What quality variation should I expect across CoreML, ONNX Runtime, TensorRT for the same quantised model?

INT8 CNN classifier: 1-3% accuracy variation across CoreML/ONNX Runtime CPU/CUDA/TensorRT, occasional 5%+ on specific operators or transformers. Variation systematic per-platform based on which operators each runtime quantises aggressively vs higher-precision fallback. Transformers larger variation: attention quantisation-sensitive, KV-cache quantisation differs, embedding lookups differ. Detection/segmentation: variation manifests in IoU and confidence rather than top-line accuracy. Discipline: before multi-platform deploy, run same quantised model on each platform same test set; tabulate accuracy, latency, per-class behaviour — reveals portable vs needs per-platform calibration vs needs distillation.

Real-Time Edge Processing with GPU Acceleration

Q: When should I choose distillation over quantisation for edge inference?

Distillation when: deployment matrix wide (3+ platforms with different runtime stacks); numerical determinism across platforms required (identical or tight-tolerance output); source model too large for smallest target even quantised. Smaller architecture runs natively at full precision everywhere; QA cost is distillation training + per-platform smoke tests not per-platform numerical revalidation. Quantisation when: matrix narrow (1-2 platforms with mature toolchains); peak compute matters (INT8 TensorRT 2-4x FP16); re-training not feasible (third-party artifact, no training data, prohibitive cost). Mixed common: distil to smaller architecture, then quantise that for capable platforms.

Q: Why does INT8 quantisation behave differently on CoreML, ONNX Runtime, and WebGL — and what does that mean for QA?

INT8 is not a single spec — each runtime has own rounding, accumulator widths, calibration, operator behaviour. CoreML uses Apple's scheme; ONNX Runtime supports multiple modes (per-tensor/channel/dynamic/static) per-provider; WebGL limited INT8 varying by browser/GPU; TensorRT has fusion-aware calibration. Same INT8 model produces slightly different outputs per platform; accuracy varies 1-5% on held-out set per model/task sensitivity; edge cases near decision boundaries differ. QA implications: per-platform numerical validation required — run on each target with representative eval set, measure per platform, accept/reject per tolerance. Doubling platforms roughly doubles validation effort.

Q: How many edge platforms before distillation's portability advantage outweighs quantisation's compute savings?

Crossover usually 3-4 platforms. Below: per-platform quantisation QA manageable. Above: QA grows faster than per-platform compute savings. Depends on model quantisation sensitivity (sensitive models hit wall earlier) and runtime diversity. Per-platform quantisation costs: calibration (representative data sets quantisation parameters), numerical validation (separate per platform), performance tuning (operator support, fusion, fallbacks), regression testing on model/runtime updates. Distillation costs: training (one-time per model version) + per-platform smoke testing (lightweight, portable architecture). 1-2 platforms quantisation wins; 3-4 distillation starts winning; 5+ distillation usually right unless specific platform needs INT8 throughput.

Q: How do I evaluate model-compression options against my deployment matrix without re-validating per platform?

Define matrix: rows = platforms (CoreML, ONNX Runtime CPU/CUDA, TensorRT, WebGL/WebGPU), columns = variants (FP16, INT8 PTQ, INT8 QAT, distilled FP16, distilled INT8), cells = accuracy/latency/cost. Sequence: FP16 baseline on each platform (upper bound, identifies operator gaps); INT8 PTQ on mature platforms (easy compute savings); INT8 QAT where PTQ insufficient; distillation if matrix wide and QA cost high. Decision: best variant per platform hitting accuracy target with best latency, then consolidate if managing one-per-platform exceeds operational capacity. Cache per-platform results in deployment matrix versioned with model; targeted re-validation on changed platforms only.

Q: Where do ONNX models fit in a multi-platform pipeline — and what are the real performance-vs-portability tradeoffs?

ONNX is portable interchange format — convert from PyTorch/TF, deploy via ONNX Runtime or convert to platform-native (CoreML, TensorRT). Portability at representation level (standardised operator set, portable graph format); inference performance depends on runtime/provider. ONNX Runtime reference runtime with execution providers (CPU, CUDA, TensorRT, DirectML, CoreML, OpenVINO, ROCm) — same model runs, performance varies 2-10x per provider quality for specific model; operator support varies with fallbacks hurting performance. Tradeoff: format portability not performance portability — each platform needs validation and tuning. Most useful as pipeline interchange (training → ONNX → multiple targets), not final runtime (platform-native typically wins on performance). Production: PyTorch → ONNX canonical artifact → ONNX Runtime cross-platform consistency, convert to CoreML/TensorRT where peak matters.

