Evidence

Benchmark results and methodology.

All figures on this site were measured on NVIDIA H100 with vLLM 0.23, against a baseline with the engine's prefix and KV caching enabled. Latencies are medians of five runs; output identity is verified byte-for-byte against the unmodified engine.

Latency

New request
2,939 ms
Similar request, verified
214 ms
Exact repeat
0.8 ms
Response time by request type, Mistral-7B. Bars to linear scale, truncated.

The exact-reuse path returns in 0.8 ms (0.62 ms at scale) without an engine call. Verified acceleration on similar requests measured at 13.8× on Mistral-7B, 9.2× on Qwen2.5-7B and 3.6–4.9× on Llama-3.1-8B, with zero code changes between model families; the multiple depends on how far into the answer the similarity holds. Non-repeating traffic carries approximately 1% overhead.


Correctness

Every reused answer is confirmed by the serving engine before it is returned. Across three model families, 100% of verified answers were byte-identical to the unmodified engine's output. A 200-request mixed-traffic soak completed with zero failures, and identity held in full under 16-way concurrent load.

100%
byte-identical to the engine, every verified answer
362/362 verified answers · 16/16 under 16-way concurrency

Persistence

574,000 tokens of unrelated traffic time to answer → engine cache — evicted, 60 ms AIe · 0.62 ms — unchanged before after
Exact-repeat latency, before and after 574,000 tokens of unrelated traffic. The engine's cache forgets under pressure; the fabric doesn't.

The memory holds under pressure. After 574,000 tokens of unrelated traffic, AIe's exact-reuse latency was unchanged at 0.62 ms, while the engine's prefix cache had evicted the entry (60 ms) and disk-offloaded KV recovered partially (21.3 ms). Following a full process restart, 89.5% of the memory answered warm.


Scale

The memory answers in about a millisecond with more than a million entries stored — on commodity hardware, with capacity bounded by storage rather than GPU memory. That is the difference in kind: a serving engine's cache holds what fits beside the model in GPU memory and evicts in minutes under load; the fabric is designed to hold months of traffic, and grows with use. On traffic composition: a 50,000-request public corpus of raw consumer chat showed a 15.1% exact-duplicate rate before any similar-request counting; per-deployment rates are measured by the profiler on the customer's own logs.

1,000,000+
entries measured, still answering in about a millisecond
Engine KV cacheGPU memory · minutes of traffic · evicts under load
AIe fabric memorycommodity storage · months of traffic · grows with use

Validation scope

ValidatedIn progress
Open-weight dense models (Mistral-7B, Qwen2.5-7B, Llama-3.1-8B) · vLLM 0.23 · NVIDIA H100 · single-node deployment · exact and similar-request reuse Frontier-scale and mixture-of-experts models · multi-node shared memory · professional-domain reuse profiling · partial reuse within dissimilar requests

Validation status is updated as results complete, each with its date and method.

Run it yourself.

The evaluation pack includes the reproducible harness, the methodology, and shadow-mode deployment on your own traffic — so the numbers you present internally are measured on your workload.