gemma4-research/docs/reference/gpu-bakeoff-2026-04-20.md

# GPU Bakeoff — Gemma 4 Throughput Across Three Architectures

**Date:** 2026-04-20
**Host matrix:** steel141 (RTX 3090 Ti) · pve197 CT 105 (Tesla V100) · matt-strix (AMD Strix Halo iGPU)
**Models:** `gemma4:26b` (MoE Q4_K_M) · `gemma4:31b-it-q4_K_M` (dense Q4_K_M)
**Harness:** `scripts/gpu-bakeoff/harness.py`
**Raw data:** `scripts/gpu-bakeoff/runs/`

---

## TL;DR

| GPU | 26B (MoE) decode | 31B (dense) decode | Long-prompt prefill (26B) |
|-----|------------------|--------------------|-----------------------|
| **RTX 3090 Ti** (steel141) | **128 tok/s** | **27 tok/s** | **23,849 tok/s** |
| **AMD Strix Halo iGPU** (matt-strix) | 54 tok/s (42%) | 11 tok/s (39%) | 14,326 tok/s (60%) |
| **Tesla V100** (pve197) ⚠ | 8 tok/s (6%) | 1.6 tok/s (6%) | 2,696 tok/s (11%) |

> ⚠ **V100 numbers reflect degraded conditions — SDXL on CT 167 occupies
> 31.7 / 32.7 GB VRAM, forcing Ollama's Gemma 4 models 95% onto CPU.**
> Under isolation, V100 should land between 3090 Ti and Strix based on
> raw specs (HBM2 ~900 GB/s). See § "V100 caveat" for the evidence.

### Headline findings

1. **MoE changes everything.** `gemma4:26b` decodes **~4.7× faster** than
   `gemma4:31b` on every GPU tested, because only ~4 B of its 25.8 B
   parameters activate per token. Total parameter counts (26 B vs 31 B)
   don't predict latency; *active* parameters do.
2. **3090 Ti wins decisively on decode.** For inference workloads the
   memory-bandwidth-flop ratio of consumer Ampere GDDR6X is hard to
   beat at this price point.
3. **Strix Halo punches above its bandwidth.** Gets 42 % of 3090 Ti
   decode speed on only ~25 % of the memory bandwidth (~256 GB/s vs
   ~1008 GB/s) — good SIMD utilization, especially on the MoE model.
4. **V100 is held back by shared VRAM.** Its spec should put it closer
   to 3090 Ti than to Strix, but coresident SDXL crowds out Ollama's
   layer offload. The V100 column in this doc is an *as-is* reading,
   not a *peak-capability* reading.

---

## Hardware inventory

| Host | GPU | VRAM | Bandwidth | Compute cap | Notes |
|------|-----|------|-----------|-------------|-------|
| steel141 | RTX 3090 Ti | 24 GB GDDR6X | ~1008 GB/s | 8.6 (Ampere) | Seth's workstation. Also has a GTX 1660 SUPER as aux display card — not used for inference. Ollama on 127.0.0.1:11434. |
| pve197 CT 105 | Tesla V100-PCIE-32GB | 32 GB HBM2 | ~900 GB/s | 7.0 (Volta) | LXC with GPU passthrough. Ollama on 192.168.0.179:11434. **Coresident with CT 167 ai-visualizer (SDXL) which consumes most of the VRAM.** |
| matt-strix | AMD Strix Halo (Radeon 890M iGPU + XDNA 2 NPU) | Shared LPDDR5X | ~256 GB/s | — | Unified memory lets it fit models a 24 GB card can't. Ollama on 100.117.155.64:11434 via Tailscale. |

---

## Methodology

- Each (host × model × prompt-length) cell:
  - 1 warm-up call (discarded, absorbs model load time and JIT warm-up)
  - 3 measurement calls
  - `temperature: 0.0`, `top_k: 1` (greedy), `num_predict: 256`, `num_ctx: 4096`
  - `keep_alive: 10m` so the model stays resident between runs
- Two prompt lengths:
  - **short** (~15 tokens) — isolates decode performance, prefill time is negligible
  - **long** (~500 tokens) — stresses prefill (prompt evaluation)
- All timings come from Ollama's own `/api/generate` response fields
  (`prompt_eval_duration`, `eval_duration`, etc.), so HTTP and wall-clock
  jitter are excluded from the rates.
- Median of the 3 measurement runs is reported in tables; min/max are in
  the raw JSON.
- **No network-introduced variance** — all three hosts exposed HTTP
  Ollama endpoints (matt-strix via Tailscale). The timings reported are
  computed server-side from `prompt_eval_count / prompt_eval_duration`
  and `eval_count / eval_duration`.

