◆ full specifications

Every number, and how we got it.

Below is the complete spec sheet for the pre-production gpuvera terminal. Figures marked target are engineering goals, not yet independently verified. When a unit is final, this page changes and we note what moved.

Last updated with v0 skeleton · numbers pending bench sign-off

gpuvera terminal three-quarter view showing the full chassis, machined top vents and rear I/O — concept render, pre-production.

Concept render — pre-production.

Compute
Inference accelerator[ SoC / NPU / GPU — TBD ]
Peak compute[ __ TOPS (INT8) / __ TFLOPS (FP16) — target ]
Sustained throughput[ __ tokens/sec on model X @ quant Y ]1
Memory & storage
Unified / model memory[ __ GB — target ]
Memory bandwidth[ __ GB/s — target ]
On-device storage[ __ GB/TB NVMe — TBD ]
Largest model class[ up to __B params @ quant — target ]2
Thermals & power
Cooling[ passive + active fan — TBD ]
Idle noise[ __ dBA @ 1m — target ]3
Power draw (idle / peak)[ __ W / __ W — target ]
Power input[ USB-C PD __ W / DC barrel — TBD ]
Connectivity
Wired network[ 1/2.5 GbE — TBD ]
Wireless[ Wi-Fi 6/6E · BT — TBD ]
Data / display ports[ USB-C ×n · USB-A · HDMI? — TBD ]
Local APIOpenAI-compatible HTTP endpoint over LAN4
Physical
Dimensions[ __ × __ × __ mm — target ]
Weight[ __ g — target ]
Chassis[ machined aluminium — TBD ]
FootprintFits in one hand · sits beside a monitor
Software & ownership
Model formats[ GGUF / open-weight — TBD ]
Accounts requiredNone — runs without a vendor account
SubscriptionNone — no per-token or recurring fee
Offline operationFull — inference runs with network unplugged

1 Throughput will be published with the exact model, quantization, context length, and batch size it was measured at. No single "hero" number without its test conditions.
2 "Largest model class" depends on quantization; the specs page will list concrete model + quant combinations that fit in memory, not a marketing ceiling.
3 Noise measured at 1 metre in a quiet room under sustained load; method disclosed alongside the figure.
4 Local endpoint is served on your LAN only by default; exposing it beyond the LAN is your choice, not the default.
Status: This is a v0 skeleton. All bracketed values are placeholders pending bench verification. When finalized, changed figures will be annotated.