AI Chip Power and Exact Compute

SARC-E

Run AI chips at lower power and keep every answer exact, even when a whole power line fails.

Lowering the voltage on an AI chip saves power, until the chip starts making mistakes. The worst arrive in bursts: one sagging power line can corrupt a whole row of results at once. SARC-E repairs them inside the chip's maths engine, so every answer stays exact and nothing wrong slips through silently.

UK Patent Pending · GB2627785.5

Illustrative example

Repaired

Maths engine

Line sensorsRepair budgetResult status
Line 01Normal
Line 02Normal
Line 03Normal
Line 04Repaired
Line 05Normal
Line 06Normal
Line fault detected · repaired in place
EXACTCHECKED TWICENO REPLAY NEEDED

The Challenge

Lower voltage saves power. Faults decide how far you can go.

Scattered faults

At the edge of timing, single results start to go wrong. Standard error checks are built for these, and handle them well.

A sagging power line

When one supply rail droops, every unit on it can fail together, and a whole row of results goes wrong in the same instant.

A slipping clock branch

When one clock branch arrives late, a whole column of results is corrupted at once, far more than a standard check can repair.

Today's error checks are built for scattered faults, and a failing line overwhelms them. So chips keep wide safety margins, and that margin costs power on every calculation. In a car or a medical device, one silent wrong answer is one too many.

How It Works

One Maths Engine. Every Line Watched. Every Answer Exact.

01

Shape

Check groups follow the chip's physical wiring, so no single power line or clock branch can put more errors into a group than the group can repair.

02

Flag

Small sensors watch every line. A line that starts to sag is flagged, so the decoder knows exactly where the damage is and the same checks repair far more.

03

Confirm

When the flags use up the repair budget, a second, independent check confirms the result before it leaves the chip. SARC-E never guesses.

04

Recover

Anything still in doubt is replayed at safe voltage, and the voltage controller keeps the chip where every result stays recoverable.

Works alongside the computation inside the maths engine. Nothing in the AI model has to change.

Applications

Three Ways SARC-E Protects Your Silicon

Datacentre accelerators

Hold the power budget on silicon you already run at scale.

  • Whole failing lines repaired in place, not just scattered errors
  • Voltage set by what the chip can still repair, not by a fixed margin
  • No change to the AI models running on the silicon

Edge AI silicon

Exact answers inside a tight power and thermal envelope.

  • Exact compute at reduced voltage
  • Line sensors that point the decoder straight to the damage
  • Safe replay whenever the repair budget runs out

Automotive and medical AI

One silent wrong answer is one too many.

  • A second, independent check before any result is released
  • SARC-E never guesses
  • Built for safety-critical AI compute

Watch the explainer

One walkthrough covers all three.

Illustrative modelled behaviour

SARC-E walkthrough

01 / 05

SARC // RUN LEANSAFE REDOVCORE0.67V
Run lean
Fault appears
Shape it
Correct it
Exact result

Outputs

What SARC-E Delivers

Line-Fault Correction

Whole failing lines repaired in place, not just scattered errors.

Line Monitoring

Sensors that point the decoder straight to the damage.

Independent Second Check

A separate confirmation whenever the repair budget is used up.

Safe Replay

Anything still in doubt recomputed at safe voltage.

Recoverability-Targeted Voltage

Voltage set by what the chip can still repair, not by a fixed margin.

No Model Changes

Works inside the maths engine; AI models run as they are.

Why SARC-E

Built for exact answers, not approximate ones.

Exact, not approximate

Every result is either exactly correct or replayed at safe voltage. Nothing wrong leaves the chip silently.

Checked twice, never guessed

Any result that uses up the repair budget is confirmed by a second, independent check before it is released.

Inside the maths engine

SARC-E works alongside the computation. The AI model does not change.

Built For

AI Accelerator and GPU MakersHyperscalers and DatacentresAutomotive Chip SuppliersMedical Device MakersMobile and Edge SoC DesignersChip IP VendorsFPGA and ASIC Design Houses

Status

Validated in simulation · Validation programme open · UK patent pending GB2627785.5

Physivitis is selecting a small number of chip partners to validate SARC-E on FPGA and test silicon, measuring real line-fault behaviour on their own designs before licensing. Validation partners receive preferential licensing terms, including the option of an exclusive licence.

Validation Programme

Apply for the SARC-E Validation Programme

Validation partners shape SARC-E for their silicon and receive preferential licensing terms.

Prefer email? contact@physivitis.tech

SARC-E · UK Patent Pending GB2627785.5 (priority GB2616933.4) · Physivitis Ltd

We use a small number of cookies to make this website work and to understand how it is used. You can accept them or continue with only the essential ones. Read our Cookie Policy.