---

## Full results

### Decode rate (tok/s, median of 3 runs)

Decode is the metric that matters most for interactive LLM use — it's
the speed of token generation after the prompt has been processed.

| Model | 3090 Ti | V100 ⚠ | Strix Halo |
|-------|---------|-------|------------|
| gemma4:26b (MoE, ~4 B active) | **128.20** | 8.34 | 53.86 |
| gemma4:31b (dense, 31.3 B active) | **27.15** | 1.55 | 10.64 |

### Prefill rate (tok/s, long ~500-token prompt, median)

Prefill is the cost of ingesting the prompt and populating the KV cache
before decode begins. Batched per-token, so short-prompt prefill numbers
are noisy (dominated by fixed overhead — see raw JSON for those); the
long-prompt numbers below are the ones to reason from.

| Model | 3090 Ti | V100 ⚠ | Strix Halo |
|-------|---------|-------|------------|
| gemma4:26b (long) | **23,849** | 2,696 | 14,326 |
| gemma4:31b (long) | **7,716** | 436 | 3,278 |

### Short-prompt prefill (for reference)

On a 15-token prompt, prefill tokens/sec is meaningless — prompt is too
small to amortize overhead. Included only to confirm no regression.

| Model | 3090 Ti | V100 ⚠ | Strix Halo |
|-------|---------|-------|------------|
| gemma4:26b (short) | 2,063 | 240 | 1,276 |
| gemma4:31b (short) | 661 | 41 | 292 |

---

## V100 caveat — why the numbers are degraded

Mid-bakeoff I probed `GET /api/ps` on pve197 while the V100's Q8 MoE was
loaded:

```
gemma4:26b-a4b-it-q8_0   size: 30.5 GB   size_vram: 1.57 GB
```

**Only 1.57 GB of the 30.5 GB model is actually resident on the V100;**
the other 28.9 GB is running on CPU via Ollama's CPU-offload fallback.
`nvidia-smi` corroborated: 31,754 / 32,768 MiB used, 0 % utilization
at probe time. That remaining ~29 GB of VRAM isn't free — it's held by
the SDXL pipeline on CT 167 (claude-avatar + ai-visualizer).

Impact on every V100 number in this doc:
- `gemma4:26b` Q4_K_M is 18 GB — doesn't fit in the ~1 GB of headroom
  SDXL leaves, so it runs largely on CPU. Observed 8.3 tok/s is
  consistent with CPU inference of a MoE 26B Q4 model.
- `gemma4:31b` Q4_K_M is 19.9 GB — same fate. Observed 1.55 tok/s is
  consistent with dense 31B on CPU (dense kills you on CPU; only
  ~4 B params activate on the MoE, so the MoE suffers less).
- The Q8 variant (28 GB) never had a chance on the V100 while SDXL is
  loaded. Bakeoff did not attempt it.

**To get isolated V100 numbers**, stop SDXL on CT 167 (or stop CT 167
entirely) and re-run `scripts/gpu-bakeoff/harness.py --host pve197`.
Left as a follow-up — whether that's worth the ai-visualizer
interruption is a judgment call. See "Open questions" below.

---

## Why 26B decodes 4.7× faster than 31B

`gemma4:26b` is the MoE variant ("A4B" in Google's naming = *activated
4B*). Per-token inference routes through only ~4 B of its 25.8 B total
parameters. `gemma4:31b` is dense: every one of its 31.3 B parameters
participates in every token's forward pass. Memory bandwidth is the
binding constraint for decode, so the ratio of *active* params is what
you actually pay for.

Rough math (3090 Ti, 1008 GB/s, Q4_K_M ≈ 0.5 bytes/param):
- 26B MoE: 4 B × 0.5 B = 2 GB per token. Theoretical max ≈ 504 tok/s.
  Observed 128 tok/s = **25 % utilization**.
- 31B dense: 31.3 B × 0.5 B = 15.65 GB per token. Theoretical max ≈
  64 tok/s. Observed 27 tok/s = **42 % utilization**.

