Same GA104, higher power — and 19 Gbps GDDR6X doubles the 3070's bandwidth to 912 GB/s. For LLMs, this is the one "Ti" upgrade that really counts.
The 3070 Ti keeps the 3070's 256-bit bus but moves to 19 Gbps GDDR6X, doubling bandwidth to 912 GB/s. Since token generation is bandwidth-bound, this card runs roughly 2× faster than the plain 3070 and lands in the 3080's speed class (760 GB/s), just behind the 3080 Ti (960 GB/s). The extra CUDA cores (6144 vs 5888) add a modest prompt processing gain (~900 t/s vs ~800).
Same ceiling: 8B Q4/Q5 with plenty of context, 13B only at Q3 or with aggressive KV-cache trimming. The story on this card is identical to the 3070, just with more watts (290 W TDP — a full 70 W more for the same LLM ceiling).
In a server running 24/7, 290 W vs 220 W for roughly 2× the LLM throughput is defensible on its own — the watts buy real speed here. In multi-GPU builds the extra heat still compounds: tiling two of them pulls ~580 W of GPU power for 16 GB and ~55 t/s on a 13B Q4.
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
+40–60% (estimated) on token generation; no single-card 3070 Ti A/B is in the community record yet. It costs ~0.6–2 GB extra VRAM and
does not change output quality, and it 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 | ~90–118 t/s | ~126–189 t/s | ~800–1000 t/s | ✅ comfortable, 16K+ ctx |
| 8B Q6_K | ~6.3 GB | ~42–54 t/s | ~59–86 t/s | ~800 t/s | ✅ with 8K ctx |
| 13B Q4_K_M | ~8.5 GB | — | — | — | ❌ just over budget |
| 13B Q3_K_M | ~6.7 GB | ~55–72 t/s | ~77–115 t/s | ~600 t/s | ⚠️ quality loss |
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. Token gen ≈ 2× RTX 3070 (GDDR6X); prompt proc. ≈10–15% higher.
| Quant | GGUF size | 1× 3070 Ti (8 GB) | 2× 3070 Ti (16 GB) |
|---|---|---|---|
| Q4_K_M | ~19 GB | ❌ ~11 GB over | ❌ ~3 GB over |
| Q5_K_M | ~22 GB | ❌ | ❌ |
| Q6_K | ~25 GB | ❌ | ❌ |
| Q8_0 | ~33 GB | ❌ | ❌ |
Identical outcome to the 3070 — the Ti's extra cores don't add VRAM. Partial CPU offload (≈14 layers on GPU) runs 27B Q4_K_M at ~4–6 t/s, marginally faster than the plain 3070's offload. The 290 W TDP makes this the least efficient way in the lineup to do anything that 8 GB can't hold. For 27B: the card to pair two of is the 3060 12GB, and the card to add one of is the 3090.
Single-card numbers above. Tiling Tis adds a few % of speed over plain 3070s and 70 W of heat per card — the 8 GB ceiling and the ~30% PCIe split tax (no NVLink) are unchanged.
| Build | GPU power | VRAM | 13B Q4 | 27B Q4_K_M | 27B Q5_K_M | 27B Q6/Q8 |
|---|---|---|---|---|---|---|
| 2× 3070 Ti | 580 W | 16 GB | ~42 t/s | ❌ ~3 GB over | ❌ | ❌ |
| 3× 3070 Ti | 870 W | 24 GB | ~40–45 t/s | ~16–18 t/s, 8K ctx | ~14 t/s | ❌ |
| 4× 3070 Ti | 1160 W | 32 GB | ~52 t/s | ~21–23 t/s | ~18–20 t/s | ❌ |
Same story as the 3070 with worse power numbers: 3× reaches the 24 GB 27B-Q4 line at ~17 t/s for 870 W (a 3090 does it at 350 W), and 4× (32 GB) runs 27B Q5 for a kW of GPU. The 12VHPWR adapter and thermals at 4× make this a build you do out of affection, not arithmetic. If you own Tis and want 27B: the 3-card config is the sweet spot; beyond that, add a 3090 instead of a fourth Ti.
When a model doesn't fit, -ngl N keeps the last layers on the CPU instead.
The rule to internalize: offloaded layers run at CPU speed, not a percentage of GPU
speed. A 27B model that's half offloaded is not 50% of GPU speed — it's ~3–5 t/s
total on a modern 8-core. Practical patterns: drop just the last 2–4 layers to the CPU 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 to save another ~0.5 GB on big
models. Budget 32 GB of system RAM for 27B-class offload, 64 GB for 70B. Never offload the
KV cache — the conversation gets unusably slow. Two Tis = 16 GB, so 27B Q4 needs a 3 GB offload — doable with the last ~14 layers on CPU, but expect ~10–12 t/s rather than the 15–17 of the 3-card build. Offloading here is a bridge, not a destination.
llama-server -m 27b-q4_k_m.gguf -ngl 99 -sm layer -c 8192 --tensor-split 1,1,1llama-server -m 27b-q5_k_m.gguf -ngl 99 -sm layer -c 8192 --tensor-split 1,1,1,1
llama-server -m llama-3.1-8b-q5_k_m.gguf -ngl 99 -c 16384 -b 2048 -ub 2048 --host 127.0.0.1 --port 8080
Finally, the "Ti" upgrade that matters for LLMs: GDDR6X doubles the bandwidth to 912 GB/s — 2× the 3070 and near 3080 Ti speed — but the ceiling is still 8 GB. If you own it, it's the fastest 8 GB 30-series server: run an 8B/Q6 or Q4-13B and it'll keep up. If you're choosing, the 3070 saves 70 W but loses 2× the speed, and the 3080 10 GB gives strictly more VRAM at close speed — that one is the better buy for most people.