News: Optoelectronics
15 November 2024
Ayar Labs showcasing future of AI infrastructure with Fujitsu, Intel Foundry, Corning and Altera
In booth #810 at Supercomputing 2024 (SC24) in Atlanta, GA, USA (17–22 November), silicon photonics-based chip-to-chip optical connectivity firm Ayar Labs of San Jose, CA, USA is demonstrating a broad range of optical I/O implementations for AI scale-up fabrics. The firm’s optical I/O solution is said to eliminate the inefficiencies, high power consumption and increasing costs of copper interconnects and pluggable optics, boosting the performance and economics of AI infrastructure. At SC24, Ayar Labs is featuring its technology in various AI infrastructure concepts with Fujitsu Ltd, Intel Foundry, Corning Inc, Altera, and others.
“Copper interconnects and pluggable optics have already failed to support large-scale AI workloads,” says CEO & co-founder Mark Wade. “Optical I/O overcomes these bottlenecks by enabling scale-up fabrics that connect more GPUs or accelerators per cluster across greater distances. This increases cluster performance without increasing power per rack, which paves the way for more efficient, cost-effective, and profitable AI infrastructure.”
Ayar Labs and Fujitsu are showcasing the future of hardware and system architecture design in response to the growing demands of AI and HPC workloads. A conceptual model of Ayar Labs’ optical I/O solution integrated with the Fujitsu A64FX processors in a PRIMEHPC FX700 chassis is on display in Ayar Labs’ booth representing up to 16Tbps of bi-directional bandwidth with the integration of two TeraPHY optical I/O chiplets per CPU.
“Collaborating with Ayar Labs allows us to showcase a new era of AI and HPC system design,” says Naoki Shinjo, senior VP, head of the Advanced Technology Development Unit at Fujitsu. “Our joint architectural concept at SC24 highlights our commitment to push the boundaries of innovation and deliver unparalleled solutions that meet tomorrow’s most demanding computational challenges.”
In addition to the model on display, on 19 November at 11:30am (in room B206) Ayar Labs and Fujitsu are presenting a joint talk ‘New Optical-Based Scale-Up Fabrics to Meet the Performance and TCO Requirements of Next-Generation AI/HPC System Architectures’. Ayar Labs’ chief technology officer & co-founder Vladimir Stojanovic and Miki Atsushi, a member of Fujitsu’s Advanced Technology Development Division, are discussing the key challenges in current architectures and the requirements for future systems, while Ayar Labs is explaining how optical I/O-based scale-up fabrics enhance system-level performance and reduce total cost of ownership for these systems.
Other demonstrations and displays planned for SC24 include:
- Innovative AI System Architecture Tool: Developed by Ayar Labs, this new tool can predict the performance and economics of AI inference workloads, including future agentic models. It helps to assess next-gen AI infrastructure by simulating various GPU and network configurations. Visitors to the booth can see the impact Ayar Labs’ optical I/O has on scaling large language model (LLM) inference, with profitability gains of up to 20x and a 3–4x interactivity improvement in future GPT models.
- Advancing AI with Optical FPGAs: Ayar Labs, Altera and Corning have partnered to pioneer a new optical field-programmable gate array (FPGA) solution. This proof of concept provides 4Tbps of bi-directional I/O data transfer by integrating Ayar Labs’ TeraPHY optical I/O chiplets with Corning’s advanced glass waveguide modules to Altera’s FPGA fabric. This unique solution provides what is claimed to be unprecedented bandwidth and latency for compute-intensive applications such as generative AI, LLMs and AI scale-up fabrics.
- 4Tbps In-Package Optical I/O Solution: SC24 guests can experience a live demonstration of Ayar Labs’ 4Tbps optical I/O solution powered by its SuperNova light source. Visitors can see first-hand ultra-efficient data transfer at extremely low latency without the need for forward error correction (FEC). Ayar Labs optical I/O provides a high level of energy efficiency, power density and performance per watt that is suitable for AI models with trillions of parameters, advanced HPC designs, and more.
- Detachable Optical Connector Ecosystem: Ayar Labs is highlighting the latest advancements in detachable optical connectors for optical I/O, including the latest Intel Foundry detachable photonic glass interconnect and the Teramount TeraVERSE detachable connector. These connectors enable flexible, efficient and cost-effective deployment and maintenance of Ayar Labs’ in-package optical I/O solutions for AI.
Ayar Labs showcasing optical interconnect solutions for AI infrastructure at OFC
Ayar demos first 4Tbps optical solution, paving way for next-gen AI and data-center designs