So dense workloads actually extract *higher* bandwidth utilization —
they're less overhead-dominated per token. But in absolute terms, MoE
wins by a large factor because the active-parameter bill is much
smaller. For interactive chat this is decisive: Seth's `mort-bot`
running `gemma4:26b` gets ~4.7× the responsiveness it would on
`gemma4:31b`, even though the models are near-equal in total params.

Why the ratio holds on every GPU: **memory bandwidth is the bottleneck**
across all three cards. Strix gets 42 % of 3090 Ti on 26B and 39 % of
3090 Ti on 31B — identical ratios — because it has ~25 % of the
bandwidth and matches or exceeds proportionally.

---

## When to use which GPU

**Interactive chat / agent workloads (decode-heavy).**
  - Primary: **3090 Ti** — by a wide margin. 128 tok/s on 26B is
    comfortable for real-time responses.
  - Fallback: **Strix Halo** — 54 tok/s is usable. Benefit is unified
    memory can host larger models the 24 GB 3090 Ti can't.
  - Avoid: V100 *while SDXL is coresident.* Without SDXL it should be
    competitive.

**Long-context / prompt-heavy workloads (prefill-heavy).**
  - Primary: **3090 Ti** again — 23,849 tok/s prefill means a
    500-token prompt ingests in ~21 ms.
  - Strix at 14,326 tok/s is ~35 ms — still interactive.

**Running models that don't fit elsewhere.**
  - Strix Halo. Unified LPDDR5X can hold 80 GB+ models that 24 GB and
    32 GB discrete cards can't — at the cost of lower bandwidth.
  - The largest model tested here (`gemma4:31b` Q4 at 19.9 GB) fits
    all three. Q8 variants (28 GB+) only fit the V100 and Strix.

**Fine-tuning / training.**
  - Not measured here. 3090 Ti's 24 GB limits batch size on 20 B+
    models; V100's 32 GB HBM2 is much more forgiving *if* isolated.

---

## Open questions / follow-ups

1. **Isolated V100 re-run.** Stop SDXL, re-run the harness. Expected
   outcome: V100 decode lands between 3090 Ti and Strix (probably
   ~70-90 tok/s on 26B given HBM2 bandwidth ~900 GB/s vs 3090 Ti's
   ~1008 GB/s). That would settle the V100's actual rank.
2. **V100 Q8 baseline.** `gemma4:26b-a4b-it-q8_0` (28 GB) is the Q8
   MoE variant Seth pulled on pve197 — worth measuring once isolated.
   Q8 vs Q4 quality/speed tradeoff for the same model would be useful.
3. **Strix max-model fit.** Strix can probably host models that
   wouldn't fit the discrete cards. A follow-up would pull a larger
   model (70 B+ quantized) on matt-strix and see the Strix-only
   performance ceiling.
4. **Contention behavior.** The V100 finding generalizes — whenever
   the homelab is running coresident AI workloads, Gemma 4 inference
   falls off a cliff. A "contention-aware routing" decision (don't
   send latency-sensitive Ollama traffic to a card with SDXL running)
   may be worth building into the mort-bot / openwebui gateway.

---

## Raw data

All per-run JSON traces are under `scripts/gpu-bakeoff/runs/`:

```
runs/
├── steel141/
│   ├── gemma4-26b/{short,long}.json
│   ├── gemma4-31b/{short,long}.json
│   └── gemma4-26b-q8/{short,long}.json    # skipped — model not on host
├── pve197/
│   ├── gemma4-26b/{short,long}.json        # ⚠ degraded, see caveat
│   └── gemma4-31b/{short,long}.json        # ⚠ degraded, see caveat
└── matt-strix/
    ├── gemma4-26b/{short,long}.json
    ├── gemma4-31b/{short,long}.json
    └── gemma4-26b-q8/{short,long}.json    # skipped — model not on host
```

Each JSON contains the warmup call and all 3 measurement calls with
every field Ollama's `/api/generate` returns (token counts, durations,
loaded-at, context length), plus a `summary` with min/median/max for
prefill and decode rates.