Skip to content
Slow Variables
Ask

Methods · Compute & physical

Hardware price-performance growth (doubling time)

Doubling time of FP16 tensor throughput per release-price dollar across accelerators with a public price, fitted log-linearly from 2012 on Epoch's hardware table. Derived metric perf_per_dollar_doubling_days.

Reason through this

The reading

Only accelerators with a public list price enter, and recent data-centre GPUs have none, so the series can end years before today.

Every readingeach dot links to its record
days
-7.9 21 0.361 · 1331 days · 2026-08-02 · reported
a readingnormal rangefast range
reported

1 reading drawn from the derived metric perf_per_dollar_doubling_days; 42 observations in the evidence.

Derived (perf_per_dollar_doubling_days) from: Epoch AI ML hardware. Epoch AI, Machine Learning Hardware (CC BY 4.0).

The numbers
as ofreadingvaluerecord
2026-08-02-7.9 21 0.3611331 daysreported

Why it matters

The floor under falling inference prices; the brief's reading is +49% a year.

What cuts against it

List prices for a handful of parts; real transaction prices for H100-class GPUs are not public; FP16 peak throughput is not delivered throughput. The fit covers only accelerators with a public list price, about 22% a year, slower than trends fitted on wider hardware sets; list prices overstate what hyperscalers pay.

How it is measured
  • derived metric perf_per_dollar_doubling_days (formula in the semantic layer)
  • source Epoch AI ML hardware · default tier 2 · CC BY 4.0
  • proxy price, model_release · cadence annual · valve invention to product

Normal = doubling no faster than every 500 days (roughly the +49% a year Epoch reports); fast = yearly or faster. Between is emerging.

Applied to metric:perf_per_dollar_doubling_days.

Derived rows (1)
as ofdimsvalueinputs
2026-08-02-7.9 21 0.3611331 dayssources123+39 more
Latest readings
  • 2026-08-02 · 1331 daysas of 2026-08-02(42 obs)
Why this status, and its history

Evaluator: 1282 (Epoch AI ML hardware, 2025-11-06) is inside the consistent band (lo=500 ). Auto-reason; band rationale: Normal = doubling no faster than every 500 days (roughly the +49% a year Epoch reports); fast = yearly or faster. Between is emerging.

  1. 2026-09-10unmeasured to consistent with normalconf — → 40 · the evaluator (automatic)

    Evaluator: 1282 (Epoch AI ML hardware, 2025-11-06) is inside the consistent band (lo=500 ). Auto-reason; band rationale: Normal = doubling no faster than every 500 days (roughly the +49% a year Epoch reports); fast = yearly or faster. Between is emerging.

    sources123456+14 more
Why this confidence

Rests on 42 observations from Epoch AI ML hardware; best evidence tier 2, grade B; current.

Confidence is independent of status: 90–95 multiple strong independent sources; 70–89 good evidence, some ambiguity; 50–69 mixed or hard to operationalise; 0–49 limited or vague evidence.

Predictions that lean on this