Gyroscope report

A compiler that knows where the fan is

A Sutro-group pitch, written Tue 22 Sep 2026 (week 39), from an idea dictated Mon 21 Sep 13:11 โ€” and independently named in the room the same evening: "It's like a thermally aware compiler." Prior art checked, own measurements re-verified, and the experiment that would kill it.

What you would be re-inventing (say this first, before someone says it for you)

What actually survives

  1. Data-dependent energy as a compiler cost model. ALUPower measured it as a phenomenon; nobody compiles against it. A scheduler that knows zeros โ‰ช constants โ‰ช random can choose tilings, layouts and encodings to minimise bit flips. That lane is open.
  2. Doing it on silicon with no DVFS loop. Every prior system schedules against a controller. The ET-SoC-1 has none, so the physics lands on the energy axis undistorted. You have the only bench like this in the room.
  3. The framing โ€” your own line from yesterday's post brief: you cannot remove the leak, you can only choose which axis it comes out on. Regulate power and it lands on frequency; pin frequency and it lands entirely on joules. That is sharper than "heat-aware compiler," which is 2003.

Your three measurements, re-verified against the raw files

Measurement Status
A100 8192ยณ: fp16 sustains 1,216.2 MHz vs the 1,410 MHz boost (โˆ’13.7%, +15.9% wall clock); int8 1,274.3 MHz; fp32 and the b1 AND+POPC kernel hold exactly 1,410.0 MHz in every sample at the same 54โ€“60 ยฐC โœ… confirmed in clocks_sm_MHz
At the measured clock the kernel is at 93.7% of scaled peak, not 80.8% of datasheet peak โ€” two thirds of "cuBLAS inefficiency" is the clock, not the code โœ…
Two A100s, same part number: 57.7 W vs 37.5 W idle (55%), ~1% speed difference โš ๏ธ true, but the boards differ in power cap (400 W vs 320 W), driver and VBIOS โ€” not a clean part-to-part comparison, and the 1.19% speed delta is inside your own 4.2% noise
ET-SoC-1: clock pinned at 600 MHz, fp32 TensorFMA 546 cycles/op regardless of data, 38.0 W zeros โ†’ 47.6 W constants โ†’ 64.5 W uniform โ†’ 67.3 W random (~1.8ร—) โœ… two rounds agree within 1.1 W
"The power cap caused the 1,216 MHz" ๐Ÿ›‘ inference, not measurement. All 90 recorded throttle-reason masks decode to NONE because every one is an idle snapshot. Your hottest sample anywhere is 66 ยฐC against an 85 ยฐC target, so no thermal event ever fired. One 100 ms-sampled nvmlDeviceGetCurrentClocksThrottleReasons run closes this โ€” and it is the same one-line fix yesterday's post brief flagged

The pitch, as it should be spoken

A compiler that knows where the fan is.

We built a wall: the chip designer guarantees the chip stays up, the compiler pretends physics doesn't exist. Both sides then lie to each other. Three lies I measured on my own hardware this month.

One: my A100 runs an fp16 tensor-core GEMM at 1,216 MHz, not the 1,410 every cycle-count model assumes โ€” 15.9% of wall clock, gone. Two: on the same board, in the same run, at the same temperature, fp32 and a binary AND+POPC kernel lose zero megahertz. Temperature isn't the variable; power is. Three: on the ET-SoC-1 โ€” a chip with no DVFS loop at all โ€” fp32 matmul takes 546 cycles per op regardless of the data, and burns 38 W on zeros, 67 W on random normal. Same instructions, same speed, 1.8ร— the energy, purely from how many transistors flip.

So the statement isn't "abstractions leak." It's: you cannot remove the leak, you can only choose which axis it comes out on.

Why now: thermal-aware compilation has been patented since 2015 and nobody uses it, because the payoff was milliseconds on a laptop. That changed โ€” training clusters now swing hundreds of megawatts and the grid is the binding constraint โ€” and the compiler's search space is no longer human-sized. We can afford to search placement ร— schedule ร— datatype ร— bit-flip count, which is exactly the chunk that was "too hard" in 2003.

Three questions it must answer: (1) what is the cost model โ€” joules per op as a function of operand values, not opcode? (2) what does it control โ€” placement, clock, encoding, or the shape of the power waveform in time? (3) what does it optimise โ€” time, energy, peak power, or dP/dt?

The smallest experiment that kills it, one afternoon, on hardware you already reach: pin the A100's SM clock across seven steps on the existing Modal harness, rerun the 8192ยณ fp16 GEMM, and log throttle reasons inside the 100 ms sampler. Clean line in 1/f + SwPowerCap โ‡’ there is a budget to schedule against. HwThermalSlowdown โ‡’ the idea is about cooling, not code. In parallel on the ET-SoC-1: same matmul, operands permuted to minimise Hamming distance. If joules move and cycles don't, a compiler has a free lunch it is not taking.

The strongest objection, and the answer

Rupert Wu already gave it out loud at #32: "you're only shaving on milliseconds." Naveen Rao's version would be blunter โ€” the hyperscalers solved this at the layer where the money is, without breaking any abstraction.

Concede the time axis entirely. Yes: on speed this is milliseconds and the datacenter people own it. The claim is on the energy axis, where nobody is scheduling: two identically-named A100s differ 55% in idle power, >99% of a short MNIST run's joules are the idle floor, and on the ET-SoC-1 the same instruction sequence at the same cycle count costs 1.8ร— depending on operand bits. Runtime DVFS cannot touch that โ€” by the time the runtime sees the workload, the bits are already chosen. Only a compiler picks the layout, the tiling and the encoding that decide how many transistors flip.

"An instruction takes N cycles" leaks 14%. "An instruction costs N picojoules" leaks 80%.

That sentence is the talk, and it is the one number in your data nobody else in that room has measured.