TechNode
Technology · Measurement code, public

See for yourself whether the numbers are real
in the code.

Every core number at TechNode is backed by real measurements, logs, and code. Not claims — *evidence*. We publish the code and ledgers so you can verify it yourself.

~1 ms
Yield latency (NtSuspend, process-internal)
5.94 ms
dispatch p99 (process-internal, n=300, network excluded)
1.26M
Preemption-yield soak cycles · 0 failures (19.6h)
Live now · Control Plane Loading…
Ready nodes
Parked workers
Total VRAM
State

Live /status API snapshot · published, real-usage data.

The architecture at a glance

Three layers · all modular · all measurable.

Control Plane (Brain)

Provider router + session broker + HTTP API. Handles worker registration, session assignment, and reassignment decisions. The current beta pool is a few personal PCs (RTX 4060 Ti / 4070 Super) in scale.

Provider Agent

A native PowerShell agent on each host PC. Input detection (Raw Input) → workload freeze (NtSuspend) → heartbeat. Average 1ms response.

Pre-Warmed Pool

Workers are spun up in advance and wait in a frozen state. On dispatch: thaw + payload injection → zero cold-start cost. Measured at n=300.

What's different — and what's still lacking

What's genuinely different by design, and what's still lacking — we write both down as they are.

Genuinely different by design

Pre-warming workers in advance, and instant yield that pauses the workload the moment the provider presses a key. That's why dispatch and yield latency are in the µs–ms range (measured process-internally).

Still lacking

Scale is a few personal PCs (at launch). If a node drops it reconnects automatically, but in-progress state is not preserved. Isolation is process-level (running on a personal PC, so it isn't fully separated from the host — no sensitive data). High availability ("99.99%"-type numbers) has never been measured, so we don't claim it.

In the measurement methodology below we've written out the actual ledger paths, n counts, and distributions exactly as they are. Verify it yourself.

How far does isolation go

Current isolation is process-level. The workload runs as a separate process, but because it runs on a personal PC, it isn't fully separated from the host (the PC's owner).

Confidential computing (CVM + GPU Confidential Compute), which would also block host access, requires H100-class datacenter hardware and is not supported on consumer GPUs. So do not upload sensitive data, personal information, or proprietary models. It's well suited to public and experimental workloads.

The transport is encrypted with TLS. Strong isolation such as microVM and CVM is on the roadmap for the formal stage, when datacenter-grade nodes come online.

Measurement methodology

We answer "how did you measure this?" with *code*.

bench_failover_soak.py
24-hour failover soak Repeats failover at 50ms intervals. Appends to the ledger line by line at .technode_runtime/soak/failover_soak.jsonl. 19.62 hours / 1,258,509 cycles · 0 failures · 5 SLA breaches.
prewarm_bench.py
Pre-warm dispatch measurement Measured on a PC2 RTX 4060 Ti, all the way through NtSuspend + stdin payload + matmul. n=300. p50 430µs · p95 2.71ms · p99 5.94ms (process-internal, network not included).
bench_runpod_setup.py · bench_vast_setup.py
Competitor cold-start measurement Boots RTX 4090 containers via the official RunPod and Vast APIs. RunPod p50 28.5s / p95 29.8s (8/8). Vast about 70s / ready in 2 of 12 runs. Ledger retained.
confidential_compute.py
Isolation-stage definitions (single source) Defines the current stage (freeze) and the higher stages (microVM · CVM · GPU CC) in one place. The higher stages activate after attestation verification once datacenter-grade hardware is introduced.

Check the code and ledgers yourself

We take due-diligence and technical-review requests by email. We'll share the raw ledgers and the measurement code.

Request technical materials →   To the Customers page