Company News

We are looking for an FPGA partner to prove SARC in silicon

SARC: AI chips that run lean and stay provably exact, or safely redo the work. Never silently wrong.

Patent pending, GB2616933.4

12 September 20266 min read

The story

01 · The problem

The biggest way to save energy in an AI chip is the one nobody uses: just turn the power down. Every engineer knows it. And every engineer also knows what happens next. Push the voltage far enough and the arithmetic starts to wobble. Tiny mistakes. A flipped bit here, a value that comes back almost right. So the industry protects itself the only ways it can. It keeps power in reserve, running the chip well above what it truly needs, or it does the whole calculation twice and checks. Both are honest. Both work. And both quietly burn the very energy they were supposed to save. I kept thinking about how much sits in that gap.

02 · The turn

At some point we stopped asking how to avoid the errors, and asked a different question instead. What if they are not random noise at all? What if a fault has a physical shape, a signature we can measure, predict and work with, the way a sailor reads the sea instead of commanding it?

03 · The solution

That question became SARC. The chip runs lean. And when a small error shows up, SARC reconstructs the correct answer mathematically, right there, in place. Nothing is redone. And it checks itself two independent ways, so every result is either exactly right, or safely recomputed. Never silently wrong. We think that is the honest version of low-power AI.

SARC in action

Press play for the narrated walkthrough. On alternate loops, a fault cluster is too big to correct in place, so one tile takes a detour through the safe redo lane and comes back green.

SARC // RUN LEANSAFE REDOVCORE0.85V
Run lean
Fault appears
Shape it
Correct it
Exact result

Narration uses a pre-recorded natural voice. The labels carry the meaning with sound off.

The benefit, if validated

Provably correct, never silently wrong

Every result is verified and either exactly right or safely recomputed. No silent errors reach the output.

Lower energy in the compute engine

The chip runs in the low-power regime others avoid, because correctness no longer depends on generous voltage headroom.

Safety-grade integrity without duplicating the hardware

Error recovery happens mathematically, in place. No doubled silicon, no full duplicate run to compare against.

These are design goals we are now seeking to prove in hardware.

What we have achieved so far

Modelled and filed, now seeking hardware proof

  1. Approach defined and modelled

    The SARC error-recovery approach is fully defined and modelled in simulation.

  2. Correctness mechanism validated in simulation

    Validated in simulation, with near-zero undetected errors.

  3. UK patent application filed

    Patent application GB2616933.4 filed with the UK Intellectual Property Office.

  4. FPGA hardware validation

    The next step is FPGA hardware validation. That is the step we are seeking a partner for.

If it validates in silicon, this is a rare thing: real energy savings with a mathematical guarantee of correctness, in a regime the industry has always avoided. That is the result we are looking for a partner to help us prove.

The partnership

We are seeking a partner: an FPGA design lab or a university hardware group. Together we would build a small integer systolic array with our error-recovery path on an FPGA, run it first under injected faults and then under real reduced-voltage faults, and measure it against defined thresholds.

You bring the hardware bench and the FPGA toolchain. We bring the method, the models and the maths.

Detailed test specification shared under NDA.

Physivitis Ltd, deep-tech IP for the AI and quantum era. Patent application GB2616933.4. This describes an approach in hardware validation; benefits stated are design goals, not measured results.