PREDICTED
Analytical Core output derived from workload, mapping, traffic, and timing assumptions.
CANONICAL HARDWARE CAPTURE
Analytical ML performance modeling,
tested against hardware.
TensorForge predicts accelerator performance, measures the same workload with PyTorch CUDA, and keeps the disagreement visible.
ANALYTICAL · PREDICTED—
MEASURED · P50—
ABSOLUTE PERCENTAGE ERROR—
Five labels. Five different claims.
MEASURED
Physical PyTorch latency, throughput, and CUDA memory from this device.
CALIBRATED
Empirical compute and memory-copy ceilings measured by dedicated probes.
DIAGNOSTIC
NVML context from a separate telemetry execution window.
POLICY
Decisions that exist only when their required evidence exists.
The model is useful because its boundary is measurable.
PRIMARY
—gemm_large_square
Loading primary comparison.
On this capture, TensorForge’s calibrated analytical lower bound tracked the large square GEMM closely. This is one workload on one GPU—not a claim about all GEMMs or devices.
CONTRAST
—gemm_tiny
Loading contrast comparison.
The tiny workload diverges sharply. This is consistent with fixed framework or kernel-launch overhead absent from the lower-bound model, but this experiment does not establish that explanation as causal.
Measured ceilings, not marketing peaks.
—
—
From workload description to lower bound.
WORKLOAD
↓FLOPs + TRAFFIC
↓MAPPING + TILING
↓CALIBRATION CEILINGS
↓MAX(COMPUTE, MEMORY)
↓PREDICTED LATENCY
ANALYTICAL CLASSIFICATION Predicted bottleneck: —. NVML does not prove this classification.
What the GPU reported in a separate run.
Telemetry was collected in a separate execution phase from latency measurement. It supplies diagnostic context and never determines a regression or performance verdict.
NVML reported SwPowerCap during parts of the telemetry window. This is observed context, not proof that the benchmark was power-limited.
No evidence, no verdict.
—
—
—
Ops wraps Core. Core remains independent.
TENSORFORGE CORE
Workloads
GEMM / Conv2D / Transformer
↓GEMM / Conv2D / Transformer
Mapping / PE array / SRAM
↓Tiling + residency schedules
↓Timing + bounded exploration
↓Deterministic analytical result
tensorforge_ops→tensorforgeDEPENDENCY DIRECTION
TENSORFORGE OPS
PyTorch benchmark
Calibration + validation
Regression gate
NVML telemetry
Right-sizing
Impact report
Core imports none of PyTorch, MLflow, NVML, or TensorForge Ops. The dependency arrow never reverses.
A captured experiment with a verifiable source.
What this capture does not establish.
- One physical GPU and one capture session
- Laptop thermal and power state affect measurements
- Analytical lower bound—not cycle accurate
- Not a GPU simulator
- No measured-bottleneck claim
- No generalization to every RTX 3050 laptop
- No production capacity claim
- No model-quality assessment