AI-TEST-RIG

48-LQFP · HOME INFERENCE EVALUATION

Model_Test_Array

2025.01.15

03:47 UTC

LLAMA-3.1-8B-INSTR

META AI · OPEN WEIGHTS

RTX-3090-24GB

OLLAMA-0.4.0
Q4_K_M

MMLU-Pro GPQA IFEval

NET_OUTPUT

Peak 14.2GB VRAM. 47 tok/s. Six-hour burn-in stable. Quantization noise below measurement threshold. Temp 0.7, top_p 0.9.

2025.01.12

22:15 UTC

QWEN2.5-14B-INSTR

ALIBABA DAMO ACADEMY

RTX-3090+64GB

LLAMA.CPP-B4000
Q5_K_M · MIXED

HumanEval MBPP MultiPL-E LiveCodeBench

NET_OUTPUT

Eight layers CPU-offloaded. 23 tok/s with latency. Code gen exceptional in C++/Python. Rust fail: context window, not model.

2025.01.08

17:33 UTC

MISTRAL-SMALL-24B

MISTRAL AI · PARIS

DUAL-3090-NVLINK

VLLM-0.6.5
BF16 · NATIVE

MATH GSM8K BBH

NET_OUTPUT

24B native, no compression. 48GB aggregate. 89 tok/s — record. Chain-of-thought exceeds all quantized. NVLink overhead negligible.

Roleplaying traces, creative signal paths, character impedance matching, and experimental routing that defies standard netlist constraints.

Drill To Bottom

AI-TEST-RIG-v1.4 · NEOCITIES FAB · LEAD-FREE PROCESS

DATE: 2025 REV: 1.4 ROHS: YES TELEMETRY: NONE