Skip to content

Conversation

@Yifei-yang7
Copy link

This PR adds Sirius, a GPU-native SQL engine which can provide drop-in acceleration for DuckDB.

The structure and files are similar to DuckDB, and the changes are mainly results and benchmarking script. The script is tested on the lambda-GH200 instance.

Please let me know if you need any adjustments or additional artifacts. Thanks!

@CLAassistant
Copy link

CLAassistant commented Dec 4, 2025

CLA assistant check
All committers have signed the CLA.

@rschu1ze
Copy link
Member

rschu1ze commented Dec 5, 2025

@Yifei-yang7 Thanks - looking forward to merging this.

Before doing so, I had a question. Basically all ClickBench submissions are tested on AWS EC2 instances. This submission uses lambda.ai. Running Sirius on EC2 would make it easier to compare against other databases (also, my EC2 account is covered by my company whereas lambda.ai asks me for my credit card).

Would you consider submitting results for a EC2 instance? They provide various P* boxes with Nvidia GPUs. Perhaps the compilation script needs some changes as well.

@rschu1ze rschu1ze self-assigned this Dec 5, 2025
@Yifei-yang7
Copy link
Author

@rschu1ze Thanks for your quick response!
The issue with AWS EC2 is that they don’t offer GH200 instances, which is why we’re using Lambda instead. Using other GPU instances (e.g., H100 without Grace CPU) results in worse performance — something we observed when testing via Lambda.

@rschu1ze
Copy link
Member

rschu1ze commented Dec 5, 2025

@Yifei-yang7 For sure, however long/mid-term I think that AWS will offer GH200 instances as well. What about this: you test that script runs on EC2 as well (H100) and we submit both entries - this will be more future-proof going forward.

@Yifei-yang7
Copy link
Author

Yifei-yang7 commented Dec 5, 2025

@rschu1ze I just tried to launch an H100 instance on EC2 but kept getting "insufficient capacity". It seems to be much harder to reserve an instance than Lambda.
In this case, would it be possible that you start verifying on Lambda first, and I can see if I can get an EC2 H100 instance at the same time?

@Yifei-yang7
Copy link
Author

@rschu1ze I managed to reserve a p5.4xlarge instance on EC2 and confirmed that the benchmark script runs correctly there. I’ve added the new results to this PR. Let me know if there’s anything else you’d like me to update. Thanks.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants