The refresh: 7168 cores, same 12 GB and 672 GB/s envelope. Token generation is unchanged — prompt processing is the gain.
Spec note: the RTX 5070 Super is not listed on NVIDIA’s official comparison page (knowledge/rtx-50-series.md, fetched 2026-09-24). The values above are the standard published board specs, kept as supplementary.
The Super refresh adds 17% more CUDA cores (7168 vs 6144) on the same 192-bit GDDR7 bus. That splits the two phases of inference apart:
The VRAM story is identical to the 5070: 13B Q4_K_M at 32K+ is the comfort zone; 27B is ~7 GB over. If you already own a 5070, the Super is not an upgrade for LLM work — the token speed you live with every day doesn't move.
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
+40–70% (estimated) on token generation; no per-card community A/B exists for this card yet; the estimate follows the community record, where mid-bandwidth cards gain most at n-max 2. 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 Q8_0 | ~9.7 GB | ~75–100 t/s | ~105–170 t/s | ~800–1050 t/s | ✅ 32K ctx |
| 13B Q4_K_M | ~8.5 GB | ~40–55 t/s | ~57–93 t/s | ~850–1150 t/s | ✅ 32K+ ctx |
| 13B Q5_K_M | ~11.9 GB | ~28–38 t/s | ~40–64 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 →
Estimates for full GPU offload (-ngl 99), ~8K context, batch 2048. Token generation scales roughly with memory bandwidth; individual runs vary by model architecture (GQA vs MHA, MoE) and llama.cpp build. Token gen matches the RTX 5070; prompt proc. is ~15% higher.
| Quant | GGUF size | 1× 5070 S (12 GB) | 2× 5070 S (24 GB) |
|---|---|---|---|
| Q4_K_M | ~19 GB | ❌ ~7 GB over | ⚠️ ~22–28 t/s, 8K ctx |
| Q5_K_M | ~22 GB | ❌ | ❌ ~2 GB over |
| Q6_K | ~25 GB | ❌ | ❌ |
| Q8_0 | ~33 GB | ❌ | ❌ |
Identical to the RTX 5070: 27B Q4 fits at 2× (24 GB, 8K context, modest speed); 3× (36 GB) runs Q5 at 8–16K. Token generation in multi-card builds matches the 5070 — the extra cores help prompt processing, not token streaming.
--cache-type-k q8_0 --cache-type-v q8_0) stretches 13B context past 32K.No NVLink: tensor split over PCIe 5.0 with a ~20–30% tax. Same 12 GB tiles as the 5070 — the table below matches its token speeds.
| Build | GPU power | VRAM | 13B Q8 | 27B Q4_K_M | 27B Q5_K_M |
|---|---|---|---|---|---|
| 2× 5070 S | 500 W | 24 GB | ~45–58 t/s | ⚠️ ~22–28 t/s, 8K ctx | ❌ ~2 GB over |
| 3× 5070 S | 750 W | 36 GB | (overkill) | ~26–34 t/s, 16K+ ctx | ~20–25 t/s, 8–16K ctx |
3× (36 GB) is the interesting point — 27B at Q5 with a real context. Same board requirements as the 5070: three working x8/x16 slots, workstation-class in most cases.
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. 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 ~0.5–1 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.
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-q4_k_m.gguf -ngl 99 -c 32768 -b 2048 -ub 2048 --host 127.0.0.1 --port 8080
The RTX 5070 with a ~15% prompt-processing bump and no change to the thing that matters in chat: token speed. As a purchase, it's simple — take the Super when it's the same price, skip it when it isn't. As a platform, everything the 5070 page says applies: a 13B-class card with a 12 GB ceiling.