Ada's sweet spot: 672 GB/s and 16 GB — 13B at Q6, IQ8 and MoE 30B on the fastest 16 GB bus of the generation, at 285 W.
The Ti Super is the only Ada card that pairs a 16 GB pool with a fast GDDR6X bus. What that combination buys, at 672 GB/s:
Token speed tracks the 4070 Ti's bus; the VRAM is the upgrade. Prompt processing (~900–1200 t/s on 13B) is the best in the 16 GB tier.
27B Q4 needs 19 GB — 3 GB over. 13B plain Q8_0 (17.4 GB) misses by 1.4 GB. Those are the two lines this card sits just below, and a pair crosses both.
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 per-card A/B in the record; the estimate matches its 672 GB/s bus class at n-max 2. 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 | ~900–1200 t/s | ✅ 32K+ ctx |
| 13B Q4_K_M | ~8.5 GB | ~45–58 t/s | ~65–95 t/s | ~900–1200 t/s | ✅ 32K+ ctx |
| 13B Q6_K | ~13.8 GB | ~32–42 t/s | ~46–69 t/s | ~750–950 t/s | ✅ 8–16K ctx |
| 27B Q3_K_M | ~15.8 GB | ~11–15 t/s | ~16–25 t/s | ~500–650 t/s | ⚠️ 4–8K ctx |
| MoE 30B-A3B (Q3) | ~14.5 GB | ~32–42 t/s | ~46–70 t/s | ~800–1000 t/s | ⚠️ 4–8K ctx, 3B active |
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. The Q6_K and IQ8_0 lines are what the extra 4 GB over the 4070 Ti actually buys.
| Quant | GGUF size | 1× 4070 Ti S (16 GB) | 2× 4070 Ti S (32 GB) |
|---|---|---|---|
| Q3_K_M | ~15.8 GB | ⚠️ ~11–15 t/s, 4–8K ctx | ✅ (overkill) |
| Q4_K_M | ~19 GB | ❌ ~3 GB over | ✅ ~22–28 t/s, 8–16K ctx |
| Q5_K_M | ~22 GB | ❌ | ✅ ~19–24 t/s, 8K ctx |
| Q6_K | ~25 GB | ❌ | ✅ ~16–21 t/s, 8K ctx |
| Q8_0 | ~33 GB | ❌ | ❌ ~1 GB over |
Single-card 27B is Q3 at short context — the honest ceiling. The pair (570 W, 32 GB) is Ada's best single-rig 27B story: Q4 at 8–16K, Q5/Q6 at 8K, and 13B at full Q8_0 (17.4 GB) with room to spare. It sits ~40% behind a 2× 3090 NVLink rig at the same models — and ~3× behind a 2× 5070 Ti — but it's the fastest 16 GB pair of its generation at street prices.
PCIe 4.0, ~25–35% tax. 16 GB tiles: the pair lands on the 27B Q4 line with 32 GB.
| Build | GPU power | VRAM | 13B Q8_0 | 27B Q4_K_M | 27B Q5_K_M |
|---|---|---|---|---|---|
| 2× 4070 Ti S | 570 W | 32 GB | ~55–70 t/s, 16K+ ctx | ✅ ~22–28 t/s, 8–16K ctx | ~19–24 t/s, 8K ctx |
Two slots on a normal ATX board close the job. At 570 W total it's a serious rig — but it fits in a regular case with a 2×8-pin-800W PSU.
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-q4_k_m.gguf -ngl 99 -sm layer -c 16384 --tensor-split 1,1
llama-server -m llama-3.1-13b-q6_k.gguf -ngl 99 -c 16384 -b 2048 -ub 2048 --host 127.0.0.1 --port 8080
The 4070 Ti Super is the card the 40 series should have started with: 16 GB and 672 GB/s in one envelope, at 285 W. It's the generation's single-card answer for 13B quality work and MoE models, and — in pairs — its 27B answer. It loses to the 3090 on speed-per-token and to the 50-series on everything, but at street prices it's the best Ada LLM card made.