Former AMD and Intel graphics executive Raja Koduri seems to be one of those creators that doesn't let go of a project they see potential in, judging from how his new company, Mihira AI, has licensed projects that Intel chose to shelve. In this case, it seems that Mihira is looking to acquire the IP left behind by the company's work on Project Endgame - a network-based solution meant to deliver additional graphics power towards local hardware at the push of a button. The report comes courtesy of EE Times.
Project Endgame was announced by Intel back in 2022, but work on the feature went silent even throughout the period where it was supposed to be made available in beta form. Yet in July this year (about four months after Raja Koduri jumped ship from Intel towards his new company, Mihira AI), Intel announced via Twitter that the project was in an "indefinite hold".
One man's trash is another man's gold, however, and it seems Raja and Mihira AI believe that the IP behind Project Endgame will be critical for the company's aspirations of becoming a software provider for the highly-lucrative data center space. According to the company, Mihira AI aims to build heterogeneous data center architectures meant to deliver graphics and AI workload acceleration.
The strategy is an integrated one: Mihira AI is aiming to open a content creation studio that's meant to be illustrative of the company's ability to orchestrate data center workloads through its software stack. The major goal here - and it really is major - is to offer an agnostic solution that competes with both Nvidia's CUDA (the green team's crown jewel), and AMD's ROCm.
Heterogeneous workloads, in this case, mean orchestrating communication and workloads among general CPU compute clusters for rendering; heterogeneous accelerators for AI and gaming GPU workloads; and perhaps yet another future, fourth cluster, to be optimized for low-power AI inference.
Through that objective, it does seem like the company could find some use for Project Endgame - the software solution was indeed designed to allow for workload orchestration by pulling in additional graphics grunt when needed. At the very least, that's one piece of Mihira AI's puzzle - especially after the company builds its own IP on top of the solution.
It's perhaps also telling that Mihira AI has decided not to delve too much into actual hardware design and fabrication, despite its founder's experience in the field within both AMD and Intel. According to Raja, it's a matter of pragmatism to "forget beating Nvidia on performance[.]" The real opportunity, Mihira AI thinks, lies in lowering and democratizing entry barriers towards cloud and network-based compute. On that perspective, it seems, canned projects have the most potential.
