# Performance and benchmarks

> Benchmark results and the measured LLM token savings.

URL: https://docs.hotkv.com/performance/

HotKV's performance numbers live as interactive charts on the homepage,
where you can filter by server and pipeline depth rather than read a
static table; this page summarizes what is there and links to the full
detail.

## Throughput benchmarks

The [Performance section](https://hotkv.com/#performance) of the homepage benchmarks HotKV
head-to-head against Redis 8.6.2 and Valkey 9.0.3 using `memtier_benchmark`,
with identical parameters, across 6 cloud servers (AWS, Hetzner, and four
Google Cloud shapes including an Arm instance), 14 commands, and 2
pipeline depths. It reports commands won, peak throughput, and the
percentage difference against each competitor, filterable per server. Full
methodology and raw results are available on request from
[sales@hotkv.com](mailto:sales@hotkv.com).

## Token savings benchmark

We measured how much HotKV's exact-match and semantic caching save on a
realistic LLM workload, rather than estimating it. The
[token savings section](https://hotkv.com/#token-savings) of the homepage has the measured
numbers; the full methodology, including where the result does and does
not generalize, is in the blog post
[We measured it: HotKV cut GPT and Claude token use by 60%](https://hotkv.com/blog/token-savings-benchmark).

In short: 1,500 customer-support questions from the Banking77 dataset were
sent to `gpt-6-sol` and `claude-sonnet-5`, twice each, once straight to
the model and once with HotKV's exact and semantic cache in front, with
the provider's own prompt caching left on in both runs. Across both
models, HotKV cut the tokens sent to the models by 60% on top of whatever
the providers' own caching was already saving, and cost fell by more than
half. The blog post also covers where this does not hold: a workload with
little or no repeated traffic saves nothing, and a flatter, less realistic
traffic mix in the same test only reached a 41% cache hit rate instead of
the measured 60%.

## Next steps

- [AI features guide](https://docs.hotkv.com/ai-features/index.md) to put the prompt and semantic
  cache to work in your own application.
- [Introduction](https://docs.hotkv.com/introduction/index.md) for an overview of what else ships
  in every plan.