Didier Lasserre : So you’re talking about the bit processing?
Orin Hirschman : Yes.
Didier Lasserre : Yes. So there’s a lot of research going on right now. And they’re finding out that some efficiencies aren’t coming in like the expected 16-bit and 32-bit. As I mentioned, 5-bit was one that came up recently. And candidly, I don’t remember the exact application or the research on that. But 5 bit was one of them that was important. And so one of the benefits with this bit processing is that as our customers told us, we’re future-proof. And what I mean is if there’s an application that comes in that needs 5 bit, we’re there. If there’s one that needs 128-bit, 256 bit, we’re there. We don’t have to redesign our part to be able to address those needs because of the fact that we have 2 million bit processors on our Gemini-I. And they can work, as I mentioned, as 1 bit or make up a number between 1 million and 2 million. And so we’re future-proof from that respect, which is obviously eye opening to them.
Lee-Lean Shu : Yes. Just to comment. Right now, the GenAI, the large language model, is a very big model. It’s a very, very memory-intensive application. So they will lead to a big memory. They will lead to very fast memory. So the effort — one of the effort is they try to reduce the data format, if you have less number of bit to calculate, then you got less memory, you require less memory too. Okay. So we have big processing capability, which you can — just like Didier mentioned 5 bit, 4 bit, 3 bit, 2 bit, we will be there. And with this is we have more capability to keep up with the software innovation and that’s what some of the customer likes about this issue.
Orin Hirschman : And just going back to benchmarking, have you done benchmarking for some of the data center and hyperscaler customers?
Didier Lasserre : So candidly, most of the benchmarking we’ve done have been based essentially off of POCs, where customers have come in and said, “This is what I’m using. Here’s my data set. What do your numbers look like?” And so we’ve done real life case POCs. We’ve done that for one of our Israeli customers that we talked about and then also one of our recent SAR customers. As far as broad market benchmarking, we haven’t done a lot, but we know that’s critical and that’s something that’s on our agenda to start doing. It’s just been a resource issue for us to be able to make that happen. But we understand it’s important, and that’s what we’re next on the list.
Orin Hirschman : And I see. Two more questions, if I may, and then I’ll let other people ask. So in terms of proof of concept using customer data, have any of the hyperscalers given you data sets to actually show them what you can and can’t do with it?
Didier Lasserre : No. These are other applications. These have been SAR applications and other applications. The hyperscalers know. Because as I mentioned, the discussions early on have been revolving around this next-generation device. And so we’ve given them — we have given the benchmarking what we anticipate we could do, absolutely, but they’re not true benchmarks, they’re calculated benchmarking.
Orin Hirschman : Okay. And last question, just on the you mentioned as you’re using the Process-as-a-Service for the people’s data, and you had some early tests going back a while ago and that, that needed to be revamped or some additional APIs. Where are you up to with any of the potential customers offering it as a plug-in or anything like that?
