Arc B580 12GB vs H100 80GB SXM5 for Local LLMs

The H100 80GB SXM5 has about 635% more memory bandwidth (3,350 vs 456 GB/s). Batch-1 decode tracks bandwidth closely, so a model that fits both cards should decode roughly 635% faster on the H100 80GB SXM5. The H100 80GB SXM5 fits more of the catalog at Q4/8K — 46 vs 18 of 61 models (80 vs 12 GB VRAM). Largest model that fits the H100 80GB SXM5 but not the Arc B580 12GB: gpt-oss-120b (120.4B). MSRP is $249 for the Arc B580 12GB vs $30,000 for the H100 80GB SXM5. MSRP is a launch list price, not live retail — Amazon prices move; click through for the current price. Not retail-shoppable: H100 80GB SXM5 (datacenter-class; cloud rental is the realistic way to use it). In the Q4/8K head-to-head below (40 shared models), the H100 80GB SXM5 posts the higher tok/s in 40 rows vs 0. Methodology.

Spec comparison

	Arc B580 12GB	H100 80GB SXM5
Vendor	Intel	NVIDIA
VRAM	12 GB	80 GB
Memory type	GDDR6	HBM3
Memory bandwidth	456 GB/s	3,350 GB/s
TDP	190 W	700 W
MSRP (list)	$249	$30,000
Buy	Search Amazon (affiliate)	—

MSRP is the launch list price, not live retail. Buy links are Amazon affiliate links — the price after you click is the live price.

Q4 / 8K head-to-head — top 15 of 40 shared models

Model	Params (B)	Arc B580 12GB tok/s	Arc B580 12GB fits	H100 80GB SXM5 tok/s	H100 80GB SXM5 fits	Winner
Qwen3-0.6B	0.8	362.7 est.	Yes	2664.8 est.	Yes	H100 80GB SXM5
gemma-3-1b-it	1.0	290.2 est.	Yes	2131.8 est.	Yes	H100 80GB SXM5
Llama-3.2-1B-Instruct	1.2	241.8 est.	Yes	1776.5 est.	Yes	H100 80GB SXM5
Qwen3-1.7B	2.0	145.1 est.	Yes	1065.9 est.	Yes	H100 80GB SXM5
Qwen3-30B-A3B	30.5	117.8 est.	No	710.6 est.	Yes	H100 80GB SXM5
Qwen3-Coder-30B-A3B-Instruct	30.5	117.8 est.	No	710.6 est.	Yes	H100 80GB SXM5
GLM-4.7-Flash	31.2	113.1 est.	No	710.6 est.	Yes	H100 80GB SXM5
Qwen3-30B-A3B-Instruct-2507	30.5	117.8 est.	No	710.6 est.	Yes	H100 80GB SXM5
Qwen3-Next-80B-A3B-Instruct	81.3	18.5 est.	No	710.6 est.	Yes	H100 80GB SXM5
Qwen3-Coder-Next	79.7	19.3 est.	No	710.6 est.	Yes	H100 80GB SXM5
Llama-3.2-3B-Instruct	3.2	47.4 measured	Yes	666.2 est.	Yes	H100 80GB SXM5
gpt-oss-20b	21.5	180.6 est.	No	592.2 est.	Yes	H100 80GB SXM5
Phi-4-mini-instruct	3.8	76.4 est.	Yes	561.0 est.	Yes	H100 80GB SXM5
Phi-4-mini-reasoning	3.8	76.4 est.	Yes	561.0 est.	Yes	H100 80GB SXM5
gemma-3-4b-it	4.3	74.4 est.	Yes	546.6 est.	Yes	H100 80GB SXM5

est. = bandwidth-based estimate; measured = curated real benchmark that overrides the estimate. Fits = weights + KV(8K) + 1 GB overhead ≤ 95% of VRAM.

Fit coverage at Q4 / 8K

	Arc B580 12GB	H100 80GB SXM5
Catalog models that fit	18 of 61	46 of 61
Largest exclusive fit	—	gpt-oss-120b (120.4B)

Arc B580 12GB details → · H100 80GB SXM5 details →