Two RTX 3060s: 16 tok/s Dense, 94 tok/s MoE

Two RTX 3060 12GB cards, bought new, give you 24 GB of VRAM for around $990 — 2 × $495 in the 2026-09-28 snapshot — against one used RTX 3090 at $1,300, with roughly the same power draw (2 × 170 W against 350 W). On paper they are the same build with a different number of boxes.

Our board has 29 sourced runs on an RTX 3060 x2 rig, more than any other consumer configuration we track. Read down them and the interesting number is not the VRAM. It is the 6× spread between two models of the same size that the same rig renders at completely different speeds.

As of 2026-09-15. Every number is an imported, sourced row on our board with engine, quantization and context attached; we did not run these tests. Sources at the end.

What a dual-3060 rig actually runs

The 24 GB ceiling is real and the rig sits just under it — peak measured VRAM across these runs is 20.1 GB. Here is the top of its size ladder:

Two models sit at the top of that table and one of them is six times the other. GLM-Z1-32B at 33B dense: 16.6 tok/s — technically running, painful to use. Qwen3-30B-A3B-Thinking at 31B but with 3B active parameters per token: 94.3 tok/s. Same rig, same quantization family, same month.

Why the shape matters more than the size

Decode throughput is set by how many bytes the engine must read to produce one token. For a dense model that is the whole weight set every token. For a mixture-of-experts model it is only the routed experts — roughly the active parameter count.

• Dense 31–33B at Q4: ~16–20 GB read per token. The dual-3060 rig moves 24 GB of weight traffic per token over two cards' memory bandwidth, and it lands at 16 tok/s.
• MoE 31B with 3B active at Q4: ~2 GB read per token. Ten times less traffic for a model that answers like a 30B class model, and the same rig produces 94 tok/s.

The DRAM bandwidth of a 3060 is 360 GB/s; two cards do not pool it — the second card's bandwidth only helps for the layers that live on it. But that does not matter much when the per-token traffic is ten times smaller. The dual-card rig's weakness — split memory, PCIe transfers between layers — is mostly hidden by a sparse model and fully exposed by a dense one.

That is the decision rule this build turns on, and it is not the one people apply when they compare "24 GB for $990" against "24 GB for $1,300" (2026-09-28 snapshot).

How it compares to the single 3090

No model on our board was measured on both the dual-3060 rig and a single 3090, so this is an architecture-to-architecture comparison rather than a controlled one — we are comparing a 30B MoE on the 3060 pair against a 30B MoE on the 3090:

Three things to take from that table, in order of importance:

1. The single 3090 is 3.75× the dual 3060 on the same shape of model — 353.7 against 94.3. Bandwidth accounts for about 2.6× of it (936 GB/s against 360), and the rest is format and kernel: NVFP4 has a fast path on Ada-class silicon, Q4_K_M on Ampere does not enjoy the same advantage.
2. The 3090 pair is slower than the single 3090 (197 against 353.7) — because it is running a 16× larger context on a different engine. This is the same trap as our 552B MoE analysis: the box count is not the variable.
3. The 3090's figure is at 8K context and the 3060 pair's is at 65K. If you need the window, the comparison narrows considerably — we have no 3090 row for this model at 65K.

Should you build it?

If your workload is MoE models, the dual 3060 is the cheapest way into 30B-class quality we can see on the board. 94 tok/s at 65K context for around $990 in new cards (2026-09-28 snapshot), with two-thirds of the power draw spread across two standard 8-pin connectors. Models like Qwen3-30B-A3B and gemma-4-26B-A4B are exactly what this rig is for.

If your workload is dense, do not. 16 tok/s is below the threshold where most people keep using a model. A single used 3090 costs about 31% more (one 3090 at ~$1,300 against $990 for the pair, 2026-09-28 snapshot), fits the same class of dense model, and runs it several times faster — plus it is one card, one power connector, one thing to cool.

And if you already own one 3060, the second card is a reasonable $495 upgrade (2026-09-28 snapshot) that opens up 30B-class MoE models you cannot fit today. That is the strongest case for this build: it is an upgrade path, not a shopping list.

Two footnotes worth keeping:

• QAT versions of the same model are not the same row. gemma-4-26B-A4B-it-qat shows 95.3 tok/s against 62.1 for the plain quantized build — but the plain build is measured at a 65K context and the QAT row at 512, so most of that gap is the context, not the quantization. Compare within a context column.
• The 24 GB question has a second answer we have not measured. Nobody has put a 3060 pair against a 3090-pair on the same model and context. If you have both, submit the run — it is the comparison this decision deserves.

FAQ

Are two RTX 3060s as good as one RTX 3090?

For mixture-of-experts models at the same total VRAM, no — the single 3090 ran a 30B MoE at 353.7 tok/s against the dual-3060's 94.3 on a comparable model shape. For capacity they are equivalent: both hold roughly 32B at 4-bit.

What is the fastest 30B-class model on a dual-3060 rig?

On our board, Qwen3-30B-A3B-Thinking at 94.3 tok/s (Q4_K_M, 65K context) and gemma-4-26B-A4B-it-qat at 95.3 tok/s — both mixture-of-experts models with a few billion active parameters.

Why is a 30B MoE so much faster than a 30B dense model?

Decode reads the active weights once per token. A 30B dense model reads ~16–20 GB per token at 4-bit; a 30B MoE with 3B active reads about a tenth of that, which is why the same rig shows 94 tok/s against 16 tok/s.

Does a second GPU add its memory bandwidth?

No. Each card serves the layers resident on it; the combined figure is not 2 × 360 GB/s. The split also adds inter-card transfers for every layer boundary — costs that a sparse model largely hides.

Is a used RTX 3090 better value than two new RTX 3060s?

Per dollar on dense models, yes: one 3090 costs ~31% more than two 3060s and runs dense 30B-class models several times faster. On MoE models the gap narrows to the bandwidth difference, and the 3060 pair wins on warranty and on spreading the purchase.

Sources

• Dual-3060 rows (29 sourced runs): RTX 3060 x2 · RTX 3060 · RTX 3090.
• Source runs: Qwen3-30B-A3B-Thinking on 3060 ×2 and Nemotron-3.5-Lightning-30B-A3B on RTX 3090 / 3090 ×2 are imported, sourced rows linked to their original posts from the board.
• Price context: US street prices — RTX 3060 at $495 (new, 2026-09-28 snapshot), RTX 3090 at $1,300 (used, 2026-09-28 snapshot).
• Related: 3.5 Bits Beat 4 Bits · One Spark, 552B · full board · methodology.

Imported upstream measurements as of 2026-09-14. DiggTop did not run these tests; each row carries engine, quantization and context, and states the fields the methodology rules require.