The massive image: As Nvidia begins rolling out its hotly anticipated Blackwell server GPUs, data relating to an upcoming lineup refresh is beginning to emerge. Set for launch in late 2025, GB300 options quite a few upgrades in reminiscence, connectivity, cooling, and different areas in comparison with its predecessor.
Provide chain sources have revealed detailed specs of Nvidia’s GB300 information middle GPU to UDN (translated by TechPowerUp). Codenamed “Blackwell Extremely,” the chip goals to supply substantial efficiency enhancements over the corporate’s lately launched GB200.
One of the crucial notable enhancements is a reminiscence enhance, with GB300 rising from its predecessor’s 192 GB to 288 GB of HBM3e RAM. Moreover, Nvidia shifted the structure from eight layers to 12, and the computing board makes use of LPCAMM reminiscence.
Every B300 chip on the coronary heart of the GB300 would require 1,400W. Furthermore, networking has been upgraded from ConnectX 7 to ConnectX 8, and optical modules have expanded from 800G to 1.6T.
Every card guarantees a 50 p.c FP4 efficiency enchancment over GB200. Nvidia’s utilization of FP4, which is well-suited to inference workloads, is likely one of the main causes behind the heightened anticipation for GB300.
Nevertheless, the elevated efficiency comes with an additional price. Every cupboard requires over 300 supercapacitors, which carry a roughly $20 to $25 manufacturing price per unit. Nvidia may also provide non-obligatory battery backup items costing roughly $1,500 for a whole unit.
Prior experiences indicated that the corporate is contemplating a socket configuration for GB300, which might simplify manufacturing and upkeep necessities. Sadly, the choice would additionally enhance energy and cooling calls for.
Mass manufacturing and shipments of GB200 are anticipated to peak in mid-2025 following delays stemming from late-stage design flaws and overheating issues. Regardless of the setbacks, orders from Nvidia’s shoppers, together with Microsoft, are backlogged for a yr. The AI growth has created intense demand for Nvidia’s server GPUs, rising its market cap to over $3 trillion and making it one of many world’s three most dear firms.
Blackwell Extremely is predicted to reach in late 2025, nevertheless it stays unclear whether or not this schedule will collide with Nvidia’s subsequent technology, Rubin, which could emerge across the identical time. Rubin will make the most of TSMC’s 3nm N3 course of node and HBM4 reminiscence.
Nvidia can also be getting ready to launch its next-generation shopper Blackwell GPUs subsequent yr with a probable reveal coming throughout CES in January.
