3. Hardware Resource Investment Plan#
Throughout chip adaptation, the chip company must provide physical devices at each stage to support CI/CD, validation, and release. For the first two batches, two machines are recommended for each batch. KernelGen is expected to use two machines so that operators can be specialized, tested in batches, and used for operator competitions. Single-card, four-card, or eight-card configurations should be confirmed according to the scenario. Additional machines are usually required for Day 0, distribution, cloud-native, or showroom plans. Virtualization can supplement multi-OS validation, but critical validation still requires physical devices.
Stage ID |
Stage |
Required devices |
Purpose |
Configuration requirements |
Availability timing |
|---|---|---|---|---|---|
Stage 1 |
Environment readiness |
1 chip development board or server with remote access for the FlagOS team |
Single-card operation |
Ubuntu OS, 64 GB or more memory |
Prepare immediately after the Stage 0 MOU is signed |
Stage 2 |
FlagTree compiler integration |
1 device added to the CI resource pool |
Compiler build validation and CI pipeline triggers |
Single card, 200 GB or more disk space, network access to the FlagOS CI cluster |
Connect to CI when Stage 2 starts |
Stage 3 |
FlagGems operator adaptation |
1 CI machine and 1 debug machine |
Automated CI tests and operator development/debugging |
The CI machine must stay reliably online; the vendor may keep the debug machine locally |
Confirm the CI machine SLA when Stage 3 starts |
Stage 4a |
FlagCX communication adaptation |
At least 2 machines, each with at least 4 cards |
Collective communication tests and multi-node communication validation |
High-speed interconnect such as IB or RoCE; homogeneous cluster |
Available when Stage 4a starts |
Stage 4b |
FlagScale training and inference |
At least 2 machines, reused from Stage 4a |
Distributed training and inference service validation |
Same as Stage 4a, with persistent storage mounted |
Reuse Stage 4a devices |
Stage 5 |
End-to-end validation |
1 CI machine and 2 SVT machines |
Continuous CI regression and system validation testing |
SVT machines must be isolated from CI machines to avoid interference |
Available when Stage 5 starts |
Stage 6 |
FlagRelease release |
1 dedicated FR machine |
Model migration, image builds, and release validation |
Single card or above, 200 GB or more disk space, high-speed network |
Available when Stage 6 starts |
Stage 7 |
KernelGen |
2 machines |
Automatic operator generation, tuning, and operator competitions |
Single-card operation with FlagTree compiler and operator development/tuning environments installed |
After Stage 2 is complete |