Phillip Burr, Head of Product at Lumai – Interview Series

Phillip Burr is the head of the Lumai product, with more than 25 years of experience in managing global products, roles to go to the market and driving in pioneering semiconductor and technology companies, a busy record of products and expansion of products.

Lumai is a deep -based technical company that develops 3D optical computing treatments to accelerate the burdens of artificial intelligence work. By performing the complications of the matrix-matrix using rafters of light in three dimensions, its technology provides up to 50x performance and energy consumption of 90 % less compared to traditional silicone accelerators. This makes it particularly suitable for artificial intelligence inference tasks, including large language models, while significantly reducing energy costs and environmental impact.

What inspired the establishment of Lumai, and how did the idea of ​​Oxford University’s research evolved into a commercial project?

The initial spark was ignited when one of the founders of Lumai, Dr. Xianxin Guo, got a 1851 research fellowship at Oxford University. Understanding interviews the possibility of visual computing and asked whether Xianxin will consider patents and overcome a company if his research is successful. This creative mind of Xianxin has made, and when, along with one of the other founders of Lumai, Dr. James Spal, has proven that the use of light to do the account in the heart of artificial intelligence can significantly enhance the performance of artificial intelligence and reduce energy. They have realized that only the current Silicon intelligence devices (and still) are struggling to increase performance without a significant increase in energy and cost, and therefore, if they can solve this problem using the visual account, they can create a product that customers want. They took this idea for some VCS who supported them to form Lumai. Lumai recently closed its second tour of financing, raising more than $ 10 million, and brought additional investors who also believe that the visual account can continue to expand the scope of increased demand for artificial intelligence without increasing energy.

You have a great profession across ARM, the semi -independent conductors, and more – what enriched you to join Lumai at this stage?

The brief answer is the team and technology. Lumai has an impressive team from the experts of the visual and automatic centers and the data center, as they bring experience such as Meta, Intel, Altera, Maxiel, Seaagate and IBM (along with my own experience in ARM, Indie, Mentor Motions and Motorola). I knew that a team of great people focus on solving the challenge of reducing the cost of artificial inferences can do amazing things.

I believe a firm belief that the future of artificial intelligence requires new and innovative computing. The promise was to be able to provide 50x AI’s performance for an AI account in addition to reducing the cost of artificial intelligence conclusion to 1/10 compared to today a very good chance to wrap it.

What are some of the early technical or commercial challenges that your founding team in the scaling faced to penetrate a search for a company ready for products?

The search penetration has proven that optics can be used to multiply the rapid and very effective matrix. Despite technical breakthroughs, the biggest challenge was to persuade people that Lemay could succeed as other visual startups failed. We had to spend some time explaining that the Lumai approach was completely different and that instead of relying on a single -dimensional segment, we used 3D optics to reach the levels of size and efficiency. There are of course many steps that you can get from laboratory research to technology that can be widely published in the data center. We have realized early that the key to success is to bring engineers who have experience in developing products in large size and in data centers. The other field is a program – it is necessary to benefit the standards of models and standard materials from artificial intelligence from the Lumai processor, and to provide tools and frameworks to make this smooth as possible for artificial intelligence software engineers.

Lumai is said to use 3D optical matrix proliferation. Can you break this in simple phrases for the general public?

Artificial intelligence systems need to do a lot of sports accounts called the matrix. These accounts are the engine that works on artificial intelligence responses. In Lumai, we do this using light instead of electricity. Here is how to work:

1. We purify information in symptoms of light
2. These light symptoms are transmitted through a three -dimensional area
3. Light interacts with lenses and special materials
4. These reactions complement the sports process

Using the three dimensions of the area, we can process more information with each beam of light. This makes our approaches very effective – reducing the energy, time and cost needed to operate the artificial intelligence systems.

What are the main advantages of optical computing on conventional and even optical optical graphics processing units?

Since the rate of progress in silicone technology has slowed down, every step in the performance of the artificial intelligence processor of Silicon only (such as GPU) leads to a significant increase in energy. Silicon solutions only consume an incredible amount of strength and chase the decreasing returns, making them incredibly complex and expensive. The advantage of using optics is that once there is any power in the visual field it is not practically consumed. Energy is used to reach the visual field, but, for example, in the Lumai processor, we can achieve more than 1,000 mathematical processes for each beam of light, each cycle, which makes it very effective. This expansion can not be achieved using integrated guides due to both material size restrictions and sign noise, with the number of silicone solutions calculations and the ritual stroke in only 1/8, which can achieve Lumai today.

How does the Lumai processor achieve almost zero inference, and why is this decisive factor of the burdens of modern artificial intelligence?

Although we do not claim that the Lumai processor provides a zero proliferation, it carries out a very large matrix processor (1024 x 1024) in one cycle. Silicon solutions are usually divided only into a smaller matrix, which are treated individually step by step, then the results should be combined. This takes time and leads to more memory and energy. Reducing time, energy and the cost of treating artificial intelligence is essential to allowing more companies to take advantage of artificial intelligence and to enable advanced artificial intelligence in the most sustainable ways.

Can you walk for us by how to combine the PCIE -compatible shape factor with the current data center?

The Lumai processor uses PCIE cards along with the Standard CPU, all within a standard 4U shelf. We are working with a group of data center shelf suppliers so that the Lumai processor integrates with their own equipment. We use standard interfaces, standard programs, etc., so that an external Lumai processor looks like any other data center processor.
Energy use in the data center is an increased global anxiety. How does Lumai put itself as a sustainable solution to the artificial intelligence account?

Energy consumption in the data center increases at a risk rate. According to a report issued by the Lawrence Berkeley National Laboratory, the use of the United States’ data center energy is expected to lead to three times by 2028, and consume up to 12 % of the country’s strength. Some data center operators consider stabilizing the core power to provide the necessary energy. The industry needs to consider different approaches to artificial intelligence, and we believe that optics are the answer to this energy crisis.

Can you explain how you avoid the structure of Lumaii, the expansion of the current methods of silicon and light?

The performance of the first Lumai processor is just the beginning of what can be achieved. We expect our solution to continue to provide tremendous leaps in performance: by increasing the visual clock speeds and displaying vectors, all without a similar increase in the energy consumed. There is no other solution that can achieve this. Digital silicone methods will only continue to consume more, more cost and energy for each performance increase. Silicon guns cannot achieve the required vector width, and thus the companies that were looking for integrated spoils to calculate the data center to process other parts of the data center – for example, optical or visual switching.

What role do you see visual computing in the future of artificial intelligence – and more broadly, in computing as a whole?

Optics as a whole will play a big role in data centers to move forward – from visual delivery, visual networks, visual switching, and visual artificial intelligence processing, of course. The demands for placing artificial intelligence in the data center are the main engine of this step to visual. Visual communication will enable faster connections between artificial intelligence treatments, which is necessary for large artificial intelligence models. The optical switch will enable more efficient networks, and the visual account will enable artificial intelligence faster, more efficient in energy and less expensive. Collectively, they will help enable artificial intelligence more advanced, overcoming the challenges of slowing silicone on the calculation side and restrictions of copper speed on the interconnected side.

Thank you for the wonderful interview, readers who want to know more Lumai.