At the AI Hardware Summit held in the Bay Area, Allyson Klein from TechArena had an insightful conversation with Letizia Giuliano from Alphawave Semi, discussing the latest advancements in AI connectivity, chiplet designs, and the move towards open standards. The exchange highlighted how these technological developments are shaping the AI industry and the challenges that still need to be addressed.
Letizia began by explaining the role of Alphawave Semi in delivering high-performance connectivity solutions. With a focus on cutting-edge silicon IP, the company is a leader in high-speed interfaces like 100 gig, 200 gig, and PCIe Gen 7. Additionally, they leverage a strong custom silicon business, supported by partnerships with industry leaders such as ARM and foundry ecosystems like TSMC. At Alphawave, Letizia is responsible for product marketing and management, overseeing the development of products that power various customer systems.
When it came to the topic of chiplets, Letizia noted that the industry had moved beyond just talking about them. Currently, designs are being executed, and next-generation products are being powered by chiplets, especially in closed systems. However, the shift towards open ecosystems for chiplets is still a work in progress. According to Letizia, Alphawave is committed to accelerating this transition by working with standards like UCI, which enables a universal chiplet interconnect standard for die-to-die connectivity. This allows customers to integrate IO chiplets with a variety of silicon-proven IPs, improving their cycle time to market.
Customisation is a significant challenge in chiplet design, particularly in the AI space where there is no one-size-fits-all solution. Each customer has unique needs depending on their workloads and where their systems are positioned in the data centre. Alphawave has tackled this by creating a reference platform that can be adapted for different form factors, price points, and system requirements. By standardising the physical layer and offering flexible building blocks, Alphawave can deliver chiplets that suit a range of applications, from high-performance AI systems to more cost-effective setups.
The conversation also touched on the industry’s move towards open standards for chiplet ecosystems. Letizia explained that while there is a push from some sectors to fast-track development, others are more focused on fostering collaboration. She pointed out that collaborations like those in the UCI and ARM ecosystems are examples of how the industry can build on existing standards rather than reinventing the wheel. The focus is on reusing established platforms, such as PCIe and Ethernet, to create new protocols tailored to AI workloads.
In discussing connectivity trends, Letizia emphasised that AI systems require different solutions compared to general-purpose compute systems. While Ethernet remains the standard for front-end networks, AI’s growing computational demands mean that new, more efficient protocols are necessary. For instance, the need for faster GPU-to-GPU connections has driven the adoption of custom interfaces, as existing standards like PCIe aren’t evolving quickly enough to keep pace with AI’s growth.
Letizia also highlighted Alphawave’s recent product launches in 2024, including the AlphaChip 1600, a multi-standard IO product chiplet designed specifically for AI connectivity. This chiplet integrates PCIe Gen 6, CXL 3.0, and 100 gig Ethernet, making it an ideal solution for AI systems requiring high-speed connectivity between NPUs, CPUs, and GPUs. Additionally, the company has launched HBM3 running at 9.6 gig, with further developments in HBM4 on the horizon, providing essential memory components for next-generation AI systems.
As the conversation concluded, Letizia invited those interested in learning more about Alphawave’s products and solutions to visit their website and LinkedIn page, where they regularly share resources, webinars, and updates on their latest innovations.
This discussion provided a comprehensive overview of the advancements in AI connectivity and chiplet designs, shedding light on the challenges and progress in developing open standards. With industry leaders like Alphawave Semi at the forefront, the path towards greater standardisation and interoperability in AI hardware looks promising.
Alphawave IP Group plc (LON:AWE) is a semiconductor IP company focused on providing DSP based, multi-standard connectivity Silicon IP solutions targeting both data processing in the Datacenter and data generation by IoT end devices.
