Menu

Blog

The FPGA margin problem.

Vendors must guarantee timing for every chip they ship, so their models describe the worst one. That is the right call for a vendor. It also means every card in your fleet carries that chip's margin, for every year it runs, and here is what per-chip measurement changes.

September 2026Fluid Silicon5 minute read

Variation is a first-order design constraint

Defective and weak transistors show up as variation in the logic you build on: lookup tables (LUTs), carry blocks and interconnect. At 16, 14 and 7 nm, random variation alone exceeds 20%, on top of systematic variation across the die.

  • Random variation. Chip-to-chip and element-to-element scatter. More than 20% at 16, 14 and 7 nm.
  • Systematic variation. Smooth gradients across the die from the manufacturing process.
  • Spatial variation. Neighbouring resources that drift together, so one weak region can hold several critical paths.

These variations force conservative timing models that guarantee the slowest operating corner. Every other chip pays for that guarantee in energy and performance.

Margins are set once, for a chip that is statistically rare

Vendor timing models cover the worst die: highly defective, at elevated voltage, at high temperature and at the end of its life. Most of your chips are nowhere near that corner, yet every one of them runs as if it were. On production shells at realistic utilization we measure 30–54% of timing margin carried for that corner.*

One margin for the fleet, compared with a margin per chipLeft: a wide distribution of path delays across a fleet, with the worst-case corner far to the right of the typical chip, so every chip carries 30 to 54 percent timing margin. Right: three chips measured separately, each with narrow distributions for LUT links and carry links, and each signed off at its own measured delay plus under 20 picoseconds of error and drift.Today: one margin for the whole fleetWorst-case corner30–54% timing margin*typical chipPath delay →Measured: a margin per chipFleet corner, no longer neededChip AChip BChip CPer-chip margin: error + drift, under 20 psPath delay →

Scroll sideways to see the whole chart.

Fleet of chipsLUT linksCarry linksPer-chip sign-off
Illustrative, not to scale. *Fluid Silicon measurements.

A fleet-wide margin

Has to cover the worst chip in the whole population, including chips that never reach your racks.

A per-chip margin

Only has to cover its own measurement error, plus drift until the next measurement. Fluid Silicon measures each link to under 20 ps.

At scale

Measuring more chips finds more tails. It does not add margin to the chips that don't have them.

Service life keeps stretching

Cards now run six years and more in production, beyond what vendors planned for. Large operators keep several board generations in service at the same time. The FPGA 2026 paper on hyperscale FPGA engineering at Microsoft puts it plainly:

“Microsoft therefore maintains many FPGA parts and families in production simultaneously, supporting multiple board generations for a decade or longer. Currently, Azure has 5 generations of FPGA boards actively in deployment.”
Rydberg et al., “Hyperscale FPGA Engineering Systems at Microsoft,” FPGA 2026

The same operator reported more than a million hosts with FPGA-based SmartNICs back in 2018 (Firestone et al., “Azure Accelerated Networking: SmartNICs in the Public Cloud,” NSDI 2018). At that scale, a margin set for the worst chip in the population is a line item.

Silicon as a black box

Without visibility into the real characteristics of each logic component, a failure has no attribution. Teams replace boards on a schedule or after an outage. They cannot prove which cards are healthy, which are drifting or which logic caused a fault. An error in one card can take capacity offline or travel downstream before anyone notices.

Per-element measurement turns each card into evidence: how much margin it has today, how fast it is changing and when it will need attention.

What the uncertainty costs

  • Energy spent on voltage the chip does not need
  • Performance left behind a clock set for the worst die
  • Spare hardware held in reserve, just in case
  • Surprise failures that remove capacity, and errors that propagate before anyone catches them
  • Maintenance and downtime planned around guesses

What changes with measurement

Measure every chip and every element at speed, and the margin you need shrinks to measurement error plus drift until the next measurement. Power and performance is where most fleets start.

*Fluid Silicon measurements. They vary with variation pattern, workload, vendor and device age. Cited figures are for context; Fluid Silicon is not affiliated with Microsoft.

Measure a device at speed, live.

A 30-minute session: in-system slack measurements on a supported AMD or Altera device, and what an evaluation on your boards involves.

Request a demo