One board per job. A single "best at AI" score hides the thing that matters: the card that wins at running a language model is often not the card that wins at upscaling, and the gap between them can be an order of magnitude.
Every model we have measured, grouped by the VRAM it actually needs. The figure is the lowest requirement any card recorded for that model, so it is the honest floor rather than a worst case.
Every language model we have measured, grouped by the family it belongs to. Useful when you have picked a family and want to know which size to run.
Running an LLM locally: tokens per second on the models people actually run.
1392 measured results across 121 models · fastest so far: NVIDIA GeForce RTX 5090
Turning a prompt into a picture, timed end to end at 1024px.
135 measured results across 11 models · fastest so far: NVIDIA B200
Rendering a clip from a prompt, measured in frames per second of output.
49 measured results across 2 models · fastest so far: NVIDIA B300
Training throughput in tokens per second: LoRA on a fixed batch, the number that decides whether a fine-tune is an evening or a week.
43 measured results across 4 models · fastest so far: NVIDIA B300
Throughput with 32 requests in flight at once, which is the question when a model sits behind an API rather than on your desk.
39 measured results across 4 models · fastest so far: NVIDIA B200
Changing an existing image rather than making a new one.
26 measured results across 2 models · fastest so far: NVIDIA B300
Reading 3D structure out of a flat photo, a stage in most 3D and relighting pipelines.
22 measured results across 2 models · fastest so far: NVIDIA RTX PRO 6000 Blackwell Workstation Edition
Isolating objects in an image, the groundwork for masking and compositing.
22 measured results across 2 models · fastest so far: NVIDIA B300
Describing and reasoning about a picture: captioning, OCR and visual question answering.
12 measured results across 2 models · fastest so far: NVIDIA H100 80GB HBM3
Cutting the subject out of a photo, the most common production task there is.
11 measured results across 1 model · fastest so far: NVIDIA H200
Enlarging an image without it falling apart.
11 measured results across 1 model · fastest so far: NVIDIA B300
Turning a single picture into a 3D asset.
9 measured results across 2 models · fastest so far: NVIDIA H100 80GB HBM3
Animating a still you already have, the stage that turns keyframes into footage.
9 measured results across 3 models · fastest so far: NVIDIA H100 80GB HBM3
Generating spoken audio, measured against realtime: 10x means ten seconds of speech every second.
8 measured results across 1 model · fastest so far: NVIDIA L40S
Generating music from a prompt, measured against realtime.
8 measured results across 1 model · fastest so far: NVIDIA H100 80GB HBM3
Transcribing audio, measured against realtime, which is what decides whether a long recording is a coffee break or an afternoon.
8 measured results across 1 model · fastest so far: NVIDIA L40S
Pairing a small draft model with a large one is sold as a free speedup. Measured per card, it is often a slowdown. Above 1.0x it helped, below 1.0x it cost you.
6 measured results across 1 model · fastest so far: NVIDIA L4