Benchmarking vLLM on 2× RTX 3090 — Post-NVLink: The Verdict Flips

The NVLink Bridge

Verifying NVLink in vLLM

The Post-NVLink Sweep

Axis 1: TTFT — Halved Across the Board

Model	PCIe TTFT	NVLink TTFT	Δ	Budget (2500 ms)
qwen3-coder-30b	1112 ms	415 ms	−63%	✅ → ✅
qwen36-35b-a3b	997 ms	412 ms	−59%	✅ → ✅
qwen3-coder-next-60b-ream	1203 ms	529 ms	−56%	✅ → ✅
glm47-flash (†)	~1181 ms	511 ms	−57%	✅ → ✅
devstral-24b-gptq	3235 ms	2013 ms	−38%	❌ → ✅
qwen36-27b	4222 ms	2199 ms	−48%	❌ → ✅

Axis 2: Decode — Uniformly Small Gains

Model	PCIe tok/s	NVLink tok/s	Δ
qwen3-coder-30b	150.6	167.7	+11.3%
qwen36-35b-a3b	136.3	148.2	+8.7%
qwen3-coder-next-60b-ream	118.6	128.0	+7.9%
glm47-flash (†)	~111.0	117.5	+5.9%
qwen36-27b	59.8	62.7	+4.8%
devstral-24b-gptq	54.9	54.2	−1.3% (flat)

Axis 3: Maximum Usable Context — The Same Ceilings

Model	PCIe max ctx	NVLink max ctx	Δ
qwen3-coder-30b	256K	256K	=
qwen3-coder-next-60b-ream	256K	256K	=
qwen36-35b-a3b	192K	192K	=
qwen36-27b	192K	192K	=
glm47-flash (†)	—	192K	new (ran at TP=2 first time)
devstral-24b-gptq	128K	32K ‡	run artifact (timeout stopped probe early)

The Verdict Flips

Post-NVLink: What You’d Actually Deploy

Tier 1 — Daily driver (unchanged):

• qwen3-coder-30b (TP=2, 168 tok/s, 256K ctx, non-thinking mode)
• qwen36-35b-a3b (TP=2, 148 tok/s, 192K ctx, multimodal)

Tier 2 — Diversity / fallback (one new member):

• glm47-flash (TP=2 now possible: 118 tok/s, 192K ctx — first run at TP=2)
• glm47-flash (single GPU, 107 tok/s, 89K ctx) — still co-locatable
• qwen3-coder-next-60b-ream (TP=2, 128 tok/s, 256K ctx)
• gpt-oss-20b-nightly (TP=2, 65 tok/s, 128K ctx)

Tier 3 — Previously dead, now alive:

• devstral-24b (TP=2, TTFT now 2013ms — under budget!) — still general assistant, not coding specialist
• qwen36-27b (TP=2, TTFT now 2199ms — under budget!) — still general-purpose, but now runnable

Tier 4 — Still dead:

• qwen3-coder-next-80b (OOM on 2×24 GB — raw capacity, not interconnect)

Why the Three Axes Behave Differently

What’s Still Open

TL;DR