Data storage and memory power AI at CES 2025

This is the last of my blogs on CES 2025 in Las Vegas. In this blog I will talk about meetings with Phison, Weebit Nano, Seagate Technology and OWC. I also discuss my meetings with Ambiq, Aleksera and Hailo. Digital storage and memory technologies are enabling AI applications in consumer devices.

I met with Phison at CES 2025 and they talked to me about their partnership to supply controllers for Micron SSD products that include Micron Crucial NVMe Gen5 SSDs. Phison said they started supplying NVMe controllers for Crucial products starting with Gen4. Gen5 products are made on advanced 7nm nodes and enable lower power and higher performance.

In particular, at CES 2025, Phison controllers were in the Crucial PS5028-E28 consumer SSD with sequential read and write performance of up to 14.5 GB/s with random synchronous read and write performance of 3M IOPS and average power consumption of 8, 5 W. They were also testing the PS5031-E31T PCIe Gen5 SSD for desktop and laptop use.

Phison was also showing the PS2251-21, the U21, a USB4 SoC NAND flash controller with read and write speeds of up to 4GB/s and supporting up to 8TB without a bridge chip. The U21 is a compact form factor that enables new product designs in large and small enclosures, see below. Phison is also selling its Pascari controllers for enterprise applications.

I met Coby Hanoch from Weebit Nano in a Venetian hotel room during CES and we talked about the latest progress they’ve made with their non-volatile resistive memory technology. The biggest news is that Weebit has licensed its onsemi memory technology. Onsemi plans to use Weebit as a memory to replace eFlash, NOR flash, on its 65nm Treo platform. This platform targets advanced automotive and industrial applications. Onsemi is interested in using Weebit’s ReRAM for the technology’s cost and energy efficiency. This agreement increases the number of fab licensees for Weebit to three, including Onsemi, DB HiTek and SkyWater.

Weebit’s ReRAM technology is an end-of-the-line, BEOL process with many advantages over eFlash. It uses only two masks, with known fab materials and using conventional process equipment. Weebit has also demonstrated product manufactured by Global Foundries on a 22nm node.

Seagate was in another Venetian hotel room showing off their latest products and talking to analysts about future plans. In particular, product displays were from the Seagate, LaCie and Firecuda brands, shown below.

Lance Ohara of Seagate and Dave Helmly of Adobe gave presentations focusing on digital storage for creators in both media and entertainment. Lance said intelligent data curation enables creation, and like many other industries, Seagate is looking at how AI can help with future digital storage solutions.

OWC had an exhibit at Showstoppers where they were showing off some of their products. In particular, its company Thunderblade X12 and active optical cable, as well as the general availability of a Thunderbolt 5 Hub, shown below.

The ThunderBlade X12 is a RAID transport production solution for use with up to 12K RAW video and stereoscopic 360-degree VR, special video, content and will be available in March. The OWC USB4 40Gb/s active optical cable allows reliable and high-performance long-distance connection of Thunderbolt 4/3 and USB4/3/2 devices. The fiber optic cable offers bandwidth up to 40 Gb/s and power delivery up to 240 W, as well as 8K video transmission up to 15 feet. It features USB-C connectivity and with the fiber optic connection eliminates the 2 meter distance limit of Cu-based Thunderbolt and USB4 cables.

The OWC Thunderbolt 5 Hub is now generally available and offers 3 Thunderbolt 5 ports and one USB-A port with up to 80 Gb/s bi-directional data speeds and up to 120 Gb/s for higher display bandwidth needs.

I also had a chance to speak with 3 AI chip companies targeting the embedded consumer and other markets at or just before CES 2025, including Ambiq, Alexera, and Hailo. Ambiq announced that their Ambiq 4 SoC chips with MRAM memory were being used in the ThinkAR AILens, see below, as well as some lower-cost, lower-power hearing aids.

AILens weighs only 35 grams and has over 10 hours of battery life, due to the lowest possible power modes with a non-volatile memory for AI training. The Arm Cortex®M4F microprocessor in the Ambiq 4 reaches up to 192 MHz for processing graphics, audio and AI models.

Additionally, the company said their Apollo 5 SoC chips doubled the available MRAM memory to 4MB. Ambiq’s chips are manufactured by TSMC and are also used in various Fitbits and other wearables.

I met with Fabrizio Del Maffeo, CEO and co-founder of Alexera AI, a Netherlands-based startup working on AI inference solutions at the edge. He showed me three generations of their AI chips, the Metis AI Accelerator, below with the earliest on the left and the latest on the right.

The company is offering an M.2 AI Edge accelerator card with a quad-core Metis AIPU and a variety of evaluation systems working with ARM, aetina, Dell, Lenovo and Advantech. The company’s products are focused on AI computing and video analytics. The overall AI platform integrates Voyager SDK and the company’s hardware.

In December 2024 Alexera and Arduino announced a collaboration that combines Axelera AI’s Metis AI Platform with Arduino Pro SOMs, providing developers and businesses with easy-to-use and cost-effective tools to power AI innovation in all industries. Although I didn’t see it myself, Alexera was demonstrating an AI-powered industrial monitoring system featuring pre-trained LLM (Phi-3) and Arduino Portenta X8 to analyze sensor data (temperature, humidity, CO2, etc.) in real time.

I also visited Hailo’s demo room at the Venetian where I saw their AI chip, see below, and various applications. They were showcasing their Hailo-8 AI processor, the Hailo-10H M.2 AI acceleration module for comprehensive generative AI, and a Hailo-15 chip, introduced in April 2024, offering sophisticated capabilities of vision processing for smart cameras and offering up to TOP 10 performance. . The company said they had over 300 customer qualifications with 90% of their applications based on vision.

Hailo-8 is available as a Chip On Board, COB, M.2 and PCIe card configurations. Many of the Hailo demonstrations included a Hailo AI HAT+ for a Raspberry Pi. In particular, there were demonstrations illustrating several vision processing applications, see below, as well as interesting applications such as a radar edge processing device and video processing for aerial drones.

There was also a demo unit with an AIC storage server, as well as an exhibit with a solid state SSD storage server.

Digital storage and memory capability play a crucial role in modern computing, including AI. The 2025 CES show showed advances in NAND flash controllers, non-volatile memory and various AI applications.