Innovating Ethernet for High-Performance AI and HPC Workloads
Broadcom has announced the release of its new
Tomahawk Ultra switch chip, a significant step forward in networking for artificial intelligence (AI) and high-performance computing (HPC). Designed specifically to address the increasing demands of these environments, Tomahawk Ultra delivers industry-leading ultra-low latency and massive throughput, redefining Ethernet's role in data centers focused on AI and HPC workloads[1][3].
Key Features of Tomahawk Ultra
- Ultra-Low Latency: Achieves latency as low as 250 nanoseconds, crucial for AI and HPC systems that rely on rapid communication between thousands of chips and servers[2][3].
- Massive Throughput: Provides 51.2 terabits per second (Tbps) switching capacity, specifically optimized for small packet sizes (64 bytes), a typical data format in AI model training and inference[2][3].
- Scale-up Support: Designed to interconnect up to 512 devices at 100 Gbps each, enabling the creation of large, tightly-coupled AI clusters and HPC systems[2].
- Advanced Congestion Control: Includes forward error correction, credit-based flow control, and compatibility with existing Ethernet network interface cards (NICs) and data processing units (DPUs)[2].
- In-Network Compute: Incorporates in-network collectives (INC), a feature inspired by technologies used in Nvidia switches, to minimize node-to-node traffic and speed up job completion[3].
Competing with Nvidia's Proprietary Technologies
Broadcom positions the Tomahawk Ultra as a direct competitor to
Nvidia's NVLink Switch chip, which has become a standard connecting hundreds of GPUs in AI supercomputers. The Tomahawk Ultra not only meets the requirements for bandwidth and latency but also allows for clustering up to four times the number of chips using standard Ethernet protocols—eschewing proprietary networking in favor of open compatibility and industry standards[4].
Broader Impact and Industry Adoption
Broadcom’s Tomahawk Ultra signals a shift in the networking landscape, with Ethernet—long seen as a higher-latency, general-purpose protocol—now optimized to rival specialized AI and HPC interconnects such as NVLink and InfiniBand[2][3]. This advancement is particularly relevant as cloud providers and enterprises build ever-larger clusters for training advanced models, from autonomous systems to large language models like
Chatgpt.
"Tomahawk Ultra is a testament to innovation, involving a multi-year effort by hundreds of engineers who reimagined every aspect of the Ethernet switch," said Ram Velaga, senior vice president and general manager of Broadcom's Core Switching Group[1].
Looking Ahead
With its breakthrough in speed and scalability, Tomahawk Ultra marks Broadcom's commitment to delivering open, high-performance networking solutions for the next generation of AI and HPC infrastructure. The introduction of this chip could redefine architecture choices for hyperscalers and supercomputing centers as they visualize the future of AI at scale[1][3][4].
