Phillip Burr is the Head of Product at Lumai, with over 25 years of experience in global product management, go-to-market and leadership roles inside leading semiconductor and technology firms, and a proven track record of constructing and scaling services.
Lumai is a UK-based deep tech company developing 3D optical computing processors to speed up artificial intelligence workloads. By performing matrix-vector multiplications using beams of sunshine in three dimensions, their technology offers as much as 50x the performance and 90% less power consumption in comparison with traditional silicon-based accelerators. This makes it particularly well-suited for AI inference tasks, including large language models, while significantly reducing energy costs and environmental impact.
What inspired the founding of Lumai, and the way did the concept evolve from University of Oxford research right into a industrial enterprise?
The initial spark was ignited when considered one of the founders of Lumai, Dr. Xianxin Guo, was awarded an 1851 Research Fellowship on the University of Oxford. The interviewers understood the potential for optical computing and asked whether Xianxin would consider patents and spinning out an organization if his research was successful. This got Xianxin’s creative mind firing and when he, alongside considered one of Lumai’s other co-founders Dr. James Spall, had proven that using light to do the computation at the guts of AI could each dramatically boost AI performance and reduce the energy, the stage was set. They knew that existing silicon-only AI hardware was (and still is) struggling to extend performance without significantly increasing power and value and, hence, if they might solve this problem using optical compute, they might create a product that customers wanted. They took this concept to some VCs who backed them to form Lumai. Lumai recently closed its second round of funding, raising over $10m, and bringing in additional investors who also imagine that optical compute can proceed to scale and meet increasing AI performance demand without increasing power.
You’ve had a formidable profession across Arm, indie Semiconductor, and more — what drew you to affix Lumai at this stage?
The short answer is team and technology. Lumai has a formidable team of optical, machine learning and data center experts, bringing in experience from the likes of Meta, Intel, Altera, Maxeler, Seagate and IBM (together with my very own experience in Arm, indie, Mentor Graphics and Motorola). I knew that a team of remarkable people so focused on solving the challenge of slashing the fee of AI inference could do amazing things.
I firmly imagine that way forward for AI demands recent, progressive breakthroughs in computing. The promise of with the ability to offer 50x the AI compute performance in addition to cutting the fee of AI inference to 1/tenth in comparison with today’s solutions was just too good a chance to miss.
What were a few of the early technical or business challenges your founding team faced in scaling from a research breakthrough to a product-ready company?
The research breakthrough proved that optics could possibly be used for fast and really efficient matrix-vector multiplication. Despite the technical breakthroughs, the largest challenge was convincing those that Lumai could succeed where other optical computing startups had failed. We needed to spend time explaining that Lumai’s approach was very different and that as an alternative of counting on a single 2D chip, we used 3D optics to succeed in the degrees of scale and efficiency. There are in fact many steps to get from lab research to technology that will be deployed at scale in an information center. We recognized very early on that the important thing to success was bringing in engineers who’ve experience in developing products in high volume and in data centers. The opposite area is software – it is crucial that the usual AI frameworks and models can profit from Lumai’s processor, and that we offer the tools and frameworks to make this as seamless as possible for AI software engineers.
Lumai’s technology is claimed to make use of 3D optical matrix-vector multiplication. Are you able to break that down in easy terms for a general audience?
AI systems have to do quite a lot of mathematical calculations called matrix-vector multiplication. These calculations are the engine that powers AI responses. At Lumai, we do that using light as an alternative of electricity. Here’s how it really works:
- We encode information into beams of sunshine
- These light beams travel through 3D space
- The sunshine interacts with lenses and special materials
- These interactions complete the mathematical operation
Through the use of all three dimensions of space, we will process more information with each beam of sunshine. This makes our approach very efficient – reducing the energy, time and value needed to run AI systems.
What are the predominant benefits of optical computing over traditional silicon-based GPUs and even integrated photonics?
Since the rate of advancement in silicon technology has significantly slowed, each step up in performance of a silicon-only AI processor (like a GPU) leads to a major increase in power. Silicon-only solutions eat an incredible amount of power and are chasing diminishing returns, which makes them incredibly complex and expensive. The advantage of using optics is that after within the optical domain there may be practically no power being consumed. Energy is used to get into the optical domain but, for instance, in Lumai’s processor we will achieve over 1,000 computation operations for every beam of sunshine, each cycle, thus making it very efficient. This scalability can’t be achieved using integrated photonics as a result of each physical size constraints and signal noise, with the variety of computation operations of silicon-photonic solution at only at 1/eighth of what Lumai can achieve today.
How does Lumai’s processor achieve near-zero latency inference, and why is that such a critical factor for contemporary AI workloads?
Although we wouldn’t claim that the Lumai processor offers zero-latency, it does execute a really large (1024 x 1024) matrix vector operation in a single cycle. Silicon-only solutions typically divide up a matrix into smaller matrices, that are individually processed step-by-step after which the outcomes should be combined. This takes time and leads to more memory and energy getting used. Reducing the time, energy and value of AI processing is critical to each allowing more businesses to profit from AI and for enabling advanced AI in essentially the most sustainable way.
Are you able to walk us through how your PCIe-compatible form factor integrates with existing data center infrastructure?
The Lumai processor uses PCIe form factor cards alongside a regular CPU, all inside a regular 4U shelf. We’re working with a variety of information center rack equipment suppliers in order that the Lumai processor integrates with their very own equipment. We use standard network interfaces, standard software, etc. in order that externally the Lumai processor will just seem like another data center processor.
Data center energy usage is a growing global concern. How does Lumai position itself as a sustainable solution for AI compute?
Data center energy consumption is increasing at an alarming rate. In keeping with a report from the Lawrence Berkeley National Laboratory, data center power use within the U.S. is anticipated to triple by 2028, consuming as much as 12% of the country’s power. Some data center operators are contemplating installing nucleus power to offer the energy needed. The industry needs to have a look at different approaches to AI, and we imagine that optics is the reply to this energy crisis.
Are you able to explain how Lumai’s architecture avoids the scalability bottlenecks of current silicon and photonic approaches?
The performance of the primary Lumai processor is simply the beginning of what’s achievable. We expect that our solution will proceed to offer huge leaps in performance: by increasing optical clock speeds and vector widths, all with no corresponding increase in energy consumed. No other solution can achieve this. Standard digital silicon-only approaches will proceed to eat increasingly cost and power for each increase in performance. Silicon photonics cannot achieve the vector width needed and hence firms who were taking a look at integrated photonics for data center compute have moved to deal with other parts of the information center – for instance, optical interconnect or optical switching.
What role do you see optical computing playing in the longer term of AI — and more broadly, in computing as an entire?
Optics as an entire will play an enormous part in data centers going forward – from optical interconnect, optical networking, optical switching and naturally optical AI processing. The demands that AI is placing on the information center is the important thing driver of this move to optical. Optical interconnect will enable faster connections between AI processors, which is crucial for giant AI models. Optical switching will enable more efficient networking, and optical compute will enable faster, more power-efficient and lower-cost AI processing. Collectively they’ll help enable much more advanced AI, overcoming the challenges of the slowdown in silicon scaling on the compute side and the speed limitations of copper on the interconnect side.
