Thermal Benchmark Evidence

Recorded GPU thermal telemetry.

Validated performance

Thermal stability expressed as auditable signals.

0 μs

Throttle events recorded in the CARDIAC-PURR benchmark condition.

2.36×

Clock advantage versus the unmanaged baseline condition.

200h+

Validated runtime across two GPU architectures.

0.43%

Throughput variance reported in the benchmark evidence.

Validated telemetry / recorded benchmark run

Without CARDIAC-PURR vs With CARDIAC-PURR.

A deterministic comparison instrument built from source-supported benchmark values. Long-duration validation and controlled A/B data remain separate, so qualitative baseline states are not converted into unsupported numbers. This is recorded benchmark evidence, not a live GPU connection; interpolation is used only to render the comparison presentation.

Long-duration validation: with CARDIAC-PURR shows 200+ hours, 0 μs throttle events, 2.36× clock advantage, and sustained 1,635–1,650 MHz operation. The without condition is source-supported as degraded/throttled, not numerically reconstructed.

Without CARDIAC-PURRBaseline
GPU clockDegraded
TemperatureExceeds TDP
PowerUncontrolled
Throttle statusContinuous
Throughput var.High
Compute epochsReduced
With CARDIAC-PURRTreatment
GPU clock1,635–1,650 MHz
Temperature~82°C
Power~72W
Throttle status0 μs
Throughput var.0.43%
Compute epochs1,924,719,180
Clock trace / recorded comparisonWith: 1,635–1,650 MHz · Without: Degraded
Recorded benchmark telemetry comparison Deterministic SVG comparison of source-supported CARDIAC-PURR benchmark values for clock, temperature, and power. recorded benchmark progression clock stability

Long-duration clock mode selected. With CARDIAC-PURR: 1,635–1,650 MHz. Without CARDIAC-PURR: degraded. No unsupported numeric baseline is shown.

MD5 verified UUID traceable AWS CloudTrail corroborated 200+ hours validated

Controlled A/B validation

A/B tested. Telemetry-proven.

Controlled 12+12-hour head-to-head on live cloud GPU infrastructure. Same instance, same workload. Dual telemetry capture - NVML baseline plus DCGM treatment.

MetricBaseline - 12h unmanagedTreatment - 12h with CARDIAC-PURR
SM clock937 MHz avg (throttled)1,153 MHz avg
ThrottleSwPowerCap - continuousSwPowerCap - same, but escaped
Temperature81.3°C avg80.9°C avg (cooler)
Power72.0W avg74.5W avg (+3.5%)
StatusPermanently cappedClocks recovered
+23.1%

SM clock recovery

937 → 1,153 MHz

−0.4°C

Temperature delta

Cooler at higher clocks

+18.9%

Efficiency gain

MHz per watt improved

+4.6%

Effective throughput

L4 Ada validated

Methodology / provenance

Validation records designed for technical diligence.

Independently corroborated CARDIAC-PURR evidence across 200+ hours of validated GPU runtime, with sustained clocks, 0 μs throttle events, controlled A/B telemetry, and integrity records designed for technical diligence.

Baseline

12 hours unmanaged. NVML via nvidia-smi.

Treatment

12 hours with CARDIAC-PURR. DCGM field telemetry.

Environment

AWS EC2 g6, NVIDIA L4 Ada, eu-central-1c.

Integrity

MD5 integrity verification for telemetry records.

Traceability

UUID traceability to exact capture moments, including UUID GPU-d54fe77d-85b5-cece-5f45-58fe0444c597 where stated in the A/B source.

Corroboration

AWS CloudTrail corroboration as third-party audit context. The A/B source identifies baseline PID 10355.

Supporting evidence imagery

Thermal evidence context.

Thermal benchmark comparison visual for CARDIAC-PURR
Thermal performance — before vs. after
CARDIAC-PURR thermal dispatch and GPU telemetry visual
Thermal control — performance visualisation

CARDIAC-PURR context

Software-only control above the device driver.

CARDIAC-PURR is designed for sustained accelerator workloads where hardware thermal protection can collapse clock speed once the chip runs too hot. The benchmark evidence presents a software-layer prevention approach without clock, voltage, or firmware changes.

Read CARDIAC-PURR overview

FAQ

Benchmark interpretation.

Source-supported answers for reading the recorded CARDIAC-PURR thermal evidence without mixing the long-duration validation and controlled A/B datasets.

The 200+ hour validation is the long-duration evidence context. In the historical source, combined validated runtime exceeds 200+ hours across two GPU architectures, with telemetry points MD5-verified and UUID-traceable to their moment of capture. That same summary reports throughput variance as 0.43% for the CARDIAC-PURR condition.

0 μs refers to throttle events recorded in the CARDIAC-PURR benchmark condition for the long-duration validation. It should be read in that evidence context, not as a universal claim for every GPU or workload.

The 2.36× figure compares the CARDIAC-PURR condition against the unmanaged, throttled baseline condition in the long-duration benchmark summary. The same source reports sustained operation at 1,635–1,650 MHz with CARDIAC-PURR.

They are separate evidence contexts. The 200+ hour validation summarizes longer combined runtime across two GPU architectures, while the controlled A/B test is a 12-hour unmanaged baseline and a 12-hour CARDIAC-PURR treatment on the same cloud instance and workload.

The controlled A/B validation was conducted on an AWS EC2 g6 instance with an NVIDIA L4 Ada GPU in eu-central-1c. The source records GPU UUID GPU-d54fe77d-85b5-cece-5f45-58fe0444c597 and baseline PID 10355.

The source identifies dual telemetry capture: the unmanaged baseline used NVML via nvidia-smi, while the CARDIAC-PURR treatment used DCGM field telemetry. The page keeps those capture contexts distinct.

The source states that every telemetry point is MD5-verified for integrity and UUID-traceable to its exact moment of capture. AWS CloudTrail provides independent corroboration through a third-party audit log.

No. The telemetry module presents recorded and validated benchmark evidence. Its deterministic SVG rendering and interpolation are used only to present the comparison; they do not add raw measured samples or connect to a currently running GPU.

This page is the CARDIAC-PURR GPU thermal benchmark page. The Platform benchmark page covers the AI Control Plane family, so its benchmark content and numbers are intentionally separate.