Valavan Manohararajah

Founder & CEO at @ExstillAI, Prev Chief Product Architect @cerebras, Principal Engineer @Intel & @AlteraFPGA_ // HW-SW-ML Co-Design // 65 Patents

[Valavan Manohararajah @ValavanMan]

We design the silicon and the model together, so only the math remains.

First result, pre-silicon: the frontier model's co-processor. It runs the small models around frontier AI 110× faster than today's GPUs, with 690× less energy.

https://t.co/z8lHHnugTC

[Valavan Manohararajah @ValavanMan]

In October I said the work around the frontier model had earned silicon of its own. @vithursant19 and I built it. @ExstillAI is out of stealth.

Our belief: AI compute is nowhere near as efficient as it could be. Most of what you pay to run AI isn't the math. It's the mismatch.

[Valavan Manohararajah @ValavanMan]

Almost every model runs on a chip designed before the model existed. The model gets fitted to the chip, or the chip patched for the model, and every compromise costs compute, data movement and energy.

The industry treats that as physics. We think it's a design choice.

[Valavan Manohararajah @ValavanMan]

This work is settling into a few well-defined kinds, at volumes that keep growing.

It has earned silicon of its own.

[Valavan Manohararajah @ValavanMan]

Everyone is building chips to replace NVIDIA.

Hardly anyone is building for the work that surrounds the frontier model. Small models screen, route and fetch, dozens of times in a row per agent task.

The frontier model gets the headlines. That work decides how long you wait.

[Valavan Manohararajah @ValavanMan]

Most data-center chips, even the new ones, are built for the opposite: the biggest model, spread across as much memory as it takes, users batched to keep it busy.

Run a small model on one and you lose either way. Batch it and it waits; don't, and you pay for a mostly idle chip.