Intel’s has been leaked Next generation Gaudi2 hectoliter 2080 powerful AI processor
We have one thing terribly attention-grabbing for our readers nowadays. We have a tendency to recently get our hands on some internal documents that ensure Intel is getting ready to launch a next-generation AI coaching accelerator with monstrous capabilities later this year. Dubbed the Habana Antonio Gaudi two platform, it’ll take at NVIDIA’s new datacenter merchandise (launching later this week) for deep learning and aims to supply the next worth (from a TCO perspective) different. Confine mind the documents were dated earlier this year, and therefore the launch window *could* have captive throughout this point therefore take that with a grain of salt.
Intel gets able to ‘play offense’ with Habana Antonio Gaudi two deep learning processor, launching in 2022
Intel’s Habana offerings area unit split into 2 categories: AI coaching and AI logical thinks. The Antonio Gaudi platform is that the AI coaching platform and Francisco Goya is that the AI logical thinking platform. The Gaudi two chips can succeed the older Gaudi chips offered by Intel. Whereas the documentation we have a tendency to non-inheritable solely talked a couple of Antonio Gaudi two chip, [caution: editorial speculation] it’s doable that there’s a Francisco Goya two give the works moreover [/editorial speculation] – since AI coaching and logical thinking is typically a band giving by most vendors. Vendors will utilize Intel’s Antonio Gaudi platform victimization the Amazon AWS EC2DL1 instance, wherever it already offers four-hundredth higher performance per greenback than NVIDIA-based instances.
We even have the first-ever image of the Intel hectoliter two080 AI chip supported the Antonio Gaudi 2 platform:
While we have a tendency to do receive any extra info on the Antonio Gaudi two hectoliter 2080 chip, we are able to see that it as half-dozen HBM dies. If these area unit of the HBM2 kind, this can be a complete of forty eight GB memory aboard. If it’s the HBM3 kind, then the sky is that the limit and it might be ninety six GB (or more). We are able to conjointly see the bonding pads for a complete of twenty four VRMs for what seems to be a 12+12 section style. this can be roughly five hundredth additional power than the older Habana chip that had a 8+8 style for a complete of sixteen VRMs. forward these area unit customary HBM sized chips, it conjointly feels like Intel has accumulated the die size by five hundredth over the last generation of Antonio Gaudi processors. I’m not about to do a die-size estimation here and can leave that to my colleagues to inevitably do.
To get an honest plan of the rough electronic transistor count increase, however, we have a tendency to conjointly get to grasp the method the chip is constructed on. we all know that the Intel Antonio Gaudi two hectoliter 2080 chip is constructed on ‘a’ 7nm method supported this interview by Eitan Medina, COO of Habana Labs. sadly, this does not very facilitate United States of America abundant as a result of 7nm might be pertaining to the N7 method on TSMC, Intel seven (formerly Intel 10nm), or Intel four (formerly Intel 7nm and therefore the least likely). The initial Habana Antonio Gaudi methodors were engineered on the 16nm TSMC process that makes it additional doubtless for this chip to air N7 or Intel seven. regardless of the case is, considering the Antonio Gaudi two platform is clearly on a way smaller node than 16nm (which in itself provides a density increase of roughly 50%), and combined with the die size increase, we have a tendency to area unit watching AN absolute beast of a processor that ought to simply go toe to toe with NVIDIA’s forthcoming Hopper datacenter GPU – in terms of performance per greenback – and perhaps even in terms of absolute performance.
Since being non-inheritable by Intel, Habana Labs would have had a great deal additional resources to play with and it’s just like the company is preparing to induce serious regarding DL/ML applications. Intel finally encompasses a terribly real product roadmap that would ward off – and even take market share from NVIDIA
