Home / Software / Apps / Lemurian Labs is creating a cheaper compute paradigm for AI models

Lemurian Labs is creating a cheaper compute paradigm for AI models

The resource demands of generative AI models have driven demand for Nvidia GPU chips to an all-time high, but what if there was a cheaper chip with similar power? Google, Intel, and Nvidia alumni’s early-stage startup Lemurian Labs aims to build that.

It’s a moonshot idea, and it takes a lot of time and money to bring a chip idea to market, but investors are willing to take a chance on founders with a certain pedigree. The startup announced a $9 million seed investment today.

Lemurian aims to reinvent accelerated computing. We want to do that because our current computing method is ending. Jay Dawani, co-founder and CEO of Lemurian, told  that semiconductor physics are opposing that architecture and paradigm.

The company wants to create a chip and software to make AI workloads easier, faster, cheaper, and greener.

Like a computer architecture master class, Lemurian says computing is about math, memory, and movement. Connectivity is the goal. Data is stored in memories, transferred through an interconnect to a math unit, manipulated, and returned to memory. The traditional point in architecture is that data must travel, Dawani said.

Lemurian wants to reverse it. Instead of data traveling to compute resources, it wants compute to travel to data. “We need to essentially minimize that distance, so we aren’t really moving data, we’re moving around compute,” he said.

He claims that GPUs were originally designed for graphics but have since expanded into other areas due to their processing power. Because you’re designing for something and trying to do something else, and you’re not good at everything. That’s GPUs’ Achilles’ heel. But we’re trying to fix that,” Dawani said.

Lemurian plans to change the chip’s math, a massive task. In the early days of chip development, engineers chose a floating point approach because logarithmic methods failed, Dawani says. His firm claims to have resolved the issue.

Log number systems are great because they convert expensive multiplies and divides into hardware-free adds and subtractions. They save energy and space and gain speed. You also gain some exactness or precision, which is attractive when trying to reduce processing costs on large language models.

They did this how? “We discovered that by constructing and extending the definition of a large number system, you can create an exact solution, which is smaller and more accurate than floating point for the same number of bits,” he said.

For the same number of bits, dynamic range improves as bit count increases, which is fascinating. Without the number system, you’re limited, so that’s a big part of what lets us explore our architecture.”

They are releasing the software part of the stack slowly, hoping to have it generally available in Q3 next year. The hardware is harder to develop, manufacture, and test, but it should follow in the coming years.

Our 24 employees are mostly skilled technical engineers with project experience. That’s a small pool, but he wants to hire six more people in the next few months and 35 more in a year if they get a Series A.

Along with Good Growth Capital, Raptor Group, and Alumni Ventures, Oval Park Capital led the $9 million investment.

It’s difficult and expensive to build a company like this and get the chip to market, but if they can do it, it could make building generative AI models (and whatever comes next) cheaper and easier.

About Chambers

Check Also

With an eye on AI, Amazon adds Andrew Ng to its board. Former MTV executive McGrath will step down

If what corporate boards of directors do can show where a company wants to focus, …