Independent, reproducible hardware measurement for real hardware decisions.
Labs runs a fixed, versioned benchmark methodology on real hardware, using established tools (llama.cpp / llama-bench) rather than a homemade measurement engine, and publishes the exact command, raw output, and hardware/software evidence behind every number.
6 accepted runs on 3 systems · measured 27–28 September 2026 · evidence level E1 · throughput only
RX 7800 XT vs RX 9070 for local LLMs: same 16 GB, very different throughput
Both cards have 16 GB of VRAM, but the two test systems measured very different throughput. With the 8B model, the RX 9070 system processed prompts about 3.29× faster and generated text about 1.37× faster than the RX 7800 XT system. The host systems were not identical, so the cause of the difference was not isolated.
E1 Throughput only · Cause not isolated
Read the review →
standard-ai-v0.1
RigRoute Labs' first local-LLM-inference methodology. Measures prompt-processing and generation throughput at a fixed context depth, across an ~8B and an ~14B Q4-family model, using llama-bench's own repetition/statistics support. llama.cpp build 11146 (commit 7fe450e19), context 8192, 5 timing repetitions per measurement, within one session.
Apple M1 Pro (32 GB unified memory)
| Model | Prompt processing | Generation |
|---|---|---|
| llama 8B Q4_0 | 190.4 tok/s | 27.9 tok/s |
| qwen3 14B Q4_K - Medium | 97.1 tok/s | 11.7 tok/s |
E1 Measured 27 September 2026 · throughput only · what was not measured
AMD Radeon RX 9070 (16 GB dedicated VRAM)
| Model | Prompt processing | Generation |
|---|---|---|
| llama 8B Q4_0 | 2511.8 tok/s | 102.3 tok/s |
| qwen3 14B Q4_K - Medium | 1177.2 tok/s | 56.4 tok/s |
E1 Measured 27 September 2026 · throughput only · what was not measured
AMD Radeon RX 7800 XT (16 GB dedicated VRAM)
| Model | Prompt processing | Generation |
|---|---|---|
| llama 8B Q4_0 | 763.5 tok/s | 74.8 tok/s |
| qwen3 14B Q4_K - Medium | 417.3 tok/s | 43.5 tok/s |
E1 Measured 28 September 2026 · throughput only · what was not measured