# Gemma 2 27B Instruct on an A100 40GB (SXM)

Yes, quantized. Q8_0 is the largest that fits, leaving room for 8k tokens of context.

Canonical page: https://makerportal.ai/lab/llm-vram/gemma-2-27b-it/a100-40gb
Page title: Gemma 2 27B Instruct VRAM on A100 40GB — fits at Q8_0

This markdown document and the HTML page above are rendered from the same solved values at build time, by the same functions. Nothing here is written by a language model and nothing is fetched at request time.

## Key figures

- **Verdict:** Q8_0 — 36.8 GiB usable of 40 GB
- **Usable memory:** 36.8 GiB — 92% of the card's 40 GB — an assumption for a dedicated card, not a single fraction applied to every device
- **Weights at FP16:** 50.7 GiB — 27.2B parameters
- **Weights (Q4_K_M):** 15.5 GiB — 50.7 GiB at FP16 · 27.2B params · 6 of 6 quant rows are measured files
- **KV cache:** 368.0 KiB per token at FP16 — 46 layers × 16 KV heads × 128 dimensions
- **Speed ceiling:** 49 tok/s — 1555 GB/s ÷ bytes read per token
- **Largest context that fits:** 8,192 tokens at Q8_0
- **Trained context:** 8,192 — the cap the KV-cache figures are held to
- **Card bandwidth:** 1555 GB/s — 40 GB HBM2

## Every quant, against this card

Every quantization of Gemma 2 27B Instruct against the 36.8 GiB usable on an A100 40GB (SXM). "Measured" rows are the byte size of a real published file; "nominal" rows are the parameter count times the published bits-per-weight.

| Precision | Bits/weight | Weights | Source | Fits | Max context | Ceiling |
|---|---|---|---|---|---|---|
| FP16 / BF16 | 16.00 | 50.7 GiB | measured file | short by 13.9 GiB | — | 4.4 tok/s offloaded |
| Q8_0 | 8.50 | 27.0 GiB | measured file | yes | 8,192 (model cap) | 49 tok/s |
| Q6_K | 6.57 | 20.8 GiB | measured file | yes | 8,192 (model cap) | 61 tok/s |
| Q5_K_M | 5.70 | 18.1 GiB | measured file | yes | 8,192 (model cap) | 69 tok/s |
| Q4_K_M | 4.89 | 15.5 GiB | measured file | yes | 8,192 (model cap) | 79 tok/s |
| Q3_K_M | 3.94 | 12.5 GiB | measured file | yes | 8,192 (model cap) | 94 tok/s |

## Questions this page answers

### How much VRAM does Gemma 2 27B Instruct need?

50.7 GiB for the weights at FP16 — 27.2B parameters at two bytes each — and 15.5 GiB at Q4_K_M. The KV cache is on top of that and is not a fixed number: this model spends 368.0 KiB per token of context, so 8,192 tokens costs a further 2.9 GiB. An A100 40GB (SXM) makes 36.8 GiB of its 40 GB available on the assumption below.

### Can an A100 40GB (SXM) run Gemma 2 27B Instruct?

Yes, quantized. Q8_0 is the largest that fits, leaving room for 8k tokens of context. That is the weights and the KV cache together, against 36.8 GiB of usable memory.

### How fast will Gemma 2 27B Instruct run on an A100 40GB (SXM)?

No faster than 49 tokens/second at Q8_0, and in practice below it. Decoding is memory-bound: every token reads all 27.0 GiB of weights plus the cache, and this card moves 1555 GB/s. That division is the ceiling — no kernel, runtime or driver beats it, and a real runtime typically reaches 60–80% of it.

### Why does the context length change how much memory Gemma 2 27B Instruct needs?

Because the KV cache holds one key and one value vector per token, per layer, for the whole conversation, and it is allocated separately from the weights. This model has 46 layers and 16 key/value heads of 128 dimensions, shared across 32 query heads — grouped-query attention, which divides the cache by 2. That works out at 368.0 KiB per token. Parameter count tells you nothing about this number.

## Method and limits

Weight bytes are the byte size of the real published file wherever one exists, and the model's exact parameter count times the published llama.cpp bits-per-weight where it does not. That distinction is on every row above and it matters at both ends: a sub-1B model's Q4_K_M file runs a third larger than the nominal figure because k-quants keep its embedding tables at higher precision, and an already-4-bit release cannot be quantized upward at all. The KV cache is 2 · layers · kv_heads · head_dim · bytes per token, summed over layers with each sliding-window layer capped at its window. The speed figure is a roofline bound, not a benchmark: bandwidth divided by bytes read per token, which no runtime exceeds and every runtime falls short of. The usable fraction of card memory is an assumption: 92% here, which is what this lane assumes for a dedicated card — the other class assumes 75%, so it is not one number applied to every device. That is the only assumed input on this page; every other figure is computed from the model config and the card's published specification. Nothing on this page is written by a language model.

## Related tool

[LLM VRAM & KV-Cache Footprint Calculator](https://makerportal.ai/lab/llm-vram-kvcache-calculator) — Interactive VRAM planner — change the model geometry, quantization, context length and card capacity and read the weight bytes, KV-cache bytes and the bandwidth ceiling on decode speed.

---

Source: MakerPortal — https://makerportal.ai/lab/llm-vram/gemma-2-27b-it/a100-40gb. Free to quote and cite with attribution and a link to the canonical page.