NVIDIA is about to release a new generation of AI server GB300, which is expected to be released in the second quarter of 2023 and trial production in the third quarter. The server has significant improvements in thermal design, chip performance, memory specifications and interconnection technology. It will be equipped with a super chip based on the latest B300 GPU, with a memory configuration of up to 288GB, and uses a new generation of ConnectX-8 SuperNIC and 1.6Tbps optical module. Aimed at releasing more powerful AI computing power. However, the introduction of a full water-cooled heat dissipation solution may also lead to a significant increase in costs, and the price of the top configuration is expected to be far higher than the existing GB200NVL72 server.
One of the biggest highlights of the GB300 server is its heat dissipation design. Compared with the previous generation products, the cooling requirements of the GB300 server have increased significantly, and the number of fans used on the motherboard will be reduced. This means that the new generation of servers will rely more on water cooling systems to better cope with the thermal challenges brought by high-performance computing. Improvements in the cooling system will directly affect the performance and stability of the server, allowing it to maintain good working condition under high load conditions.
According to reports, NVIDIA's GB300 server, which is expected to be launched in mid-2025, will undergo a comprehensive design upgrade, covering all aspects from chips to peripherals, in order to release more powerful AI computing power. In terms of chips, the GB300 server will be equipped with a super chip based on the latest B300 GPU. Its FP4 performance will be greatly improved, and the power consumption will also increase from 1000W of B200 to 1400W, reaching twice that of the first generation B100. This change means that the GB300 server will have stronger capabilities when handling complex computing tasks.
In terms of memory, the HBM memory specifications of GB300 will also be upgraded to 288GB, using 8-stack 12Hi HBM3E technology. This will further increase data processing speed and improve overall performance. In addition, the B300GPU may adopt a slot design, which is expected to improve production yield and simplify after-sales maintenance, while the Grace CPU will use LPCAMM memory modules to replace the existing LPDDR5 memory, thus improving performance.
In terms of interconnection technology, the GB300 server will be equipped with a new generation of ConnectX-8SuperNIC and optical modules up to 1.6Tbps, which will significantly increase the data transmission speed and ensure the efficiency of the server when processing massive data. However, the introduction of full water cooling solutions will also increase the cost of servers. According to WccfTech, the top-end price of the GB300 server is expected to be far higher than the current GB200NVL72 server, which is approximately US$3 million, further consolidating its positioning in the high-end market.
Highlights:
The GB300AI server is expected to be released in the second quarter of 2023 and enter the trial production stage in the third quarter.
The new server will adopt a water-cooled heat dissipation design, with a reduced number of motherboard fans and a significant increase in heat dissipation requirements.
The price of the top-of-the-line GB300 server is expected to be far higher than the current GB200NVL72 server, positioning it in the higher-end market.
The upgrade of NVIDIA GB300AI server will significantly improve AI computing capabilities, but high performance also means high cost. Its positioning in the high-end market and its future applications in the AI field deserve continued attention.