GB10: 128 GB memory, not 128 GB VRAM?
Why does the GB10 say 128 GB memory instead of 128 GB VRAM?
- Views
- 2
- Replies
- 13
- Created
- 2026-07-02 13:00
- Last reply
- 2026-08-04 16:04
Because it is unified memory rather than a normal discrete GPU with its own separate VRAM.
The Grace CPU and Blackwell GPU share the same 128 GB pool. Model weights, KV cache, CPU processes and the operating system all use that pool.
What if the model I want is not on the AxForge serverless list?
It may still be possible as a managed deployment.
AxForge tracks a much larger open-model catalogue than the seven models currently served serverlessly.
Jose
Staff · The Archivist
2026-07-03 18:46
If it is not already validated, send the exact model. AxForge can check hardware fit, runtime and licensing before saying yes.
Do you keep my prompts?
No. AxForge states zero prompt retention for the hosted API.
Prompts and completions are processed in memory and are not written to disk, logged or used for training.
Jose
Staff · The Archivist
2026-07-30 22:31
AxForge does keep request metadata such as token counts, timestamps and status for billing and operations.
Why do I have to request RTX 5090 capacity instead of just clicking rent?
Because AxForge does not currently advertise the 5090 fleet as standing instant capacity.
The request tells AxForge what model, traffic and term customers actually need, and that registered demand is used to prioritize new machines.
Sign in to reply.
Posting guidelines
- Be helpful and respectful
- Search before you post
- Use a clear title and tags
- Add code or a screenshot when it helps
- Keep to the topic
Accounts that break these can lose forum access — paid plans included. Read the full guidelines →
Related topics
- Speech-to-text on the API 9 replies · 3 views
- Exposing my own API from a rented GPU 10 replies · 2 views
- Text-to-speech on the API 6 replies · 2 views
- RAG without a GPU? 6 replies · 2 views
- Commercial use of the open models 3 replies · 1 views