Research

Receiver reorder memory for multipath transport

• In simulation, our receiver tracked out-of-order data in about half the memory of our simulator's model of STrack, a published design, on average; a second design, not cited here, came closer and used less in nearly half the settingsIn simulation, our receiver needed about half the memory of our model of STrack on average; that model used less in some settings, an uncited design in more.

The Mooncake trace: hit-rate ceiling versus LRU

• On a test slice of our excerpt of a public trace of real AI requests, no cache that starts empty and fetches only on request can beat a proven hit-rate ceilingA machine-checked hit-rate ceiling for caches that start empty and fetch only on request, on a test slice of our excerpt of the public Mooncake trace.

Placing KV cache across a GPU fleet

• Pooled cache placement reaches a solver-certified optimum on one fixed test instanceOn one fixed test instance where hosts keep what they fetch, a pooled KV cache reaches the fewest misses the lab's model allows, confirmed by an outside solver.

Questions, answered plainly

What is a KV cache?

When a language model reads a prompt, it stores intermediate results from every layer so it does not have to recompute them for each new word it writes. That stored working state is the KV cache (short for key-value cache). Serving systems reuse it between requests that begin the same way, which saves GPU time.

What does "machine-checked" mean?

Our proofs are written in a proof assistant called Lean. A program checks every step, so a result does not depend on anyone’s reading of the argument. A proof still covers only what it states, and each result page says what it does not claim.

What is a cache hit-rate ceiling?

It is the largest share of lookups any cache of its kind could answer from memory on a given stream of traffic. A cache that starts empty and fetches only on request cannot serve more than the traffic repeats. The bound is proved in Lean; its value on a test slice of one public production trace comes from our counts of that slice, and it is a property of that trace, not of any implementation. A cache that starts warm, or one that prefetches, can go higher.

Are these simulations or measurements on hardware?

It depends on the result, and each page says which. The reorder-memory result is a network simulation, not silicon. The placement result is an optimum on one fixed test instance, confirmed by an outside solver on our model of that instance. The hit-rate ceiling belongs to one public trace.

Can I check the work myself?

Every figure on this site opens to the file it was measured from, and the checker on the home page tests a file in your own browser, without sending anything over the network. The checker confirms that a file is unaltered. Re-running the underlying experiments needs the lab’s code, which is not public yet.

Who is AxiomLimit for?

Teams that run or buy AI inference at scale, makers of network and accelerator hardware, and the auditors who examine them. To license or acquire a result, write to the address at the bottom of any page.