The most interesting point in the comments
Supermicro starts selling workstations ARS-511GD-NB-LCC-01-G2 (PDF) Based on the Nvidia GB300 Grace Blackwell Ultra chip, it sells for $92,887.47 (excluding tax and shipping). The computer costs about seven times as much as the M5 Ultra chip-based Apple Mac Studio, which sells for $12,299.
Image source: supermicro.com
The Supermicro workstation has a total memory capacity of 748 GB, while the Mac Studio has a total memory capacity of 256 GB. The Nvidia platform splits its 748 GB of memory between 496 GB of system memory on the 72-core Grace processor and 252 GB of HBM3e memory on the Blackwell Ultra graphics subsystem – a machine that allows you to run trillion-parameter models natively. In terms of data storage, two 1.92 TB M.2 NVMe hard drives and two 960 GB disk drives are used; there are additional M.2 slots supporting PCIe 5.0 and 6.0 standards. There are two 400 Gigabit Ethernet ports on the Nvidia ConnectX-8 SuperNIC as well as a dedicated LAN port for remote management. There are several USB interfaces and one interface for connecting a monitor; it supports IPMI for remote diagnosis and monitoring.
Delivering up to 20 Pflops of performance for FP4 digital formats – the platform is three times faster than the H200 accelerator. The artificial intelligence agent can run continuously with a generation rate of up to 1200 tokens per second. There are multiple fans and a liquid cooling system; the power supply is provided by a 1600W power supply with a titanium grade efficiency of up to 94%, allowing the system to be used in high load mode. The workstation weighs approximately 40 kg. Supermicro designed it for research labs and enterprises that need to run large language models locally.
Tell Google to receive more frequent links to our news about artificial intelligence
If you find an error, select it with your mouse and press CTRL+ENTER.
