Gains Calculator

Frame the opportunity before the Audit.

Use the Calculator to explore directional Compute Yield assumptions for your fleet. Results are planning estimates, not measured production gains.

Definitions

Related measures, but not interchangeable.

Throughput improvement, energy reduction, wall-clock reduction and structural-waste recovery are distinct metrics. Results from separate configurations are never combined into one claim.

Quality

Quality-equivalent

Same model and weights, agreed quality standard, service level and acceptance criteria.

Waste

Structural waste

Work that does not contribute proportionately to accepted output.

Yield

Compute Yield

Quality-equivalent accepted output per GPU-hour, joule and infrastructure dollar, measured after controller overhead.

A screening estimate is not a measured result. Published measurements identify the relevant hardware, model, workload, baseline and configuration.

— Calculator interpretation standard

Inference waste calculator

Simulation / planning model

Estimate directional inference capacity before buying more hardware.

Adjust the assumptions to model the possible magnitude of recoverable capacity across an existing inference fleet. Results are planning estimates, not guaranteed savings.

Economic and Technical use different assumptions. Economic estimates planning value; Technical estimates workload fit and proof posture. Compare directionally, not one-to-one.

Modeled annual opportunity $2.34M

Capacity value $2.23M + energy $102.5K

Planning estimate only. Actual results require workload-specific testing. Token-cost savings are shown separately to help avoid double-counting.

Recovered GPUs100

Capacity effect1.10×Based on selected assumptions

Cost / 1M tokens$9.38

Fleet profile

Adjust the inputs below to model one planning scenario at a time.

Editable

Total inference GPUscurrent deployed fleetGPUs Value per GPU-hourrevenue or avoided cost$/hr Average utilizationtypical production load% Recoverable waste scenarioplanning assumption%

Scenario presets

Optional economics

Power per GPUfacility-inclusive estimatekW Electricity rateplanning input$/kWh GPU capex per unitall-in procurement estimate$/GPU Baseline cost / 1M tokenstoken-cost lens$/M Annual token volumemillions of tokensM tokens

Modeled recovered GPU-equivalent capacity 100 GPUs

Capacity-equivalent estimate: approximately 1,100 GPUs at this recovery scenario.

Modeled

Recovered GPU-hours / year 744,600

Modeled GPU-hours available for additional throughput or capacity deferral.

Modeled

Recovered capacity value $2.23M

Annual value using the selected GPU-hour economics.

Planning value

Illustrative capex deferral $4.00M

Illustrative procurement deferral if recovered capacity substitutes for incremental demand.

Illustrative

Modeled energy opportunity $102.5K

1,139 MWh-equivalent annual opportunity if recovery reduces power proportionally.

Modeled

Token-cost lens $104.2K

Token-cost lens: this scenario implies a cost-equivalent shift from $10.42 to $9.38 per 1M tokens before overhead validation.

Lens only

This calculator is for planning purposes. Actual results depend on workload-specific baseline-versus-controlled testing, quality gates, overhead accounting, and hardware-counter review. For budgeting, use one primary economic lens at a time — capacity, capex, energy, or token cost — to avoid double-counting.

Technical calculator

No finance inputs

Estimate technical inference waste without financial assumptions.

This technical calculator uses the newer proof-stance model for workload screening. It intentionally avoids economic outputs so it does not conflict with the original economic calculator.

Base screening 1.33×

Modeled recoverable-work rate 32.0% × conversion efficiency 78.0% = Realized recovery rate 25.0%

Credible public screening estimate for candidate inference-waste workloads. This is not proof and should be followed by a read-only audit.

Target stance1.30×–1.50×Public screening range

Theoretical upper1.47×Before conversion haircut

Realized recovery rate25.0%After conversion efficiency

1.30×–1.50×Base screening

Credible public-facing calculator range.

1.50×–1.80×Aggressive state-bound

High-waste workloads after validation.

1.80×Counter-validation target

Counter-certified milestone, not broad proof.

Technical inputs

Select a proof stance and adjust workload screening signals.

Technical

Proof stance

Workload type State and KV/cache movement usually dominate long-context workloads. GPU fleet sizeGPUs Fleet-wide baseline throughputtok/s Baseline p95 latencyms Board power per GPUW KV / state memory footprintGB

Waste and conversion signals

Matrix compute waste signal% Memory traffic waste signal% State redundancy signal% Conversion efficiencydetected recovery that becomes usable throughput% Quality retention floor% Controller overhead%

Modeled screening multiplier1.33×

100,000 tok/s baseline → 133,332 tok/s modeled after conversion haircut.

Public screening range

Theoretical upper bound1.47×

Before scheduler/fleet conversion. Not a production claim.

Upper bound

Throughput-equivalent capacity333 GPUs

Derived from the technical multiplier, not from economic assumptions.

Technical

Modeled p95 latency reduction16.2%

850 ms baseline → 712 ms modeled.

Requires p95/p99

KV / state reduction opportunity25.4%

40.0 GB baseline → 29.9 GB modeled opportunity.

Counter required

Energy per 1K tokens improvement34.4%

6,500.00 J/1K tok baseline → 4,265.21 J/1K tok modeled.

Telemetry

Technical signal dashboard

Separates detected waste from the Modeled recoverable-work rate and Realized recovery rate.

Composite waste index37.0%

Modeled recoverable-work rate32.0%

Realized recovery rate25.0%

Technical screening score86.0%

Matrix compute waste signal35.0%

Memory traffic waste signal45.0%

State redundancy signal50.0%

Conversion efficiency78.0%

Technical guardrails

The technical calculator is designed to keep the 1.80× milestone separate from broad production claims.

Methodology

Useful multiplier 1 / (1 - Realized recovery rate)

A multiplier estimates useful output from avoided work after quality and overhead assumptions.

Technical recovery Modeled recoverable-work rate × conversion efficiency

The technical lens applies a conversion haircut before turning recoverable work into useful throughput.

Energy per output fleet power / fleet-wide throughput

Energy per 1K tokens uses total fleet power and fleet-wide throughput, not only per-GPU power.

Lens separation Economic ≠ Technical one-to-one

Economic estimates planning value. Technical estimates workload fit and proof posture. Compare directionally.

Screening confidence

weighted blend of quality retention, low overhead, conversion efficiency, and waste-signal strength, with a +20 baseline reflecting screening-not-proof scope.

This is a directional confidence score for a screening tool. It is not a measurement of expected production accuracy. Counter-validated benchmarks would replace this score before any external claim.

Next step

See Waveform against your own baseline.

Share your infrastructure environment, selected workload and operating objective. Vectris will define the evidence required for a customer-specific Compute Yield case.

Request an Audit →