How could the IT industry evolve in the face of growing demands for performance and energy efficiency? The answer could well lie in the use of light for calculation, an approach which is gaining interest and investment. This week, a new California-based company announced its entry into the industry with promising technology.
Opticore, a new company photonic calculationformalized its launch and obtained its first funding. Headquartered in Fremont, California, Opticore aims to scale data centers and artificial intelligence through an innovative approach to photonic computing.
The startup develops optical processing units (OPUs) that perform the same computational tasks as GPUs, but using light and waveguides instead of electrical components. This method enables calculations at orders of magnitude lower energy costs and higher clock speeds.
The advantages of this technology
« The era of computing with light has arrived thanks to Opticore technology. Their architecture overcomes memory bottlenecks in CMOS electronics and enables the processing of billions of parameters on a single chip using 100 times less power. Their chips are manufactured by standard foundry processes. This funding allows them to build a much larger chip “, declared the two co-founders Zaijun Chen and Ryan Hamerly.
As photonics becomes indispensable in data centers for communication between chips, computing through light could maximize its potential for next-generation, energy-efficient supercomputers.
The inherent limitations of CMOS electronics arise from fundamental material properties. When switching from the “off” state to the “on” state, transistors are penalized, similar to driving a car in “stop-start” mode. When turned on, they do not conduct perfectly and continue to dissipate energy as heat. Opticore solves this problem by using other fundamental properties, those of light.
- Eliminate performance bottlenecks with photonics: By leveraging photonic chips that solve the data I/O problem, Opticore enables faster data transfer, greater bandwidth and real-time reprogrammability, making it ideal for demanding AI learning tasks.
- Unmatched Power Efficiency: Opticore’s photonic chips transmit data and compute directly using light, requiring virtually no power dissipation in data communications. The result is energy-efficient computing without the cooling costs typically associated with CMOS systems. This argument would have been valid for other optical systems, but unfortunately these systems are far from accessible due to the large size of the chips and the energy cost of electro-optical conversion in I/O. These bottlenecks have recently been overcome through the establishment of a data exchange system. These bottlenecks have recently been overcome with Opticore’s unique temporal mapping architecture.
- Low manufacturing risk: unlike state-of-the-art CMOS processes (28nm feature size). This approach allows us to quickly scale up production in a cost-effective manner without having to resort to complex and costly new manufacturing processes. Opticore can deliver high-performance computing solutions faster and more affordably, while leveraging existing semiconductor manufacturing infrastructure.
- Efficient Scaling: By using time-division multiplexing in its algorithmic architecture, Opticore can improve speed by orders of magnitude over legacy optical architectures – up to a million times faster.
This enables a paradigm shift in terms of computing power and energy efficiency. The image below compares Opticore’s performance to existing cutting-edge GPUs and TPUs, highlighting the significant benefits of this change with improvements of nearly 100x in power efficiency and 100x in terms of compute density, with the potential for an additional 100x improvement in power efficiency and compute density of nearly 10,000x.
« I have never seen such intelligent computing architecture. Opticore’s technology with co-packaged optics represents the solution for scalable AI computing added Professor Mengjie Yu, co-founder of Opticore and professor at the University of California, Berkeley.
Impact on artificial intelligence
« AI models are limited by the power of current supercomputers. It is simply not economically viable to train much larger models. This technology could enable a jump directly to machine learning models that would otherwise be out of reach in the near future “, emphasized Professor Dirk Englund, advisor at Opticore, in an interview with MIT News.
The funding round was led by Alex Turnbull of Sagax Capital and Karan Danthi of Jetha Global, two recognized investors in the deep tech space. It was also backed by Dorjee Sun of Bioeconomy, a climate-focused investor, among others.
« The Opticore team has charted a path toward a state of the world where high-performance computing, climate goals, and supply chain security are fully compatible. They achieved this with a new and creative approach on existing platforms that can be manufactured in the United States, Europe and Singapore “, said Alex Turnbull of Sagax Capital.
« At Jetha Global, our extensive research shows that optics will play an important role in the evolution of data centers, starting with transforming the connection layer and eventually extending to xPUs “, concluded Karan Danthi, Chief Investment Officer at Jetha Global.
Illustration caption: Opticore’s chips are manufactured using standard foundry services, with stacked opto-electronic components and hyperbonded HBMs. The MIT startup promises “abundant, clean, and secure computing for the future.” Photo: Opticore.
Source : Opticore
