V100 16GB PCIe vs V100 32GB SXM2 for Local LLMs

Both list the same memory bandwidth (900 GB/s), so per-token decode speed should be similar whenever a model fits both. The V100 32GB SXM2 fits more of the catalog at Q4/8K — 43 vs 25 of 73 models (32 vs 16 GB VRAM). Largest model that fits the V100 32GB SXM2 but not the V100 16GB PCIe: Qwen3.6-35B-A3B-FP8 (36.0B). MSRP is $8,000 for the V100 16GB PCIe vs $10,000 for the V100 32GB SXM2. MSRP is a launch list price, not live retail — Amazon prices move; click through for the current price. Not retail-shoppable: V100 16GB PCIe, V100 32GB SXM2 (datacenter-class; cloud rental is the realistic way to use it). In the Q4/8K head-to-head below (49 shared models), the V100 32GB SXM2 posts the higher tok/s in 15 rows vs 9. Methodology.

Spec comparison

	V100 16GB PCIe	V100 32GB SXM2
Vendor	NVIDIA	NVIDIA
VRAM	16 GB	32 GB
Memory type	HBM2	HBM2
Memory bandwidth	900 GB/s	900 GB/s
TDP	250 W	300 W
MSRP (list)	$8,000	$10,000
Buy	—	—

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 49 shared models

Model	Params (B)	V100 16GB PCIe tok/s	V100 16GB PCIe fits	V100 32GB SXM2 tok/s	V100 32GB SXM2 fits	Winner
MiniCPM5-1B	1.08	895.6 est.	Yes	895.6 est.	Yes	Tie
Qwen3.5-2B	2.0	483.6 est.	Yes	483.6 est.	Yes	Tie
Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16	30.0	434.0 est.	No	322.4 est.	Yes	V100 16GB PCIe
DeepSeek-R1-0528-Qwen3-8B-layer-mix-bpw-3.8-mlx	2.3	420.5 est.	Yes	420.5 est.	Yes	Tie
Qwen3-30B-A3B-Instruct-2507	30.5	405.4 est.	No	322.4 est.	Yes	V100 16GB PCIe
Qwen3.5-35B-A3B	34.7	339.8 est.	No	322.4 est.	Yes	V100 16GB PCIe
Ornith-1.0-35B	35.0	334.2 est.	No	322.4 est.	Yes	V100 16GB PCIe
qwen35b-a3b-fable-sft-abliterated	34.7	325.2 est.	No	322.4 est.	Yes	V100 16GB PCIe
Qwen3.6-35B-A3B	36.0	316.9 est.	No	322.4 est.	Yes	V100 32GB SXM2
GLM-4.7-Flash	31.2	313.0 est.	No	322.4 est.	Yes	V100 32GB SXM2
Llama-3.2-3B-Instruct	3.2	302.3 est.	Yes	302.3 est.	Yes	Tie
gpt-oss-20b	21.5	268.7 est.	Yes	268.7 est.	Yes	Tie
Qwen3-Coder-Next	79.7	67.0 est.	No	268.3 est.	No	V100 32GB SXM2
Qwen3-Next-80B-A3B-Instruct	81.3	64.5 est.	No	258.2 est.	No	V100 32GB SXM2
gemma-4-26B-A4B-it	25.2	254.5 est.	Yes	254.5 est.	Yes	Tie

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

	V100 16GB PCIe	V100 32GB SXM2
Catalog models that fit	25 of 73	43 of 73
Largest exclusive fit	—	Qwen3.6-35B-A3B-FP8 (36.0B)

V100 16GB PCIe details → · V100 32GB SXM2 details →