Skip to content

Early access · AWS Trainium & Inferentia2

AWS Neuron accelerators

Pantheon diagnoses NVIDIA CUDA and AMD ROCm GPUs. AWS Trainium and Inferentia2 are not GPUs, they're a separate class of accelerator with their own toolchain, so we built a separate suite for them: Pantheon Neuron.

This is a status report, not a leaderboard

Pantheon Neuron is at v0.1.0. Of 25 planned workloads, 2 have a working kernel, and only one of those is verified end-to-end on hardware. Everything below comes from a single device of each type, not repeated runs across a fleet the way the GPU benchmark database is. Read it as where the suite stands today, not a comparison.

Validation status

Chip Device Status
Inferentia2 (inf2) NeuronCore-v2, 2 cores/device Verified, inf2.xlarge, 2026-08-26
Trainium1 (trn1) NeuronCore-v2, 2 cores/device Verified, trn1.2xlarge, 2026-08-27
Trainium1n (trn1n) NeuronCore-v2, 2 cores/device Assumed, same silicon as trn1 with more network; not run
Trainium2 (trn2) NeuronCore-v3, 8 cores/device Assumed, least confident; not run

Device generation does not track product naming: Trainium1 reports NeuronDevice v2 and Inferentia2 reports v3, but both run NeuronCore v2. The counter sets also differ between chips, so a kernel can't assume a counter that exists on one exists on the other.

Kernel status

Workload Status
baseline_metrics Telemetry only, no load
memory_read Verified on trn1.2xlarge
memory_write Written, primitives verified, this arrangement untested on hardware
Remaining 23 workloads No kernel yet

What's actually been measured

Two numbers below came from a real kernel doing real work, not a diagnostic probe:

Measurement Value Chip Source
Sustained HBM read bandwidth 264.15 GB/s Trainium1 (trn1.2xlarge) memory_read NKI kernel, 1024 MiB bf16 buffer, 20 passes, device barrier inside the timed region
Training throughput 23.25 train-steps/s Trainium1 (trn1.2xlarge) 2-layer 1024×1024 MLP, batch 16, SGD, 50 steps; confirms the training capability path works

Both are single runs on a single device, not medians over repeated samples. Read them as "this works, and roughly what it does," not as a Score to set against another chip.

What we're deliberately not showing

An early bring-up pass also ran on inf2.xlarge and read counters like effective_flops and mfu_estimated_percent from neuron-monitor. We're not publishing those numbers here: the load behind them was an untuned matmul running at roughly 0.005% of the chip's compute capability, and Pantheon Neuron's own data file flags them explicitly as proof the counters are readable, not a measurement of Inferentia2's throughput. A number your own source says not to compare isn't a benchmark, so it stays out until a real kernel runs on inf2 the way memory_read did on trn1.

Follow the build

Pantheon Neuron is developed in the open. Full methodology, kernel source, and raw probe data are in the repository.

View Pantheon Neuron on GitHub GPU benchmark database

This page grows into a full leaderboard as more workloads get verified kernels. Until then, it's an honest snapshot of where the suite stands.