Benchmarks & System Performance

Performance & Quality Benchmarks

Standardized, reproducible performance measurements of NeuronScope across LLMs, RAG retrieval pipelines, autonomous agents, and hardware architectures.

Hardware Device7 / 7 benchmarks

Model / System	Params	Device / Adapter	Telemetry Profile				Hallucination Rate	Commit
bert-base-uncased	110M	NVIDIA A100 40GB	Full Tensor Activations	8,400	980	72	N/A	9c1a04f
gpt-4o (API Adapter)	N/A	Cloud API Gateway	Prompt/Response & Tool Spans	4,200	320	45	0.3%	9c1a04f
claude-3-5-sonnet	N/A	Cloud API Gateway	Prompt/Response & Context Spans	3,850	290	42	0.2%	9c1a04f
distilbert-base	66M	Apple M3 Max	Full Tensor Activations	3,200	610	41	N/A	9c1a04f
llama-3.1-8b-instruct	8B	NVIDIA H100 80GB	All-Layer Activations & Traces	2,840	16,400	840	0.8%	9c1a04f
mistral-7b-v0.3	7B	NVIDIA A100 80GB	Activations + Attention Heads	1,950	14,800	620	1.1%	9c1a04f
llama-3.1-70b-instruct	70B	8x H100 SXM5	Layer Stride 4 + Tool Hooks	920	142,000	3,200	0.4%	9c1a04f

Benchmark Methodology: Measurements captured under median of 5 independent evaluation runs with warm GPU cache. Token throughput includes zero-copy streaming telemetry overhead. Every row links to the sealed .nsz artifact digest and GitHub commit hash.