Introduction

GPU-accelerated edge inference in 2026 spans phones running CoreML and Android NNAPI, edge servers with TensorRT or ONNX Runtime, browsers using WebGL/WebGPU/WebNN, and embedded NPUs from Qualcomm, Mediatek, or custom vendors. The compression decision — distillation versus quantisation — determines how cleanly the model ships across that matrix. Distillation produces a smaller architecture that runs anywhere; quantisation reduces a fixed architecture’s precision and inherits per-platform numerical variation. The right choice depends on the size of the deployment matrix and the team’s QA capacity. See GPU engineering for the broader landing this article serves.

The honest 2026 picture: single-platform deployments often prefer quantisation; multi-platform deployments often prefer distillation because per-platform QA cost grows faster than quantisation’s compute savings.

What this means in practice

Distillation produces architecture portability; quantisation produces compute savings but platform variance.
INT8 numerical behaviour differs across CoreML, ONNX Runtime, WebGL, TensorRT — same model, different outputs.
Three or more deployment platforms often flips the trade-off toward distillation.
ONNX is a portable interchange format but not a guarantee of portable performance.

When should I choose distillation over quantisation for edge inference?

Choose distillation when your deployment spans three or more platforms with incompatible runtime stacks, when outputs must match numerically across devices within tight tolerances, or when quantisation alone cannot fit the model onto your smallest target. The distilled model is a smaller architecture that runs natively on every target at full precision (or with mild quantisation if needed); the QA cost is paying for the distillation training plus per-platform smoke tests, not per-platform numerical revalidation.

Quantisation when the deployment matrix is narrow (one or two platforms with mature toolchains), when peak compute throughput on the target hardware matters (INT8 on TensorRT runs at 2-4x the FP16 throughput), or when re-training is not feasible (the model is a third-party artifact, training data is not available, or the cost of distillation training is prohibitive).

Mixed strategies are common. Distil to a smaller architecture, then quantise that architecture for the targets where quantisation is well-supported. The distilled architecture’s smaller size makes the quantisation less aggressive; the quantisation captures additional compute savings on capable targets.

Why does INT8 quantisation behave differently on CoreML, ONNX Runtime, and WebGL — and what does that mean for QA?

Vendor INT8 implementations differ in operator coverage, zero-point handling, and accumulator precision. Each runtime implements quantisation with its own rounding modes, accumulator widths, calibration methods, and operator-specific behaviour. CoreML on Apple Neural Engine uses Apple’s specific quantisation scheme; ONNX Runtime supports multiple quantisation modes (per-tensor, per-channel, dynamic, static) with execution-provider-specific implementations; WebGL has historically limited INT8 support that varies by browser and GPU; TensorRT has its own calibration pipeline with kernel-fusion-aware quantisation.

The practical consequences. The same INT8 model produces slightly different outputs on each platform. Accuracy on a held-out test set varies by 1-5% across platforms depending on model and task sensitivity. Edge cases (inputs near decision boundaries) produce different predictions on different platforms. QA strategies that validated on one platform and assumed others would match catch these gaps in production, not in test.

QA implications. Per-platform numerical validation is required: run the quantised model on each target with a representative evaluation set, measure accuracy per platform, accept or reject per platform’s tolerance. The QA cost scales with platforms; doubling the platform count roughly doubles the validation effort. Teams that did not budget for this typically ship with platform-specific accuracy gaps that surface as user complaints.

How many edge platforms before distillation’s portability advantage outweighs quantisation’s compute savings?

Three to four platforms mark the point where distillation typically outweighs per-platform quantisation efforts. Below that, per-platform quantisation QA is manageable; above that, the QA cost grows faster than the compute savings on each individual platform. The crossover depends on the model’s sensitivity to quantisation (more sensitive models hit the wall earlier) and the diversity of the runtimes (more divergent runtimes mean more per-platform issues).

The calculation. Per-platform quantisation costs include calibration (running representative data through the model to set quantisation parameters), per-platform numerical validation (separate accuracy testing per platform), per-platform performance tuning (operator support, fusion patterns, fallback handling), and per-platform regression testing when models or runtimes update. Distillation costs include distillation training (one-time per model version) and per-platform smoke testing (lightweight, since the architecture is portable).

At 1-2 platforms, quantisation’s compute savings dominate; the per-platform overhead is small enough to absorb. At 3-4 platforms, the overhead grows and distillation’s portability starts to win. At 5+ platforms, distillation is usually the right answer unless there’s a specific platform where INT8 throughput is critical.

