CCGT Heat Rate Demo  ·  FP64 vs FP32

Same optimizer. Same data.
Same loss function.
Different machine — different model.

FP64 (CPU) vs FP32 (GPU) on a quadratic heat rate curve.
Three parameters. High school algebra. $545,000/yr of consequence.

FP64  ·  CPU FP32  ·  GPU N = 350 observations 200 bootstrap samples L-BFGS-B · same optimizer
scroll to explore
01 · The Model

A gas plant heat rate curve — the simplest possible test

The Augmented Willan's Line is the industry standard model for how a gas turbine's fuel efficiency varies with power output, temperature, and equipment wear.

HR(P, T, d) = (a + bP + cP²) · (1 + α·ΔT) · (1 + β·d)

Three main parameters (a, b, c) define the U-shaped curve. P* = −b/2c is the economic optimum — the power level where the plant burns fuel most efficiently. The model is fit to 350 simulated plant operating observations with realistic measurement noise.

True P_opt
350
MW — efficiency minimum
True HR_min
7.80
MMBtu/MWh at P_opt
Observations
350
biased toward mid-load dispatch zone
Gas price
$4.50
per MMBtu
A · Heat Rate Curves — FP64 vs FP32 vs Ground Truth
True HR FP64 fit FP32 fit Thermo floor 7.5
02 · The Results

Two megawatts of disagreement. $545,000 of consequence.

The only variable between the two runs was arithmetic precision. The optimizer, loss function, data, and random seeds were identical.

FP64 Economic Optimum
MW · double precision
FP32 Economic Optimum
MW · single precision
Annual Dispatch Cost Gap
@ 350 MW · 8,000 hr/yr · $4.50/MMBtu · one plant
FP64 Dispatch Cost
$/MWh at 350 MW
FP32 Dispatch Cost
$/MWh at 350 MW
True Dispatch Cost
$/MWh — ground truth
03 · The Tradeoff

FP32 is a better statistician.
FP64 is a better engineer.

Neither platform chose this tradeoff. The arithmetic made it for them. FP32 fits the observed data more closely. FP64 satisfies the physical and regulatory constraints more tightly. These are two different models of the same plant.

FP32 — Statistical Fit
L_fit (MSE) — lower is better data fit
FP64 — Statistical Fit
FP64 fits the data worse — accepts the statistical cost
FP64 — Physics Violation
L_phys — lower means tighter constraint satisfaction
FP32 — Physics Violation
FP32 incurs more physics penalty — looser constraints
The implicit hyperparameter: Hardware acts as an unset, undocumented hyperparameter that silently decides the fit-vs-physics tradeoff. The same loss function, on the same data, produces different tradeoff positions depending on the arithmetic precision of the machine.
B · Where 30 Restarts Land: (L_fit, L_phys) scatter
FP64 restarts FP32 restarts
04 · Bootstrap Uncertainty

200 resamples. The spread is not noise.

200 bootstrap samples per platform, each refit from scratch. FP32 is consistently less certain about where the economic optimum is — a 52% wider uncertainty band on the number that determines dispatch bids.

FP64 P_opt std
MW across 200 bootstrap samples
FP32 P_opt std
MW — x wider
FP64 cost std
$/MWh dispatch cost uncertainty
FP32 cost std
$/MWh dispatch cost uncertainty
C · P_opt Distribution
D · Dispatch Cost Distribution
05 · Pareto Frontier

The same tradeoff surface — traced differently by each platform

By sweeping the weight on statistical fit vs physics penalty, we trace a Pareto frontier for each platform. Same loss function, same sweep. The two frontiers diverge — hardware is constituting which tradeoffs are reachable.

E · Pareto Frontier: Statistical Fit vs Physics Constraint Violation
FP64 frontier FP32 frontier
06 · Regulatory Compliance

FP64 sits twice as close to the violation boundary

FERC requires submitted heat rate curves to be non-decreasing above the economic minimum (dHR/dP ≥ 0). Both models satisfy this rule. But FP64 has x less margin than FP32 at rated load.

If you fit on a CPU workstation and deploy on an embedded FP32 energy management system — the standard plant configuration — that FP64 margin shrinks further on the way out the door. You could cross into a violation without changing a single line of code. The hardware changed.

FP64 Mono Margin @ rated
dHR/dP at 480 MW — must be ≥ 0
FP32 Mono Margin @ rated
FP32 more conservative — x larger margin
F · Incremental Heat Rate dHR/dP  ·  FERC must be ≥ 0 above eco-min
True FP64 FP32 Boundary (0)

Full Results Table

MetricFP64FP32|Δ|
07 · Takeaway

Hardware is an implicit hyperparameter

This is not a deep learning phenomenon. No neural networks, no stochastic gradient descent, no random initialization. A quadratic equation fit to 350 points with a deterministic optimizer — and the hardware still constitutes the model.

The desktop era of energy and transportation modeling kept this invisible: same machine builds it, same machine runs it, hardware is constant. That era is ending. Models now cross hardware boundaries at every step — trained on cloud GPUs, deployed on embedded ARM controllers, audited from laptops.

Each crossing is a silent reset of this implicit hyperparameter. The practical fix: validate on the hardware the model will actually run on, before you report results or submit a regulatory curve. Add the hardware platform to the model card alongside the software version and the data version.

The computer isn't just where your model runs. It's part of what your model is.

Methodology: Augmented Willan's Line · L-BFGS-B optimizer (identical on both paths) · FP32 simulated via explicit float32 dtype cast before each loss evaluation · N=350 synthetic CCGT observations (seed=42) · 200 bootstrap resamples · 30 random restarts per platform · multi-objective loss: MSE + physics penalty + smoothness regularisation