What is the biggest model the L40S can run?

Llama 3.3 70B, which needs about 45 GB, on the 48 GB version at about 11 tok/s.

What is the biggest model the Radeon Pro W7900 can run?

Llama 3.3 70B, which needs about 45 GB, on the 48 GB version at about 11 tok/s.

L40S vs Radeon Pro W7900 for local AI

Q: Which is better for local AI, the L40S or the Radeon Pro W7900?

They're close for local AI, so pick on price. Both fit the same models we track, up to Llama 3.3 70B. They generate at about the same speed, since their memory bandwidth is close (864 and 864 GB/s).

Both fit the same models we track, up to Llama 3.3 70B. They generate at about the same speed, since their memory bandwidth is close (864 and 864 GB/s).

The numbers that matter

Memory and bandwidth

Memory decides which models fit. Bandwidth decides how fast they generate, because every token reads the model's active weights once.

	L40S	Radeon Pro W7900
Memory	48 GB VRAM	48 GB VRAM
Bandwidth	864 GB/s	864 GB/s
Biggest model that fits	Llama 3.3 70B	Llama 3.3 70B
Best everyday pick	Qwen3 32B, ~24 tok/s	Qwen3 32B, ~24 tok/s

Model by model

Every model, side by side

Estimated generation speed at the largest memory size each was sold in (48 GB and 48 GB).

Model	L40S	Radeon Pro W7900
Qwen3 4B About 4 GB	~190 tok/s Runs great	~190 tok/s Runs great
Qwen3 8B About 7 GB	~91 tok/s Runs great	~91 tok/s Runs great
Gemma 3 12B About 10 GB	~59 tok/s Runs great	~59 tok/s Runs great
Qwen3 14B About 11 GB	~51 tok/s Runs great	~51 tok/s Runs great
gpt-oss 20B About 16 GB	~150 tok/s Runs great	~150 tok/s Runs great
Gemma 3 27B About 19 GB	~28 tok/s Usable	~28 tok/s Usable
Qwen3 30B-A3B About 21 GB	~150 tok/s Runs great	~150 tok/s Runs great
Qwen3 32B About 22 GB	~24 tok/s Usable	~24 tok/s Usable
Llama 3.3 70B About 45 GB	~11 tok/s Usable	~11 tok/s Usable
gpt-oss 120B About 67 GB	Needs 67 GB	Needs 67 GB
Qwen3 235B-A22B About 144 GB	Needs 144 GB	Needs 144 GB
DeepSeek R1 671B About 406 GB	Needs 406 GB	Needs 406 GB

Which should you get?

They're close for local AI, so pick on price. Price matters as much as either number: check used prices on the builds page, where every parts list has a measured or estimated speed.

How we estimate

Both columns use the same formula: memory bandwidth divided by the weights read per token, calibrated on our own rack, with Ollama's default 4-bit sizes. Real results are usually within about a third. Measured either one? Send us the number.

Compare

Other matchups

Every comparison