[QUESTION]通用GPU TEE开源计划

[triplewz]

您好，我在隐语平台看到通用GPU TEE的技术HyperGPU的介绍，很感兴趣，看到材料中提到预计2025 S2开源，但目前并未看到开源项目，请问HyperGPU开源计划是否有调整以及调整后的时间节点是怎样的？希望得到解答，感谢！

[Bonjourz]

Hi @triplewz , sorry for the delayed reply to your question about the HyperGPU open source timeline.

Regarding your inquiry about the open source plan:

Current Status: At this point, we don't have a confirmed date for when HyperGPU will be open sourced. The original timeline has been adjusted, and we're still working on finalizing the open source roadmap.

We'd love to learn more about your interest in HyperGPU:

• What specific use case are you considering for HyperGPU? Are you looking at confidential AI inference, privacy-preserving machine learning, or another application?
• What's your primary goal? Are you evaluating it for research purposes, product development, or something else?
  Your feedback and use cases are valuable to us as we shape the HyperGPU project and determine the open source timeline. Please feel free to share more details about your needs.

Thank you for your interest in HyperGPU!

[triplewz]

Hi, @Bonjourz,

1.HyperGPU will be mainly used for confidential AI inference and other data privacy protection scenarios on cloud GPU nodes.
2.Our goals are to use it for research and product development. Since many of our current AI inference services still rely on regular GPUs(non-cc GPUs), privacy protection for general-purpose GPUs is particularly important for us.
We would be highly interested if the HyperGPU project could be open-sourced.

Thanks!

[Bonjourz]

Hi @triplewz ,

Thanks for you interest in HyperGPU.

To guide the evolution of HyperGPU, we genuinely want to understand how users are deploying AI inference in production.

If you could share:

• Model sizes you're serving (7B, 10B+ or 100B+ LLMs, Stable Diffusion, CV models, etc.)
• Expected inference throughput (TPS, QPS, daily volume)
• GPU types (e.g., NVIDIA L20, A10, AMD MI series)

it would be incredibly valuable for us. Any additional context about your use case is welcome.

[triplewz]

Hi, @Bonjourz ,

The following is the information:

• The model sizes for our inference services range from several billion to over 100 billion parameters, covering large language models, multimodal models, and other related types.

• The expected inference throughput is dozens to hundreds of QPS.

• We mainly use NVIDIA GPUs, including the A, V, L, T, and P series.

Thanks.