NVIDIA GTC Preview: Three Key Focus Areas
NVIDIA will hold the GTC conference from March 17 to 21, with CEO Jensen Huang delivering a keynote speech on March 19 and holding an investor meeting. The conference will focus on the development of accelerated computing, upgrades to CPO switch technology, and the progress of the NVL288 rack and Rubin. The launch of new products B300 and GB300 is expected, with a 50% improvement in the computing performance of the B300. At the same time, the expansion of AI clusters will drive upgrades to CPO switch technology
Core Viewpoints
NVIDIA will hold the GTC conference from March 17 to March 21, with CEO Jensen Huang delivering a keynote speech at 1:00 AM Beijing time on March 19, followed by an investor meeting at 11:30 PM on the same day. Additionally, NVIDIA will host its first Quantum Day on March 20 to explore cutting-edge quantum computing technologies. We believe the core focus of this conference should be: 1) the development trend of accelerated computing and the release of B300 and GB300; 2) upgrades to CPO switch technology; 3) updates on the NVL288 rack and Rubin's progress. Considering that Rubin and Vera are expected to be released simultaneously, we believe this GTC may only provide subsequent chip roadmaps.
Core Focus #1: Blackwell Architecture Update, B300 Computing Performance May Increase by 50%
Following the release of B200 and GB200 at the 2024 GTC, the industry predicts that NVIDIA will launch the new B300 and GB300 products this time. Semi Analysis predicts that B300, based on TSMC's 4NP process, will have a computing power increase of about 50% compared to B200. Compared to the TDPs of GB200 and B200 at 1.2kW and 1kW respectively, part of the performance improvement comes from an additional 200W power consumption, with the TDPs of GB300 and B300 HGX increasing to 1.4kW and 1.2kW respectively, making the liquid cooling requirements for GB300 servers more urgent. Additionally, Toms Hardware predicts that B300 may feature 12-Hi HBM3E, providing up to 288GB of memory and 8TB/s bandwidth. Cloud vendors are promoting GB300 not only due to its FLOPS and memory improvements but also because of its higher flexibility. Unlike B200 and GB200, NVIDIA may no longer sell complete B300 server chassis but will only provide core components such as the SXM Puck module, Grace CPU, and Axiado HMC. Furthermore, we expect Rubin to fully adopt a 3nm process, but this time may only disclose the latest progress, as we believe the new CPU architecture Vera and the new GPU architecture Rubin may be released together.
Core Focus #2: Accelerated Expansion of AI Clusters May Drive Upgrades to NVIDIA's CPO Switch Technology
The rapid development of AI clusters makes scale-out crucial for improving data transmission efficiency and breaking bandwidth bottlenecks. This GTC should also focus on co-packaged optics (CPO) solutions driven by silicon photonics technology. NVIDIA plans to mass-produce the Quantum 3400 and Spectrum 5 CPO switches in July and December 2025, respectively. According to Semi Vision, NVIDIA's next-generation CPO switches may adopt a chiplet design to enhance switching capacity and optical engine performance Among them, the Spectrum 5 has a switching capacity of 204.8T (4 chips of 51.2T, with 16 optical engines of 3.2T per chip); the Spectrum 6 can reach a switching capacity of 409.6T (4 chips of 102.4T, with 16 optical engines of 6.4T per chip).
Core Focus #3: NVL288 Rack-Level Interconnect Architecture Upgrade, Possibly Launched in Sync with Rubin
We expect NVIDIA to unveil the NVL288 rack system at the GTC conference, which will consist of four NVL72 interconnects. Due to the GB200 NVL72 single rack power reaching up to 120kW causing heat issues, we predict that the NVL288 rack will revert to the early HGX's UBB+OAM structure or adopt a whole Bianca motherboard design, where the CPU is installed on the UBB via SMT, while the GPU connects to the OAM through slots and transmits signals to the UBB via copper wires. We believe the company has a strong barrier in high computing power, NVLink, and CUDA.
Author of this article: He Pianpian, Yi Chuyan, Source: Huatai RuiSi, Original title: "Huatai | NVIDIA GTC Preview: Three Key Focus Areas"
Risk Warning and Disclaimer
The market has risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial conditions, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Investment based on this is at one's own risk