NVIDIA launched the GB200 NLV4, and the GB300 may be released in March next year

NVIDIA's GB200 NLV4 fills a critical market demand. Additionally, NVIDIA plans to launch its next-generation GB300 AI server platform at the GTC conference from March 17 to 20, 2025, featuring the new B300 AI GPU with a power of up to 1400W, and FP4 performance that is 1.5 times that of the B200.

Recently, NVIDIA launched the mid-range platform GB200 NLV4, equipped with Grace CPU and Blackwell GPU, achieving a balance between performance and energy efficiency, specifically designed for modern data centers.

Analysts believe that the launch of the GB200 NVL4 coincides with NVIDIA's restructuring of its product lineup—gradually phasing out the old NVL platform in favor of more advanced options like NVL4.

For medium-sized organizations or those focused on energy efficiency, the GB200 NLV4 is a very suitable choice. Compared to the NVL72, which is designed for large-scale deployments and features edge connectors to support backbone configurations, NVL4 offers a more compact and energy-efficient alternative.

Balancing Performance and Energy Efficiency

At the core of the GB200 NVL4 are two Grace CPUs based on the Arm architecture, each equipped with 72 Arm Neoverse V2 cores, providing a total of 144 cores necessary for executing high-performance tasks. Additionally, the GB200 NVL4 is paired with four Blackwell GPUs.

Analysts indicate that this configuration is particularly suitable for AI, high-performance computing (HPC), and other data-intensive applications.

Furthermore, the GB200 NVL4 is equipped with six MCIO connectors beneath each CPU, ensuring fast PCIe connections to support seamless data transfer. These connectors facilitate the integration of network interface cards (NICs), solid-state drives (SSDs), and other critical components, providing the system with flexibility and scalability.

In terms of memory, the GB200 NVL4 offers up to 1.3 TB of unified memory, designed for organizations with memory-intensive workloads, ensuring efficient data processing.

In addition to its powerful performance, another highlight of the GB200 NVL4 is its energy efficiency.

NVIDIA expects that a fully configured server will consume about 6kW of power. While this power consumption is not negligible, it is approximately half that of NVIDIA's earlier systems, such as the DGX-1 or HGX-1, which consumed around 3.5kW.

Will the GB300 be Launched Next Year?

According to reports, NVIDIA plans to launch its next-generation GB300 AI server platform at the GTC conference from March 17 to 20, 2025, featuring the new B300 AI GPU with a power of up to 1400W, and FP4 performance that is 1.5 times that of the B200 Currently, NVIDIA's GB200 AI uses 192GB HBM3E memory with an 8-layer stacked configuration. According to reports from UDN, the GB300 will use 288GB HBM3E memory with a 12-layer stack, a technology developed by SK hynix.

The fast connection components and network cards of the GB300 AI server platform have also been upgraded, with the speed of the optical module increased from 800G to an ultra-fast 1.6T. UDN reports that performance and equipment have been "enhanced in every aspect," making it NVIDIA's "market-winning weapon."

There are also some upgrades, such as a slot design, the computing board will use LPCAMM, and the capacitor tray may become a standard configuration for the next-generation GB300 NVL72 AI server cabinet. UDN also mentioned that the BBU (Backup Battery Unit) "may be optional."

Undoubtedly, such a luxurious configuration is expected to result in a considerable cost for the components.

Analysts estimate that the mass production price of the BBU module is around $300, while the total price of the BBU for the GB300 AI server could reach $1,500. The production cost of supercapacitors is expected to be between $20 and $25, and the GB300 NVL72 AI server cabinet will require more than 300 supercapacitors