-
Notifications
You must be signed in to change notification settings - Fork 35
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Support for Intel's TDX, AMD's SEV-SNP and ARM's Realm CC #193
Comments
Hi Ajay, We're glad you like this project. We're working on integrating AMD SEV-SNP support for our confidential-AI solution soon. |
Hi Ajay, @ShannonSD covered most of the key points. I'll add a couple more details: You can find the BlindLlama whitepaper here: BlindLlama Whitepaper. The whitepaper is a good starting point if you want to know more (it does not talk about the plan for AMD SEV SNP support though). Also despite the name the BlindLlama approach can be used to serve not just LLM models, but any ML model. Additionally, regarding BlindLlama, our upcoming AMD SEV-SNP support will include support for Nvidia Hopper Confidential Computing mode (which can be used on the Nvidia H100). We are also interested in supporting Intel TDX once we have added support for AMD SEV SNP. We are not considering ARM Realm CC at this point. |
Thanks for your response. |
Hi @clauverjat Just for reference, as per the Medium article on CC ARM CCA support secure DMA as well it is open-source. |
Hi Ajay,
That’s a good question. Indeed, both solutions perform inference in trusted execution environments. However, they originated from different constraints and demands. That’s why we launched a new project : BlindLlama. For this project, the goal was to be able to do LLM inference and we focused on doing just that.
No, we don’t plan on consolidating the two projects.
Originally a TPM is a separate chip. Its purpose is to act as a root of trust for the system. Most modern servers come with a TPM chip. But the TPM comes in all sorts of shapes. For instance, AMD CPUs provides a firmware TPM (fTPM), which integrates the TPM functionality directly within the CPU. In virtualized environments, the TPM can be provided by the hypervisor (like the hypervisor can provide other virtual devices), in that case it is referred to as a vTPM.
First I want to point out a common misconception. You stated in your first message that “SGX is already deprecated in the 11th and 12th generations of Intel's CPUs”. This is misleading. So, Intel TDX should very much be viewed as a similar but complementary technology to SGX rather than a replacement. TDX provides isolation at the virtual machine level, whereas SGX offers application-level enclaves. Whether you want one or the other depends on the specifics of your use case. Now regarding BlindAI, we plan work depending on what our clients want. Recently BlindAI has not received a lot of interest, so we paused development. But of course if one client had a good use case for it, and wanted us to support it. We'll resume work on it. For now though we focus on BlindLlama and CVM technology support. I do expect most people to go for the CVM or TPM approach since it is more flexible. But that does not mean that there is no room for BlindAI (and Intel SGX). BlindAI will be a better choice if you have an application that demands higher security guarantees. |
Hi Darshan, Thanks for the Medium article, it is interesting.
I don't think there is anything wrong with ARM CCA if that's what you are asking. We just need to focus our efforts. Our clients have manifested interest for AMD SEV SNP and Intel TDX, but they haven't manifested interest for ARM CCA yet. I think it is likely due to the market share of ARM in the server CPU market : most server CPU are still Intel or AMD CPUs (Though usage of ARM on the server side is growing). Also (but this is related), my team has not yet tried ARM CCA. On the other hand, we've started work on AMD SEV SNP support, and we've experimented with Intel TDX. Before adding it to the roadmap, we'll need to experiment with it. That being said I really like the fact that ARM CCA is open source (this is great in terms of transparency). So we'd be interested in exploring the technology at some point |
Thank you @clauverjat for you detailed updates and plan, we will explore on BlindLLM and continue to discuss more on same. |
Hi Team,
Thanks for this work and making it open source.
I have a question regarding support for Intel's TDX, AMD's SEV-SNP, and ARM Realm CC in BlindAI. Is there any plan in place to include these new technologies since SGX is already deprecated in the 11th and 12th generations of Intel's CPUs?
The text was updated successfully, but these errors were encountered: