Open the window.
Define who can connect, what data is allowed, and which system class is safe to expose.
AI hardware evaluation desk
Peel the stack.
Prove the machine.
Onion Compute runs remote evaluations where every layer stays visible: access, hardware, runtime, and evidence. Bring the workload. Leave with a rerunnable decision packet.
01 / PEELExpose the layer that limits the workload.
02 / MEASURERun clean commands on real accelerator systems.
03 / PROVEDeliver evidence that another engineer can rerun.
layer map / public viewNDA-safe by default
01Access shellWindow, credentials, controls
02Hardware coreMemory, fabric, topology
03Runtime layerCUDA, ROCm, serving, kernels
04Evidence layerLogs, metrics, packet, decision
02 / Peel the stack
One workload. Four layers of truth.
The public site shows the architecture without leaking exact hosts, credentials, or allocations. Each evaluation moves inward until the actual bottleneck is measurable.
Find the pressure.
Memory capacity, bandwidth, topology, power, and communication determine the shape of the run.
Make it repeat.
Pin versions, serving path, kernels, precision, and the commands that turn a claim into a run.
03 / Public class table
Production useful. Redacted by design.
Exact machine identity stays private. The class, runtime path, tests, and evidence standard stay visible.
ClassRuntime pathPrimary testStatus
AMD acceleratorROCm · RCCL · PyTorchLarge memory, collectives, servingEVALUATION
NVIDIA datacenterCUDA · NCCL · containersBaseline serving, tuning, kernelsEVALUATION
Workstation / serverLocal runtime pathQuantized serving, cost / throughputBENCHMARK
Compiler acceleratorXLA · graph compileSharding, compile stability, fabricTARGET
04 / Workload menu
Different workloads expose different layers.
Tests are written for engineers: metric, failure mode, runtime path, and artifact.
LLM inference
TTFT, TPOT, throughput, batching, long context, KV behavior.
Multi-GPU
Tensor parallelism, collectives, topology sensitivity, stalls.
ROCm path
Runtime setup, RCCL behavior, compatibility, tuning knobs.
Kernel work
Bandwidth, occupancy, fusion, precision, upstreamable fixes.
Fine-tuning
Adapter and full-tune checks, checkpoints, data movement.
Reproduction
Commands, versions, metrics, traces, failures, clean reruns.
05 / Deliverable
The product is the evidence packet.
Every layer ends in an artifact another engineer can inspect, rerun, and use to make the next decision.
eval_packet/
Private by default. Public only after sensitive details are removed and the workload owner approves the note.
NDA-safererunnabledecision-ready
system_class.jsonPublic-safe class and environment shape.
runbook.mdExact commands, versions, and setup.
metrics.csvLatency, throughput, memory, metadata.
failures.logNegative evidence and compatibility gaps.
report.mdResult, bottleneck, recommendation.
patches/Scripts, kernels, and fixes when available.
06 / Book evaluation
Send the workload. We return proof.
Include the model, runtime, precision, context length, concurrency target, system class, and the decision this evaluation needs to support.
Book evaluation window→Submit benchmark→Attach capacity→access@onioncompute.website · exact infrastructure stays private by default