MemryX – Scalable, High-Performance AI at the Edge

AI has transformed the world around us. From reasoning models to agentic AI, they're all running on data centers, but very soon, they could run on our handheld devices. AI models are moving towards being developed on the cloud, and as they're becoming more efficient on smaller models, edge solutions are the more energy-efficient option. MemryX's forward-looking edge architecture prioritizes both data movement and software that's simple as well as practical to deploy.

Deployment on the edge poses three key challenges. The first is to achieve the accuracy and ease of use that customers have on the cloud. The second is to have the same unlimited performance capacity that the cloud has. Third is the power performance constraint that edge computing has. Due to space constraints, it's a challenge to get the same liquid cooling that's available at a data center.

MemryX's solution addresses all three concerns in terms of its ease of use. The team has over a thousand models running on their hardware with no need for compression, pruning, or quantization, so what you're training is what you get during inferencing. For edge AI inference, MemryX's MX3 AI chip can provide 10X performance and lower costs compared to mainstream GPUs.

What's more, all of MemryX's chips work together as a single logical unit with a single connection to the host using a standard interface. Whether it's an Arm-based processor, an Android-based system, a Linux-based system, or an x86-based processor, MemryX's processor works all the same.

The USB and PCIe interfaces are critical components that allow MemryX's chips to communicate with the host, and that's where Cadence's top-notch tools come into play. Cadence IP for PCIe 2.0 and USB Gen 3 and Gen 2 have allowed MemryX to get its products to the market without having to develop them in-house. The team also used Cadence's Xcelium Logic Simulator and Verisium Manager for IP-level and SoC-level verification, essential tools for ensuring there are no design bugs.

The Cadence Genus Synthesis Solution helps the MemryX team get the best RTL implementation in terms of power and performance, while the Cadence Allegro X Design Platform is crucial for their system design. It allows MemryX to have a really compact design on their M2 module.

Puneet Singh, VP of engineering at MemryX, said, "Cadence and MemryX have been partners since day one, since our inception, really. Cadence provides us with the best-in-class IP [and] the best-in-class design tools, and MemryX focuses on AI. And really, the two of us together can work and bring the best-in-class AI products into the market."

Learn more about how MemryX is making edge AI processing highly adaptable.

"Designed with Cadence" is a series of videos that showcases creative products and technologies that are accelerating industry innovation using Cadence tools and solutions. For more Designed with Cadence videos, check out the Cadence website and YouTube channel.