Support AMD MIxxx double-die
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 1.9k
- Forks
- 323
- Avg merge
- 1d 12h
- Merged PRs (30d)
- 12
Description
Description:
The AMD Instinct MI250 accelerator card contains two Graphics Compute Dies (GCDs) per physical card. However, when monitoring energy consumption (e.g., via rocm-smi or tools like CodeCarbon), only one GCD reports power usage, while the other shows zero values. This is problematic for accurate energy accounting, especially in HPC/SLURM environments where jobs may be allocated a single GCD.
Expected Behavior:
Both GCDs on the same MI250 card should report their individual power consumption, or the total card power should be clearly attributed to the active GCD(s).
Current Behavior:
Only one GCD provides non-zero power readings.
The second GCD always reports 0W, even when under load.
This leads to underestimated energy measurements and complicates per-job accounting.
Steps to Reproduce:
Allocate a single GCD on an MI250 card via SLURM (e.g., --gres=gpu:1).
Run a workload on the allocated GCD.
Use rocm-smi --showpower or similar tools to monitor energy.
Observe that the second GCD (on the same card) reports 0W, despite the card’s total power draw.
Impact:
Inaccurate energy tracking for jobs sharing a card.
Difficulty distinguishing per-GCD power usage.
Tools like CodeCarbon may misreport energy if they rely on per-GCD metrics.
Suggested Fix:
Provide a way to query total card power (sum of both GCDs) when monitoring a single GCD.
Alternatively, expose power readings for both GCDs, even if only one is allocated to a job.
Context:
This issue affects users in SLURM/HPC environments where fine-grained energy monitoring is critical for carbon footprint tracking and resource management.
Additional Notes:
The MI300 series may have similar behavior; clarification would be helpful.
Workarounds (e.g., manually summing GCDs) are error-prone and not scalable.
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
No source file, test, or entry point is named in the issue. Start by reproducing the MI250 behavior with rocm-smi under a single-GCD SLURM allocation, then trace how CodeCarbon reads accelerator power; done means both GCD usage or clearly attributed total card power is reported and covered by a regression test.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- hpc, observability-sre
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Quiet
- Clarity
- Needs clarification
- Newbie friendliness
- 38/100