Llama 3.3 70B · 39 cards measured first-party · Updated July 2026

How Fast Does Llama 3.3 70B Run on Each GPU?

Llama 3.3 70B is the model that separates a serious AI machine from an expensive one. At Q4_K_M it needs roughly 42GB of VRAM, and that single number disqualifies more of the GPU market than any other figure in our suite, including every consumer card ever made. We ran it on every GPU that could hold it and published a hard gate on every GPU that couldn't.

Benchmarked weights: bartowski/Llama-3.3-70B-Instruct-GGUF

Fastest we measured

NVIDIA H100 NVL

48.2 tok/s on Llama 3.3 70B. Anchored estimate. 94GB of VRAM, 400W board rating. AI Score 67.0/100 across our full 12-workload suite.

Pros

• 48.2 tok/s on Llama 3.3 70B
• 94GB, clears the Llama 3.3 70B floor
• Rentable by the hour rather than bought

Cons

• 400W board rating
• Datacenter or workstation hardware, not a retail purchase

Best for: Llama 3.3 70B work where you want the ceiling gone rather than the cheapest entry.

Runner-up

NVIDIA B300

47.97 tok/s on Llama 3.3 70B. Measured on our bench. 288GB of VRAM, 1400W board rating. AI Score 93.8/100 across our full 12-workload suite.

Pros

• 47.97 tok/s on Llama 3.3 70B
• 288GB, clears the Llama 3.3 70B floor
• Rentable by the hour rather than bought

Cons

• 1400W board rating
• Datacenter or workstation hardware, not a retail purchase

Best for: Llama 3.3 70B work where you want the ceiling gone rather than the cheapest entry.

Third

NVIDIA B200

44.54 tok/s on Llama 3.3 70B. Measured on our bench. 192GB of VRAM, 1000W board rating. AI Score 78.0/100 across our full 12-workload suite.

Pros

• 44.54 tok/s on Llama 3.3 70B
• 192GB, clears the Llama 3.3 70B floor
• Rentable by the hour rather than bought

Cons

• 1000W board rating
• Datacenter or workstation hardware, not a retail purchase

Best for: Llama 3.3 70B work where you want the ceiling gone rather than the cheapest entry.

Token generation on a 70B is bound by memory bandwidth, not compute. The model has to be read out of VRAM once per token, so the ceiling is how fast the card can move 42GB of weights, not how many tensor cores it has. That's why the ranking below tracks bandwidth almost perfectly and ignores core counts, and why an HBM card with modest compute buries a consumer flagship that can't even load the thing.

Because this workload is bandwidth-bound, the ranking above tracks memory bandwidth far more closely than core counts or price. A card with fewer tensor cores and faster memory will beat a card with the opposite.

That's the reason we run twelve workloads instead of publishing one score. A GPU isn't fast or slow. It's fast at some things and gated out of others, and which of those matters depends entirely on what you're actually going to run.

Our verdict

NVIDIA H100 NVL tops our Llama 3.3 70B leaderboard at 48.2 tok/s (anchored estimate), 357% of the way clear of the slowest card that still fits. But the number that decides most purchases isn't on the chart. It's the 41 cards that can't run Llama 3.3 70B at all. This is a bandwidth workload: buy memory speed, not tensor cores.

FAQ

What is the fastest GPU for Llama 3.3 70B?

NVIDIA H100 NVL, at 48.2 tok/s on our bench, an anchored estimate against our measured cards. It carries 94GB of VRAM. Of the 61 cards we have Llama 3.3 70B data for, 20 can run it at all.

How much VRAM do I need for Llama 3.3 70B?

42GB at Q4_K_M is the floor. There is no driver update, no optimisation and no setting that gets a 32GB card past it, the weights either fit or they don't. Cards below the line score zero on this workload in our AI Score, because a card that can't run the job doesn't get partial credit for being quick at the jobs it can.

Why does the Llama 3.3 70B ranking look different from your other benchmarks?

Because this workload is bandwidth-bound, the ranking above tracks memory bandwidth far more closely than core counts or price. A card with fewer tensor cores and faster memory will beat a card with the opposite. That's why we publish twelve separate workloads rather than one blended score, the ordering genuinely changes depending on the job.

Are these Llama 3.3 70B numbers measured or estimated?

Both, and every row says which. 39 of the 61 cards here were rented and run by us on the same harness. The remainder are anchored estimates interpolated per workload against those measurements. We never blend the two silently, if a row says Estimated, we have not run that card.

Can I rent a GPU to run Llama 3.3 70B instead of buying one?

Yes, and for the cards at the top of this leaderboard it's the only realistic option, most of them have no retail channel at all. It's also how we got these numbers: we rented the hardware by the hour rather than buying it. That's worth considering before you spend on a card to find out whether it's fast enough.

Why publish cards that can't run Llama 3.3 70B?

Because it's the most useful thing we know. A card that can't load a model doesn't run it slowly, it doesn't run it. Most benchmark sites leave that as a blank cell or quietly drop to a smaller quantisation to produce a number. We publish it as a hard gate and score it zero, because 'this card cannot do the thing you want' is the answer to the question you were actually asking.

How we test

Every ranking on this page comes from our own benchmark runs, not vendor claims. Cards marked Measured were rented and run by us; cards marked Estimated are interpolated per workload against those measured anchors and are labelled on every row, we never blend the two silently. LLMs run on llama.cpp (llama-bench) at Q4_K_M with -p 512 -n 128. Diffusion and video run on diffusers/ComfyUI at BF16, with SDXL at FP16. Each workload gets a warmup pass plus multiple timed runs (5 for small LLMs, 3 for large models and images, 2 for video); we publish the mean as the result and the minimum as the 1% low. Run-to-run variance is under 0.5%. Telemetry, power, temperature, utilisation, clocks, peak VRAM, is sampled at 1 Hz for the duration of every run. Where a model exceeds a card's VRAM we publish a hard won't-fit result rather than quietly dropping to a smaller quantisation. A card that can't run a model scores zero on it. Silently swapping precision to make a number appear would make every number on this site meaningless. All figures are single-GPU, single-stream, batch-size-1. That is the honest way to measure what one card does for one user, and it is deliberately not how a datacenter serves a model. Vendor and MLPerf figures use large batches across many GPUs and will be far higher. Neither is wrong, they answer different questions. Ours answers 'what will this card do for me'.