Ada's fastest 12 GB card: 672 GB/s with a 12 GB lid. The Super refresh later matched its speed with 16 GB — this one kept the lid.
672 GB/s is 33% ahead of the plain 4070's bus — token generation lands at ~75–100 t/s on 8B and ~45–58 t/s on 13B Q4, the best 12 GB numbers Ada produces, with prompt processing at ~850–1150 t/s. Then the 12 GB lid: 13B Q5 loads at 8K; 27B Q4 is 7 GB over; MoE 30B at full quant is out. The fastest 12 GB card in the series, and 12 GB is still a 13B tier.
Eight months later the 4070 Ti Super shipped: same 672 GB/s, 16 GB. Every LLM argument the 4070 Ti made — speed at mid price — the Super made with 4 more GB. If the two are close in price locally, the Super is the buy; the original Ti survives on discount.
285 W (NVIDIA's revised TGP; early launch materials said 260) puts it a step above the Super's envelope. A 700 W PSU requirement, 3-slot-class cooling — this is the 4070 tier's power-hungry member.
The MTP question: if your model ships a built-in multi-token-prediction head
(Qwen 3.5/3.6/3.8, Gemma 4), the one-flag speculative pass
(--spec-type draft-mtp --spec-draft-n-max 2 --parallel 1) is worth
+45–65% (estimated) on token generation here; no 4070 Ti A/B in the record; the estimate sits at the 672 GB/s class, between the measured 3070-class and 3090-class results. The price is ~0.6–2 GB of VRAM, the
output is bit-identical to the unaccelerated run, and models without MTP heads (Llama class) see
nothing. Full tuning rules in the MTP guide.
| Model | Size on GPU | Token gen | Token gen (MTP est.) | Prompt proc. | Fits? |
|---|---|---|---|---|---|
| 8B Q8_0 | ~9.7 GB | ~75–100 t/s | ~109–165 t/s | ~800–1100 t/s | ✅ 32K+ ctx |
| 13B Q4_K_M | ~8.5 GB | ~45–58 t/s | ~65–95 t/s | ~850–1150 t/s | ✅ 32K+ ctx |
| 13B Q5_K_M | ~11.9 GB | ~32–42 t/s | ~46–69 t/s | ~700–900 t/s | ⚠️ 8K ctx |
| 27B Q4_K_M | ~19 GB | — | — | — | ❌ ~7 GB 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 →
Figures assume full offload (-ngl 99), ~8K context and batch 2048. Token speed tracks the memory bus almost linearly, so treat ranges as class estimates — your llama.cpp build and the model's attention layout shift them.
| Quant | GGUF size | 1× 4070 Ti (12 GB) | 2× 4070 Ti (24 GB) |
|---|---|---|---|
| Q4_K_M | ~19 GB | ❌ ~7 GB over | ⚠️ ~20–26 t/s, 8K ctx |
| Q5_K_M | ~22 GB | ❌ | ❌ ~2 GB over |
| Q6_K | ~25 GB | ❌ | ❌ |
| Q8_0 | ~33 GB | ❌ | ❌ |
27B is a hard no single-card. A pair (24 GB) runs Q4_K_M at 8K — a clear margin over 2× 4070; 3× (36 GB) reaches Q5 at 8–16K around ~18–23 t/s. The speed story is good; the VRAM story is the 4070's, two sizes down.
PCIe 4.0, ~25–35% tax, 12 GB tiles: 2× hits the 27B Q4 line, 3× hits Q5.
| Build | GPU power | VRAM | 13B Q8 | 27B Q4_K_M | 27B Q5_K_M |
|---|---|---|---|---|---|
| 2× 4070 Ti | 570 W | 24 GB | ~52–68 t/s | ⚠️ ~20–26 t/s, 8K ctx | ❌ ~2 GB over |
| 3× 4070 Ti | 855 W | 36 GB | (overkill) | ~24–30 t/s, 16K+ ctx | ~18–23 t/s, 8–16K ctx |
The 3-way build (855 W, 36 GB) is a serious 27B Q5 box — and a serious power draw. It needs three working x8/x16 lanes, which is where consumer boards stop.
When a model won't fit, -ngl N parks the tail layers on the CPU. The mental model
that saves pain: offloaded layers don't run at a fraction of GPU speed — they run at CPU
speed. Handy moves: peel the last 2–4 layers off to reclaim ~0.5–1 GB of VRAM for almost
nothing; keep MoE experts on CPU with --n-cpu-moe; park the output embedding with
-ot output=CPU on big models. Budget 32 GB system RAM for 27B-class offload, 64 GB for
70B. And never offload the KV cache — the session becomes unusable.
llama-server -m 27b-q5_k_m.gguf -ngl 99 -sm layer -c 16384 --tensor-split 1,1,1
llama-server -m llama-3.1-13b-q5_k_m.gguf -ngl 99 -c 8192 -b 2048 -ub 2048 --host 127.0.0.1 --port 8080
The fastest 12 GB Ada, and a footnote to its own refresh: the 4070 Ti Super matches its speed with 4 GB more VRAM. At a real discount the Ti is a fine 13B-Q5 card and a fast 27B half-rig; at close prices, the Super wins by default.