SILICON
THAT REDRAWS
ITSELF

Everyone builds upward from the hardware contract.
We go the other way.

Read the thesis Talk to us
LIVE DIE · 28 × 12 BLOCKS

Flip it to SILMIR and move your cursor across the die.

MOVE CURSOR TO ADD LOAD
Occupancy
—
Work done
—
Saturated
—
Floorplan
FIXED
idle light busy saturated block size = allocated budget
PROCESS ±5–15%DATA INPUT ±20–50%THERMAL ±10–25%TEMPORAL ±5–15%SYSTEM ±15–30%SOFTWARE ±5–20%NETWORK ±25–40%SEVEN MEASURABLE SIGNALS. NONE REACHES THE SILICON. PROCESS ±5–15%DATA INPUT ±20–50%THERMAL ±10–25%TEMPORAL ±5–15%SYSTEM ±15–30%SOFTWARE ±5–20%NETWORK ±25–40%SEVEN MEASURABLE SIGNALS. NONE REACHES THE SILICON.
L7Applications & agents
L6Inference runtimes
L5Kernels & compilers
L4Drivers & firmware
ISA · THE CONTRACT
L1HardwareFIXED AT TAPE-OUTA VARIABLE

01 · THE FROZEN CONTRACT

EVERYONE OPTIMIZES UPWARD.

Compilers. Kernels. Runtimes. Schedulers. Autonomous agents that rewrite all of them. Every layer of the stack has spent a decade bending toward one fixed interface.

THE CHIP WAS FROZEN AT TAPE-OUT.

Below the instruction set the floorplan is a photograph. Compute here, memory there, interconnect between them, in proportions decided years before the workload existed. Nobody points down.

WE GO THE OTHER WAY.

Silmir works downward through the contract. First the sixty-five configuration knobs every production node already ships. Then a fabric whose boundaries move. Then a substrate that learns from the workload running on it.

THAT ASYMMETRY IS THE OPPORTUNITY.

02 · THE COST

THE MACHINE
IS ASLEEP.

<30%occupancy on agent workloads

10–100×spread between devices inside one inference step

Agent loops branch, wait on tools, backtrack. The busiest device does a hundred times the work of the idlest, and the idle ones are still drawing power.

GPU-SECONDS IDLED ON A 1,000-GPU CLUSTER SINCE YOU OPENED THIS PAGE 0
1,000 DEVICES · LIT = DOING WORK · LIVE

Every dark square is silicon you bought, powered, cooled, and idled.

03 · WHERE THE VALUE IS

STATIC SILICON SERVES
THE PREDICTABLE HALF.

Pick a workload. The line is its runtime demand. A fixed floorplan is a straight line drawn once. The further the two diverge, the more of the chip sleeps.

  • Dense pretraining±5%
  • Batch embedding±8%
  • Single-turn chat±15%
  • MoE serving±40%
  • Agentic loops±60%
  • Real-time robotics±75%
  • Scientific discovery∞
Agentic loops A FIXED FLOORPLAN CANNOT FOLLOW IT
DEMAND— FIXED FLOORPLANTIME →

The other half is where everything is going.

04 · THE OPENING

THE CONTROLS ARE
ALREADY THERE.

Every production node ships sixty-five configuration knobs. Prefetchers, cache policies, clocks, power caps, fabric routing. Two to the sixty-fifth combinations. Production searches none of them.

Knobs per node
65
Combinations
265
Searched at runtime
0
One knob, right workload
+45%AMD internal

Act one is a policy that reads kernel signals, writes those knobs, and learns in the idle gaps. Same racks. Same power. Same model. Two to three times the fleet you already own.

It does not need new silicon. It needs a policy.

NODE CONFIG · 65 KNOBS STATIC · 1 STATE
LATENCY / STEP2.00 ms
VS STATIC1.00×
STATES USED1

05 · THE ARC

THREE ACTS.
EACH ONE FUNDS THE NEXT.

  1. I SHIPPING · TODAY'S HARDWARE

    Turn the knobs

    A learned policy in production on the fleet you own. Two design partners. Training data at fleet scale.

    CONFIGURE
  2. II T + 24 MONTHS

    The adaptive fabric

    Compute, memory and interconnect stop being regions fixed at tape-out. FPGA validation, BlackParrot RISC-V harness, shuttle tape-out.

    RECONFIGURE
  3. III DESTINATION

    Hardware that learns

    An on-die learning substrate. The policy resident in hardware. Full-mask ASIC. Hardware and model stop being separate problems.

    SELF-ORGANIZE
Observe
10 µs
Decide
4 µs
Commit
1 cycle
Saturated blocks
shed budget
Idle blocks
take it

06 · WHY SILICON

SOFTWARE STOPS AT THE LINE.
WE CROSS IT.

A new class of inference optimizers rewrites kernels, retunes batching and swaps schedulers, continuously and autonomously. Some of the gains are real. Look closely and most are a bug fixed, a kernel unblocked, a price quoted from the cheapest rental. Every one of them accepts the chip as given.

ABOVE THE LINE · CROWDED · CEILING = THE CHIP AS SHIPPED
kernel rewritersbatching tunersscheduler agentsserving enginescompiler stacksquantization passesspeculative decodersautonomous perf agentsKV cache managersgraph optimizersrouter policiescustom attention kernels
ISA · THE HARDWARE CONTRACT
BELOW THE LINE · EMPTY
KNOBS · FABRIC · SUBSTRATE

OUR PROOF STANDARD

  1. 01

    Same hardware, before and after.

    Our number is the fleet you already own, running the work you already run. No cross-vendor arithmetic. No spot price from the cheapest rental. If it needs a different rack to be true, it is not our claim.

  2. 02

    Not one token changes.

    The policy writes hardware knobs. It never touches weights, kernels or numerics. Output is identical by construction, not by a benchmark run afterward.

  3. 03

    Every decision is logged and reversible.

    Each read, each write, each commit is recorded. An anomaly gate rejects a bad decision before it lands. A partner can replay any run and see exactly why the chip did what it did.

  4. 04

    Gains that compound instead of evaporate.

    A fixed bug is absorbed upstream within a release. A learned policy keeps improving on every run, on every framework, on every chip, because it lives below all of them.

Software-only optimizationSilmir
Where it actsAbove the ISA. Kernels, batching, engines, schedulers.Below the ISA. Knobs, then the fabric, then the substrate.
What it can changeThe code that runs on a fixed chip.The chip's own configuration. Later, its floorplan.
Reacts toThe workload mix, at deploy time.Process, data, thermal, network. Live, per step.
Loop timeHours to days.Microseconds, inside the inference step.
Model outputKernel rewrites touch numerics.Untouched by construction.
EndgameFaster software.Hardware that learns.

Even the loudest software wins trace back to the silicon. The chip was the answer all along. We make it a variable.

07 · THE WORLD AFTER

FOR SEVENTY YEARS
THE CHIP WAS THE FIXED THING.
WE END THAT.

TALK TO US founders@silmir.com
HIRING RTL · ML systems · Kernel/Driver Engineers
WHERE silmir.ai · San Francisco