Overlay Contact

The whole model lives on the chip.

A chip built around the models it runs. The whole model sits on the chip, so it never fetches weights from off-chip memory.

~7,000–8,000tokens/s, one userBefore siliconest. A B200 on a standard engine serves the same model at ~890measured. A tuned GPU engine narrows the gap, and can match us on structured outputest.
Watch the raceTalk to us
IN · embedOUT · head + sampler
Concept render of our first chip. Not silicon. Die count schematic. Drag to rotate.

Serving one user, a GPU waits on memory.

It fetches the weights from memory every step. Ours stay on the chip.

A GPU, serving one user

weights · every stepHBMGPUtokens
B200 · one user · standard engine~890 tok/ssame modelmeasured

The Overlay chip

weights on the chip: all SRAMOVERLAY CHIPtokens
Overlay · one user · pre-silicon~7,000–8,000 tok/ssame modelest.

Every weight on the chip, in SRAM.

Today we build the chip and its compiler around released models. Next, we co-design the model with the chip. Our first chip is all-SRAM, so every weight is rewritable.

  1. Model

    The model as released. Co-designed next.

  2. Compiler

    Places each part of the model.

  3. SRAM WEIGHTSevery weight rewritablePER-USERupdatesfine-tunes

    Our SRAM die

    Every weight in rewritable SRAM.

  4. Verified

    Bit-exact in simulation.

  5. Silicon

    Test chip next.

Same answer, one user, two machines.

An agent reasons through a step. Watch who finishes.

B200 · one user · standard engine~890 tok/smeasured
B200 · one user · tuned engine~2,600–4,400 tok/sest.
Overlay · one user · pre-silicon~7,000–8,000 tok/sest.
Illustrative, slowed 150×. Our speed is an estimate before silicon.

Simulated today. Silicon next.

The design runs bit-exact in simulation today. Test silicon is next. A white paper will follow when the time is right.

  1. NowSoftware and simulation
  2. NextTest silicon
  3. ThenThe product die

Building agents that can't wait? Talk to us.

Email usAddress coming soon