#United States #San Jose #DreamBig #AI-SuperNIC #Mercury

DreamBig Semiconductor Unveils the Mercury AI-SuperNIC Chip

DreamBig Semiconductor, Inc. has entered the tech scene with an innovative breakthrough in the form of the Mercury AI-SuperNIC chip. This state-of-the-art technology boasts an impressive 800Gbps bandwidth, making it a premier choice for the rapidly evolving landscape of artificial intelligence platforms.

Unveiled on January 6, 2025, in San Jose, California, the Mercury chip is crafted specifically to ensure ultra-efficient connections between various GPU setups, a critical feature in demanding AI applications and high-performance computing (HPC). The full hardware offload facilitates seamless functioning, which is paramount for future-proofing AI technologies.

The Mercury chip's design includes:

• - A hardened RDMA (Remote Direct Memory Access) engine that seamlessly integrates with existing standards such as RoCE v2 and the newly established UEC (Ultra Ethernet Consortium).
• - Unmatched throughput of 800Mpps, all while minimizing latency and power consumption, resulting in significant operational efficiencies.
• - Programmable congestion control that dynamically adapts to the specifications and needs of various data centers, ensuring that communication remains efficient and smooth.

Innovative Features Enhancing AI Applications

Mercury isn’t just about speed; it comes packed with features designed to tackle AI workloads effectively. Key functionalities include:

• - Multi-pathing and packet spraying that enhance data management capabilities.
• - Out-of-order packet placement while still ensuring in-order message delivery, maintaining the integrity of data flow.
• - Advanced telemetry congestion notifications and packet trimming that optimize network performance.
• - Selective retransmission support, ensuring that necessary data is retrieved without overloading the system.

With software drivers compliant with the UEC, the Mercury chip supports GPU-to-GPU communication that adheres to industry standards while providing remarkable throughput and minimized lag times.

Steve Majors, the Senior Vice President of Engineering at DreamBig, emphasized the company's legacy, stating, "This year, we celebrate 25 years of RDMA development from Infiniband through to iWARP and now UEC. Our experience informs our innovative re-architecture of RDMA, tailor-made for the AI era. With our hardware offload efficiency and sophisticated congestion control methods, DreamBig sets an exceptional benchmark for the future of AI networking."

A Major Leap for AI Technologies

Mercury is hailed as a transformative monolithic chip, redefining discrete networking capability for Tensor Processing Units (TPUs) and GPUs. DreamBig's solution combines higher efficiency levels with cost-effectiveness, essential assets for developers aiming to meet the growing demands of AI technology. The compatibility with chiplet technology means that Mercury can scale up to an impressive 12.8Tbps RDMA throughput, ensuring that it is well equipped for vast networks.

DreamBig will begin releasing limited samples of the Mercury platform in Q4 2025 to selected customers, giving early adopters a chance to successfully integrate this leading-edge technology into their systems.

Don’t miss the opportunity to see the Mercury in action! DreamBig will showcase their technology solutions at the CES 2025 event, taking place from January 7 to 10 in the Venetian Bellini Suite #2002, Las Vegas, Nevada.

About DreamBig Semiconductor

Founded in 2019, DreamBig has positioned itself as a forward-thinking company focused on building advanced semiconductor solutions. With a commitment to affordability and scalability, DreamBig aims to drive the advancement of AI technologies while addressing market needs across various sectors, including large language models, generative AI, and the automotive industry.

Their innovative Chiplet Hub technology continues to redefine possibilities within the semiconductor space, liberating developers to push new boundaries in AI and beyond.