In an exclusive interview with EE Times, Raja Koduri said that his new venture, Mihira AI, will create heterogeneous data center architectures to support clients’ AI and graphics workloads. One of the company’s key components will be a data center workload orchestration software layer. In addition, Mihira is setting up a content creation studio, which will supply the business with its initial workload for testing and feedback on its software and hardware as it advances. India will serve as the home base for both Mihira’s content creation studio and its first commercial data center.
Former executive vice president and chief architect Koduri led Intel’s Accelerated Computing and Graphics unit for five years before leaving the firm in March to launch his own venture. There, he oversaw the development of the Arc consumer graphics GPUs, the data center GPU lines, and the Intel Xe GPU architecture.According to Koduri, EE Times, he has three service tiers planned for Mihira.
The lowest layer will consist of a heterogeneous data center architecture designed to support workloads in three primary categories: GPU gaming workloads, heterogeneous accelerators for AI, and general CPU computing for rendering. According to him, a fourth cluster in the future might be tailored for low-power AI inference.
A board member of Tenstorrent, an AI chip and IP business, Koduri also intends to add a few Tenstorrent machines to the cluster, though the team is still working out the precise proportions of the various hardware types, he added.
Although Mihira has established a modest development cluster in Silicon Valley, the company plans to build its first data center in India.
“We are actively searching for a location,” Koduri declared. “Incentives for data centers are currently being offered in India, including very low pricing per kilowatt-hour [power supply].”
The company’s secret weapon, data center orchestration software, which arranges customer workloads across various compute types appropriately, will sit above the hardware. Software from Intel’s Project Endgame, a unified gaming services layer for cloud, edge, and home computers, has been licensed by Mihira. Intel has since discontinued development on this project. According to Koduri, the orchestration layer of Mihira will be built using software licensed from Intel, with the company developing its own intellectual property on top. For Mihira’s development setup, Intel has also donated some Project Endgame development hardware.
“There are some really interesting things we built there [at Intel], so it’s a good starting point,” he remarked. Under certain restrictions, real-time scheduling was required, which offers us some intriguing benefits for servicing AI models.
Customers will have two ways to access Mihira’s data centers.
“We will have clients who come and sit on top of our software infrastructure and only schedule work through that,” Koduri stated. “We will have clients for our raw compute.” “Many of our customers are unable to use the public cloud because they require dedicated compute and data infrastructure.”
Although hardware vendor stacks like CUDA and ROCm won’t be replaced by Mihira’s software stack at first, Mihira is actively looking at more dynamic solutions for these stacks in case they are needed in the future.
He stated, “If we can run some of the workloads on non-CUDA hardware, our TCO will be highly advantageous.” Python code is delivered to us in a container, and although we are aware that it can function if we enable AMD’s ROCm path or Intel’s PyTorch extension, doing so at the moment requires someone to modify the stack. These will be dynamically handled by our orchestration layer. Our intention is that the user not have to deal with it.
The pinnacle of Mihira’s offering is an Indian content production studio that will work on graphical projects like digital twins in addition to producing digital content for a range of industries, including entertainment.
According to Koduri, “there is an internal team whose workload includes rendering, gaming, and increasing amounts of AI, rather than building infrastructure and waiting for feedback.” “That puts me in a unique position to understand a variety of workloads at a deep level—while I start making some money.”
According to Koduri, AI will help people by closing the skills gap and enabling content producers and artists to leverage their existing skills. According to Koduri, content makers in isolated regions of India might lack the knowledge and expertise with resources that are readily available in other areas of the world, partly due to a lack of computing power.
“They won’t be limited by compute because of the Mihira infrastructure, which will provide them with access to cutting-edge ray-tracing GPUs,” he stated. They will also have access to AI models. There is a wealth of ability that can be tapped into if everyone has access to computers and equipment.
Approximately 150 of the 170 Indian workers at Mihira’s studio currently work remotely, frequently from tiny towns and isolated areas.
Will Mihira someday produce its own silicon, considering Koduri’s extensive background as a GPU architect?
“Is there a different architecture underneath that, in the long run, gains efficiency and makes the whole thing much easier to schedule?” he asked. “My brain’s silicon side is busily at work. I’m in a small group of three architects: one for GPUs, one for compilers, and one for AI models. We’re huddled in a corner brainstorming ideas. Although we won’t be a chip manufacturer, we might develop some [hardware] intellectual property that we could sell under license.
Mihira is concentrating on utilizing the AI software provided by suppliers in the interim to ascertain why it is so difficult to get AI models operating on different kinds of hardware.
“Our goal was to adopt a top-down strategy: let’s thoroughly comprehend the situation and identify Habana’s or AMD’s last mile problems,” he stated. “We hear a great deal of disparate information from various sources. Forget about outperforming Nvidia in terms of performance; that’s crucial, but it appears like many businesses are still attempting to implement AI.
By year’s end, Koduri hopes to have the first iterations of Mihira’s hardware and software operational, and in the first quarter of the following year, he intends to start collaborating with outside clients.
