# BTL-3 Compact — RTX PRO 6000 speed

Measured on 2026-07-19 using the native patched `llama.cpp` CUDA runtime and
the frozen `BTL-3-Compact-AVQ2.gguf` release artifact.

## Artifact

- File bytes: `8,392,369,600`
- GPU offload: all layers (`-ngl 99`)
- KV types: FP16
- CUDA architecture: `sm_120a`
- GPU: NVIDIA RTX PRO 6000 Blackwell Server Edition
- GPU memory: 97,887 MiB reported by `nvidia-smi`

## Command

```text
llama-bench \
  -m /vol/release/compact/BTL-3-Compact-AVQ2.gguf \
  -ngl 99 \
  -p 512 \
  -n 128 \
  -r 3 \
  -o json
```

## Results

| Workload | Throughput | Standard deviation |
|---|---:|---:|
| 512-token prompt processing | 84.70 tok/s | 0.37 tok/s |
| 128-token generation | 43.16 tok/s | 0.29 tok/s |

Generation samples: `42.83`, `43.35`, and `43.30` tok/s.

Prompt-processing samples: `84.28`, `84.91`, and `84.91` tok/s.

The measured benchmark body, including model load, completed in 41.03 seconds.

## Consumer-GPU projections

These are projections, not measurements. Single-stream decode is scaled
primarily from memory bandwidth, then widened to account for architecture,
clock, power, kernel occupancy, and driver differences.

| GPU | Estimated generation | Estimated prompt processing | Fit |
|---|---:|---:|---|
| RTX 5090 32GB | 39–44 tok/s | 75–88 tok/s | Yes |
| RTX 4090 24GB | 22–28 tok/s | 45–60 tok/s | Yes |
| RTX 4050 Laptop 6GB | No credible full-GPU estimate | No credible full-GPU estimate | No |

The RTX PRO 6000 and RTX 5090 both advertise 1,792 GB/s memory bandwidth and
share the Blackwell generation, so near-parity is the reasonable batch-one
decode hypothesis. The RTX 4090 advertises 1,008 GB/s, or 56.25% of that
bandwidth. A literal RTX 4050 Laptop has only 6GB of VRAM, below the 8.39GB
weight file before KV cache and runtime allocations; it therefore requires
CPU offload and should be described as a low-single-digit, system-dependent
fallback rather than a supported full-GPU target.