Nvidia has positioned itself perfectly in the booming market for GPU chips, driven by the soaring demand from generative AI models. However, what if there were a chip that could deliver similar performance at a lower price? Enter Lemurian Labs, an innovative startup founded by alumni from Google, Intel, and Nvidia. Today, the company announced a $9 million seed investment aimed at realizing this ambitious vision.
“Our mission at Lemurian is to revolutionize accelerated computing,” explained Jay Dawani, co-founder and CEO of Lemurian. “The traditional computing methods we’ve relied on are nearing their limits. This isn't just about challenging existing architectures, but responding to the fundamental physics of semiconductors.”
Lemurian is set on creating a new type of chip along with software to enhance the accessibility, efficiency, affordability, and sustainability of processing AI workloads. Dawani elaborated on their approach to computing, asserting that it involves three key components: “math, memory, and movement. The objective is interconnectivity. Data is stored in memory, navigated through interconnects to processing units where it is manipulated, and then written back to memory. This is the conventional model: data travels to the compute resources.”
Lemurian aims to disrupt this model. Instead of relocating data for processing, they propose moving computation closer to the data. “We need to minimize that distance,” Dawani stated, “so instead of moving data, we’re relocating compute resources.”
He pointed out that while GPUs were originally designed for graphics tasks, they've evolved to tackle various roles due to their raw processing power. “When a single design has to fulfill multiple functions, it often falls short in performance. This is the primary limitation of GPUs, and we aim to address that,” Dawani noted.
To tackle this challenge, Lemurian is rethinking the math behind their chip architecture. Dawani recounted that early chip developers opted for a floating-point approach, as the logarithmic method proved difficult to implement. He asserts that Lemurian has overcome that barrier.
“The advantage of a logarithmic number system is significant. It transforms costly multiplication and division operations into simpler additions and subtractions, which are far more efficient in hardware. As a result, we save on chip area and energy while enhancing speed and precision—an essential factor for reducing costs in processing large language models,” he explained.
So, how did they achieve this? “We discovered that by constructing our system in a certain manner and expanding the definition of a large number system, we can produce precise solutions that are smaller and more accurate than traditional floating-point methods for an equal number of bits,” Dawani shared.
“As we increase the number of bits, the dynamic range improves dramatically without additional resources, allowing us to explore novel architectures. The success of our unique number system is crucial; without it, we would face the same limitations as others.”
The company is taking a methodical approach by first releasing the software component of their system, which they anticipate will be widely available by Q3 of next year. The hardware development poses greater challenges, demanding substantial time and investment, but they envision this phase will follow in the subsequent years.
Currently, Lemurian employs 24 specialized engineers skilled in this type of development, but Dawani plans to recruit six more shortly. If they secure Series A funding, he hopes to expand the team by an additional 35 people within a year.
The recent $9 million funding round was spearheaded by Oval Park Capital, with additional investments from Good Growth Capital, Raptor Group, and Alumni Ventures, among others.
Launching a company like Lemurian and bringing a new chip to market is an ambitious and costly endeavor. However, if they succeed, it could significantly lower the costs and improve the efficiency of developing generative AI models and future technologies.
Ranking
