The 3080 with its one flaw corrected: the full 384-bit bus and 12 GB of GDDR6X. Near-3080 Ti speed at 3080-class money — the value pick of the pair.
Bandwidth is derived from the official spec (384-bit × 19 Gbps GDDR6X). The launch 3080 used a narrowed 320-bit bus for 10 GB; this is the variant that kept the original full-width bus — hence 912 GB/s vs 760.
The 3080 Ti's edge over this card is 48 GB/s of bandwidth (960 vs 912) and 1280 extra cores. For token generation that's roughly a 5% difference; for prompt processing a touch more. If both cards cost about the same, the Ti wins — but the 12 GB 3080 regularly sits well below it on the used market, which changes the equation.
The 10 GB 3080 was the odd one out in its class: fast, but forced to ration KV cache. This board fixes that. 13B Q4_K_M (~8.5 GB) leaves 3+ GB for context — a real 8K–16K window — and 8B models reach Q6/Q8 with room for 32K. Same ceiling as the 3060 12GB and 3080 Ti; the difference is how fast you stream.
8960 cores at ~1.71 GHz lands prompt processing around ~1000–1400 t/s on 13B. Fine for single-user work and modest llama-server fan-out; the 3080 Ti and 3090 pull ahead when you push large batch sizes.
Before the numbers: if your model ships a built-in multi-token-prediction head
(e.g. Qwen 3.5/3.6/3.8, Gemma 4), generation can be made faster with
--spec-type draft-mtp --spec-draft-n-max 2 --parallel 1 — on this card that's worth
+45–60% (estimated) on token generation, scaled from the measured 3090 Ti number
(42.0 → 60.9 t/s at n-max 2). It costs ~0.6–2 GB extra VRAM, does not change output
quality, and only applies to MTP-capable models — Llama-class models are unaffected. Details and
the tuning rules: MTP guide →
| Model | Size on GPU | Token gen | Token gen (MTP est.) | Prompt proc. | Fits? |
|---|---|---|---|---|---|
| 8B (Llama 3.1 8B, Mistral) | ~4.9 GB | ~94–123 t/s | ~136–184 t/s | ~1000–1400 t/s | ✅ Q8_0 with big ctx |
| 13B Q4_K_M | ~8.5 GB | ~55–69 t/s | ~80–114 t/s | ~750–1000 t/s | ✅ 8K–16K ctx |
| 13B Q6_K | ~10.6 GB | ~44–58 t/s | ~65–93 t/s | ~700 t/s | ✅ 4K–8K ctx |
| 20B+ dense | ~13 GB | — | — | — | ❌ just over |
MTP est. = rough prior with --spec-type draft-mtp for models that ship MTP heads (Qwen 3.5/3.6/3.8, Gemma 4); Llama-class models get no MTP speedup. Actual gains depend on model, quant, context length, llama.cpp build and your card — sweep --spec-draft-n-max, don't trust the column blindly. MTP guide →
Estimates, full GPU offload, ~8K context. Scaled from the 3080 Ti at 912/960 of the bandwidth.
| Quant | GGUF size | 1× 3080 12G (12 GB) | 2× 3080 12G (24 GB) |
|---|---|---|---|
| Q4_K_M | ~19 GB | ❌ ~7 GB over | ✅ ~24–27 t/s, 8–16K ctx |
| Q5_K_M | ~22 GB | ❌ | ⚠️ ~19–21 t/s, 4–8K ctx |
| Q6_K | ~25 GB | ❌ | ❌ ~1 GB over |
| Q8_0 | ~33 GB | ❌ | ❌ |
Same single-card wall as every other 12 GB board: 27B simply doesn't fit. In pairs the picture is 95% of a 2× 3080 Ti build — ~25 t/s on 27B Q4_K_M — and the price difference is the whole story. Beyond 24 GB you're in 3090 territory: a 3090 + 3080 12G pair gives 36 GB, enough for Q6 (see 3090).
-c 16384 is the sweet spot; ~40 t/s with a real context window.Single-card numbers above. 912 GB/s per tile — a close second to the 3080 Ti among PCIe 30-series tiles. No NVLink, so the ~30% split tax applies.
| Build | GPU power | VRAM | 13B Q4 | 27B Q4_K_M | 27B Q5_K_M | 27B Q6_K | 27B Q8_0 |
|---|---|---|---|---|---|---|---|
| 2× 3080 12G | 700 W | 24 GB | ~61 t/s | ✅ ~24–27 t/s | ⚠️ ~19–21 t/s, 4–8K | ❌ ~1 GB over | ❌ |
| 3× 3080 12G | 1050 W | 36 GB | ~74 t/s | ~31 t/s | ~26 t/s | ~24–26 t/s | ⚠️ ~17–19 t/s, 8K |
| 4× 3080 12G | 1400 W | 48 GB | ~95 t/s | ~36 t/s | ~31 t/s | ~28–30 t/s | ~21–23 t/s |
| 4× 3080 12G · 70B | 1400 W | 48 GB | — | 70B Q4_K_M at ~19–21 t/s — workable, but 2× 3090 with NVLink beats it for 200 W less | |||
Mirror-image of the 3080 Ti builds at ~95% of the speed: 3× (36 GB) is the 27B-Q6 point, 4× (48 GB) a full 70B-Q4 / 27B-Q8 rig. Any time the per-card price of a 3090 gets close, the mixed and dual-3090 builds overtake these on power and interconnect.
When a model doesn't fit, -ngl N keeps the last layers on the CPU instead.
Offloaded layers run at CPU speed, not a percentage of GPU speed: a half-offloaded
27B model is ~3–5 t/s total, not 50% of GPU speed. Practical patterns: drop just the last 2–4
layers to reclaim ~0.5–1 GB of VRAM (nearly free); keep MoE experts on CPU with
--n-cpu-moe; park the output embedding with -ot output=CPU. Budget 32 GB
of system RAM for 27B-class offload, 64 GB for 70B. Never offload the KV cache. On the 24 GB
pair the offload is only interesting for the 1-GB Q6 gap — IQ6_XS (~24 GB) fits without it.
llama-server -m 27b-q6_k.gguf -ngl 99 -sm layer -c 16384 --tensor-split 1,1,1llama-server -m 70b-q4_k_m.gguf -ngl 99 -sm layer -c 8192 --tensor-split 1,1,1,1
llama-server -m llama-3-13b-q4_k_m.gguf -ngl 99 -c 16384 -b 2048 -ub 2048 --host 127.0.0.1 --port 8080
The 13B card to buy when the 3080 Ti is priced like a 3090. Everything the Ti does, at 95% of the speed and the same 12 GB ceiling — and as a tile it's a near-drop-in substitute in any 24/36/48 GB build. If the price gap closes, the Ti's extra cores and 48 GB/s take over; if your budget stretches to 3090, its 24 GB and NVLink do. But at street prices, this is the sweet spot of the 3080 family for llama.cpp.