After the performance, it rose nearly 30%! Why is Nvidia's recent "big move" worth paying attention to?

Since the stunning quarterly report of US stocks was announced after last Wednesday's trading, Nvidia has been doing well, with its stock price soaring by 27%. Now it is about to break through the $400 mark and set a new historical high.

The continuous rise in stock prices is not only due to performance support, but also related to the various heavyweight news released by Nvidia in succession.

Guosheng Securities saw that as an absolute leader in AI computing power supply, Nvidia has had many important news recently, from better-than-expected financial reports and next quarter guidance, to the release of a series of new chips, new architecture supercomputers, switches, and industry applications at Computex 2023. These actions are boosting market confidence in computing power.

The above-mentioned securities firm believes that in the future, computing power is equivalent to productivity, and there is no ceiling for AI's pursuit of computing power. There are still many expected differences that will continue to catalyze the upward trend of the computing power sector and the optical communication market.

The following is an overview of Nvidia's recent highlights:

AI Supercomputer Server

Nvidia released the latest AI computing architecture Grace Hopper and the corresponding server solution DGX GH200 at Computex 2023.

Grace Hopper adopts the latest CPU and GPU integrated design scheme. A single GH200 Grace Hopper integrates 72 ARM architecture CPUs, with a computing power of 4 PFLOPS TE, 96GB HBM3 memory, and 576GB video memory. A DGX GH200 integrated with 256 Grace Hopper chips has a total video memory of 144TB (realized by exchanging networks), and with the blessing of the fourth-generation NVLink, the GPU interconnect bandwidth reaches 900GB/s, and the total computing power reaches 1 exaFLOPS@FP8, perfectly adapted to large-scale AI training/inference needs.

Guosheng Securities believes that with the increase in bandwidth, the demand elasticity of the new architecture for optical modules is also increasing. Grace Hopper and GH200 have been put into production, and Google Cloud, Meta, and Microsoft are early customers.

It is expected that GPU servers will dominate the data center market in the future. Data centers are moving towards accelerated computing, and generative AI has pushed this process forward. In the case of limited enterprise Capex, future investments will focus on GPU servers or CPU+GPU hybrid servers similar to Grace Hopper.

In the past, enterprises needed to spend $10 million to purchase a data center with 960 CPUs@11GWh to run a large language model. Now, they only need to spend $400,000 to purchase a small cabinet with 2 GPUs@0.13GWh to run a large language model.

Switches

NVIDIA has released the world's first AI Ethernet network platform, Spectrum-X, with RDMA technology that can be used for cross-server memory sharing. The Spectrum-4 switch, which will debut in 2022, has also officially entered mass production, providing 64 800GbE ports or 128 400GbE ports, and can connect 8192 GPUs distributed across two-tier networks, with a switching rate of 51.2T.

NVIDIA once again emphasized the importance of high-density architecture and communication networks for graphics card performance.

Guosheng Securities believes that the "barrel effect" of internal communication capabilities in supercomputing centers will be further amplified. The market generally overlooks the importance of switching networks.

NVIDIA Infiniband achieved record revenue in the first quarter. According to NVIDIA's calculations, if an enterprise spends $500 million to build a switching network, the throughput difference between Infiniband and Ethernet is 15%-20%. At the same time, Infiniband is more suitable for GPU loads, and the unit price of optical modules adapted to the Infiniband protocol is higher, which will further enhance the value of optical modules in supercomputing.

Diverse Customers for Computing Power

Vertical fields such as finance, medicine, games, and industry are all accelerating the deployment of AGI, for example:

1. Bloomberg will deploy the BloombergGPT and FINPILE datasets with 50 billion parameters, focusing on financial tasks;

2. Amgen is using DGX Cloud and NVIDIA BioNeMo large models to accelerate drug discovery;

3. NVIDIA Avatar Cloud Engine (ACE) is used to customize game AI models, which can bring intelligent NPCs through AI-supported natural language interaction, and is expected to completely change the gaming experience;

4. Nvidia Omniverse empowers the industrial Internet, bringing fast and intelligent solutions to the manufacturing industry.

The demand for computing power is penetrating into various industries, which will continuously strengthen the logic of "there is no ceiling for computing power pursuit" and "the more diverse the applications, the more diverse the computing power."