The hardware catalog shows figures like tok/s, TOPS, and a <200ms first-token target. Here's exactly what those mean, how they're derived today, and what we haven't measured ourselves yet — we'd rather you know the difference than assume more rigor than exists.
Every performance figure in the hardware catalog is labeled indicative for a reason.
Current tok/s and TOPS figures are estimates derived from public manufacturer specifications and community-published benchmarks (e.g. llama.cpp, Ollama, and Jetson benchmark reports) for comparable quantized model sizes on each SoC or accelerator — not Sovra's own controlled lab runs, unless stated otherwise.
Figures assume the specific quantized models listed per device (e.g. llama3.2:1b on a 1GB Raspberry Pi 5, llama3.1:70b-q4 on a GPU workstation) — a different model size or quant level will perform differently on the same hardware.
It's the design goal for a properly configured mid-tier edge device (e.g. Jetson Orin Nano class) running an appropriately sized quantized model — not a guaranteed number across the entire catalog, and not yet independently verified by us across all 36 platforms.
Being upfront about the gap between "indicative" and "measured."
We have not yet run and published a controlled, repeatable Sovra benchmark suite across the hardware catalog with fixed prompts, model versions, and hardware revisions. That's on the research roadmap (edge benchmark suite, targeted Q3 2026).
Power delivery, thermal throttling, storage speed (SD card vs NVMe), and OS configuration all move real-world tok/s — sometimes significantly — from the indicative figures shown.
| Field | What it means |
|---|---|
| TOPS | Manufacturer-published theoretical peak INT8/FP4 throughput for the accelerator — a ceiling, not a delivered tok/s number. |
| tok/s (indicative) | Estimated generation speed for the specific quantized model listed, based on public benchmarks for comparable hardware — see disclosure above. |
| Power (W) | Typical operating power draw from the manufacturer's own specification, not measured under a specific Sovra workload. |
| Price source | Linked directly on every hardware card — manufacturer MSRP or a named retailer listing, dated. |
See the full disclosure on the hardware configurator page — every catalog entry links its own price source.
Tell us your model, hardware, and use case — we'll give you a straight answer, including if we haven't tested it.
