Broadcom has begun shipping the Tomahawk Ultra, a 51.2-terabit-per-second Ethernet switch ASIC designed to interconnect high-performance computing and artificial-intelligence accelerators with latency as low as 250 nanoseconds. The chip serves as a traffic controller that links hundreds of processors in a single rack. Broadcom senior vice president Ram Velaga said the device can tie together four times as many chips as Nvidia’s NVLink switch while relying on an enhanced form of standard Ethernet instead of proprietary protocols. Tomahawk Ultra anchors the company’s new “scale-up Ethernet” strategy, which Broadcom claims can build fabrics of up to 1,024 accelerators, positioning it as an alternative to the UALink initiative championed by AMD and Intel. Fabricated on Taiwan Semiconductor Manufacturing Co.’s 5-nanometer node, the processor took roughly three years to develop and is already in the hands of data-center customers, Broadcom said.
Broadcom Ships Tomahawk Ultra Ethernet Switch with 250ns Latency for AI and HPC https://t.co/aylnDuKeHt #Broadcom #AIwire
$AVGO Broadcom Ships Tomahawk Ultra: Reimagining the Ethernet Switch for HPC and AI Scale-up
Broadcom launches new Tomahawk Ultra networking chip in AI battle against Nvidia https://t.co/BDhpCljyTo