Hardware
Every device gets its own page.
An agent choosing a target should know exactly what it is working with. Each page reads the device's typed profile from the Alpha source at build time, states what has been physically qualified on it, and says—tier by tier—how much of its profile has been measured rather than assumed.
Evidence boundary
Facts come from the typed profile, not from this page.
Bound to 408ac92b. Driver facts are live for every device below; the micro-benchmark tier has not been run on any card, and model-shaped calibration has a first end-to-end reading only where a page shows one. Each page says which is which.
GeForce RTX 3070
NVIDIA Ampere · GA104
The constrained SM86 qualification device: the smallest memory envelope both training systems must fit.
- Architecture
- SM86 · compute capability 8.6
- Total video memory
- 8,589,934,592 B (8 GiB)
- Profile tiers with readings
- 1 of 3
GeForce RTX 3090
NVIDIA Ampere · GA102
The large-memory SM86 device and the card the final qualification of every artifact is run on.
- Architecture
- SM86 · compute capability 8.6
- Total video memory
- 25,769,803,776 B (24 GiB)
- Profile tiers with readings
- 2 of 3
GeForce RTX 4090
NVIDIA Ada Lovelace · AD102
The Ada target: the qualified SM86 instruction payloads bound to SM89 command-stream and driver contracts.
- Architecture
- SM89 · compute capability 8.9
- Total video memory
- 25,769,803,776 B (24 GiB)
- Profile tiers with readings
- 1 of 3
DGX Spark · NVIDIA GB10
NVIDIA Grace Blackwell · GB10 (product GX10)
A measured AArch64 build and execution host whose accelerator is not yet admitted by any Alpha device lens.
- Architecture
- AArch64 host · Blackwell compute capability 12.1
- Unified memory
- 130,596,048,896 B (121.6 GiB)
- Profile tiers with readings
- 1 of 3