cache absorbs part of the memory traffic that would otherwise reach the main bus, so a larger L2 gives a modest lift to generation speed beyond what bandwidth alone predicts.

The other direction

Looking at it from the other side?

This page starts from the hardware. If you already know which model you want and need to know what it takes to run it, start from the model instead.

All GPUs