What quality variation should I expect across CoreML, ONNX Runtime, and TensorRT for the same quantised model?

A quantised INT8 CNN classifier typically shows 1–3% accuracy variation between CoreML and ONNX Runtime’s CPU, CUDA, or TensorRT execution providers; transformer architectures and certain operators can widen this gap beyond 5%. The variation is not random — it tends to be systematic per-platform based on which operators each runtime quantises aggressively vs falls back to higher precision.

For transformer-based models the variation is larger. Attention operations are quantisation-sensitive; KV-cache quantisation differs per runtime; embedding lookups have implementation differences. Multi-platform deployment of quantised transformers requires per-platform measurement against a representative test set, not a single benchmark.

For object detection and segmentation models, the variation manifests in IoU thresholds and confidence scores rather than top-line accuracy. The metric to watch depends on the downstream use — detection systems with confidence thresholds may behave differently per platform even when raw detections are similar.

The diagnostic discipline. Before deploying multi-platform, run the same quantised model on each platform against the same test set; tabulate accuracy, latency, and per-class behaviour. The table reveals whether the model is portable as quantised, whether per-platform calibration is needed, or whether distillation is a better path.

How do I evaluate model-compression options against my deployment matrix without re-validating per platform?

List every target platform, runtime, and hardware variant your model will encounter in production. Rows: platforms (CoreML, ONNX Runtime + CPU, ONNX Runtime + CUDA, TensorRT, WebGL/WebGPU, etc.). Columns: model variants (FP16, INT8 PTQ, INT8 QAT, distilled FP16, distilled INT8). Cells: per-platform expected accuracy, latency, deployment cost.

Sequence the evaluation. First, evaluate FP16 baseline on each platform — establishes the upper bound on accuracy and identifies any platform-specific operator gaps. Second, evaluate INT8 PTQ on platforms with mature INT8 support — captures the compute savings where they’re easy. Third, evaluate INT8 QAT for platforms where PTQ accuracy is insufficient — adds training cost but recovers accuracy. Fourth, evaluate distilled variants if the matrix is wide and per-platform QA cost is high.

Decision rule. Pick the model variant per platform that hits the accuracy target with the best latency, then check whether multiple variants per platform are operationally manageable. If managing one variant per platform exceeds operational capacity, consolidate to fewer variants that work across multiple platforms — usually distilled models since they’re more portable.

Cache the per-platform validation results in a deployment matrix that’s versioned with the model. Re-validating per platform on every model update is expensive; cached results plus targeted re-validation on changed platforms is feasible.

Where do ONNX models fit in a multi-platform pipeline — and what are the real performance-vs-portability tradeoffs?

Convert PyTorch or TensorFlow models to ONNX for portability, then deploy directly via ONNX Runtime or transcode to CoreML, TensorRT, or other native formats. The portability is at the model representation level: the operator set is standardised, the graph format is portable. The actual inference performance depends on the runtime and execution provider.

ONNX Runtime is the reference runtime. It runs ONNX models with multiple execution providers (CPU, CUDA, TensorRT, DirectML, CoreML, OpenVINO, ROCm). The same ONNX model runs across providers; performance varies by 2-10x (in our experience, a directionally useful rule of thumb) depending on provider quality for the specific model. Operator support varies — some ONNX operators are not implemented on all providers, requiring fallbacks that hurt performance.

The tradeoff. ONNX provides format portability that lets teams maintain a single model artifact for multiple platforms. It does not guarantee performance portability — each platform requires validation and possibly tuning. The format is most useful as a pipeline interchange (training framework → ONNX → multiple deployment targets) rather than as a final deployment runtime (where platform-native formats like CoreML or TensorRT typically win on performance).

ONNX in production. The common pattern: train in PyTorch, export to ONNX as the canonical artifact, deploy via ONNX Runtime for cross-platform consistency and convert to platform-native (CoreML, TensorRT) for platforms where peak performance matters. The artifact-management discipline (versioning, validation per platform, conversion automation) is part of the pipeline cost.

Compression, quantisation, and deployment matrix work in concert

TechnoLynx delivers production-grade edge ML compression: distillation pipelines, per-platform quantisation tuning and validation, multi-runtime deployment spanning CoreML/ONNX/TensorRT/WebNN, and QA automation that surfaces numerical drift before release. If your team is shipping ML models across edge platforms, contact us.

Image credits: Freepik