Hardware
Every device Alpha targets.
Each page reads the card's typed profile straight from the Alpha source, says what has run on it, and shows which parts of its profile have been measured rather than assumed. Systems shows the same devices from the models' side.
Where the numbers come from
From the typed profile, not from this page.
Driver facts are read from each profile at build 2877a0c1. Measured readings come from dated run records: model training on the RTX 3090 and the GB10, and memory bandwidth on the GB10. Everything else on a page is marked as not measured yet.
GeForce RTX 3070
NVIDIA Ampere · GA104
The smallest-memory qualified card (8 GiB): both training systems must fit here.
- 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
Alpha's performance card: Coppelius trains here 1.51× faster than PyTorch with CUDA graphs.
- 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 card: sm_86 instructions run under sm_89 launch descriptors 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 Blackwell GPU sharing 128 GB of memory with a 20-core Arm CPU. Coppelius trains here through the compiler's sm_121 back end.
- Architecture
- AArch64 host · Blackwell compute capability 12.1
- Unified memory
- 130,596,048,896 B (121.6 GiB)
- Profile tiers with readings
- 3 of 3