Lumai, a spin out from the University of Oxford, was founded by a team with expertise in optics over two years ago, aiming to develop a new type of AI processor.
During Computex 2024, Lumai joined the UK semiconductor delegation to Taiwan, attracting the attention of local semiconductor supply chain companies and venture capital firms.
The company's focus is on addressing the performance bottleneck caused by hardware limitations in AI. By leveraging optics, Lumai performs the core AI operations outside traditional silicon architecture.
"While we use digital silicon chips for control and data movement, the actual processing occurs in beams of light in free space, offering scalability and energy efficiency advantages," explained James Spall, co-founder of Lumai. "Optics allow us to achieve remarkable energy efficiency at large scales, making our approach unique, innovative, and disruptive in the field."
The data center market is vast, particularly focusing on large-scale compute for enterprises where thousands of processors are required. Optics also offers efficient interconnection between processors, reflecting a significant industry trend toward utilizing optics for processing and interconnect.
Energy efficiency is a critical concern in data centers, as the power needed to operate devices results in heat generation that must be dealt with, further increasing energy consumption. Enhancing energy efficiency by an order of magnitude can address these performance limitations and energy consumption challenges.
"Scaling up is not a concern for us, as we utilize standard components widely used in various industries. For instance, we incorporate transceivers and displays commonly found in communication and imaging sectors, respectively," said Spall.
Those components are manufactured at scale and are prevalent in data centers, making them readily available. By leveraging already established components and adapting them for their system, the company can significantly shorten the time required to transition from demonstrators to prototypes and eventually to production. This efficient process allows for quicker progress and promising outcomes.
Connecting with partners
"Our purpose here in Taipei is to establish connections and partnerships across various aspects of our technology. While we continue to utilize digital silicon and work on chip development, Taiwan offers significant potential for collaboration in these areas due to its expertise," said Spall. "With expertise in analog electronics and components within the optics industry, Taiwan presents clear synergies for our system."
Taiwan's core strength lies in manufacturing at scale, aligning well with Lumai's ongoing efforts to improve metrics such as speed and energy efficiency in the design process. Lumai is seeking partnerships with renowned organizations for their expertise in chip design, packaging, and manufacturing within the silicon industry, as well as the communications sector. These potential partners specialize in producing components like transceivers, optical fibers, and displays used in the company's system, making them suitable collaborators.
In modern AI algorithms, the fundamental operations involve multiplying and adding numbers on a massive scale. Traditional hardware like GPUs excels at parallel processing for these mathematical operations.
Lumai's approach involves physically using light to perform these math operations, avoiding the need to convert everything into binary strings for manipulation. By utilizing light beams to perform calculations in parallel, Lumai achieves efficiency through analog, physical computation. The optical processor will be integrated into the data center servers, along with the GPU, CPU, and NPU, according to Spall.
Potential of setting up in Taipei
While the company is currently based in Oxford, leveraging the talent and skills in the UK around hubs like Bristol and Cambridge, they acknowledge the value of establishing collaborations globally. Exploring the possibility of setting up locations like Taipei to facilitate closer cooperation with potential partners is a consideration. The company is eager to engage in face-to-face interactions to foster collaborations, recognizing the value of personal meetings over virtual means like Zoom for building meaningful partnerships and accelerating their processes.
Lumai is currently fundraising to support its growth trajectory for the next couple of years. This marks its second fundraising round, with initial backing from venture capitalists upon its spin-out from the University of Oxford over two years ago.
Companies specializing in optoelectronics, specifically modulators and receivers, have expressed interest in Lumai's innovative approach. "We aim to play a pivotal role in the compute stack, enhancing hardware performance essential for advancing AI capabilities. Our optical processors complement traditional CPUs and GPUs, offering compatibility with existing hardware and algorithms in data centers," said Spall.
While Lumai focuses on matrix multiplication with optics as a core operation, it caters to a diverse range of AI models, particularly in AI inference. The company's niche lies in providing efficient and scalable solutions for inference tasks within the expanding AI market. As the demand for AI algorithms grows across various sectors, the volume of inference workloads is poised to increase significantly. While training costs often take center stage, the importance of efficient inference cannot be understated, said Spall.
