Up 6.45%! Switch leader challenges Nvidia, who says Ethernet is not good?

Author: Zhang Yifan

Source: Hard AI

1. Arista: Ethernet Computing Cluster Network Exceeds Expectations, Soars 6.45%

InfiniBand and Ethernet are two communication protocols. Due to its fast transmission speed and good performance, InfiniBand has been widely used in cloud vendors' data centers after being acquired by NVIDIA. On the other hand, Ethernet, due to its slow speed and issues like data loss, has not been able to serve AI data centers. Therefore, the market once believed that only InfiniBand could be used in the AI field.

However, leading switch manufacturer Arista denied this claim at yesterday's conference.

During Arista's financial report meeting yesterday, the company stated that one of the two super-large-scale computing clusters launched by Meta in March this year is using Arista's Ethernet products.

This network connects 24,000 H100 GPUs and performs just as well as InfiniBand. With 8,000 H cards needed to train a GPT-4 model, a cluster of 24,000 H100 GPUs can be used in various scenarios, indicating that Ethernet's performance can fully meet the needs of AI computing clusters.

The company expects to connect 10,000 to 100,000 GPUs by 2025.

Furthermore, the company stated that according to cluster data in collaboration with Broadcom, Arista's Ethernet products have at least a 10% faster job completion speed compared to traditional InfiniBand. This is completely contrary to the market's previous belief that Ethernet performance is inferior to InfiniBand.

FY24Q1 Financial Data:

FY24Q2 Revenue Guidance $1.62-1.65 billion (consensus $1.62 billion), Non-GAAP gross margin approximately 64% (consensus 62.5%), Non-GAAP operating margin approximately 44% (consensus 42.9%);

Guidance for full-year 2024 revenue growth to exceed the previously discussed 10-12%;

2. ARM: Similar to Meta, FY25 Guidance Falls Short of Expectations, Plunges 9% After Hours

In the backdrop of high growth in AI, the market had high expectations for ARM. Following Meta's recent sharp decline due to lower-than-expected guidance, ARM needed to deliver exceptional performance and forecasts to avoid a similar situation.

However, ARM's FY25 guidance provided at this week's earnings call fell short of expectations, leading to a 9% plunge after hours.

Due to ARM's —

• 20-year new agreement with NVIDIA, authorizing the Grace CPU (shipped together with NVIDIA's computing cards);
• Collaboration with cloud vendors such as Amazon, Microsoft, Google to develop in-house data center CPU chips, supplying AI servers for cloud giants;

ARM's share in AI data centers has been gradually increasing. Coupled with the recent higher-than-expected capital expenditures by Meta, Google, Microsoft, the market had expected ARM to maintain high growth. However, the FY25 guidance falling short of expectations resulted in a significant drop after hours

Financial data for 24Q1 performance meeting:

FY25Q1: Revenue of $875-925 million, yoy +29.6%-37.0% (consensus expectation $870 million);

FY25: Revenue of $3.8-4.1 billion, yoy +17.5%~26.8% (consensus expectation $4.025 billion);

3. Citi: Intel's Market Share in Microprocessors Squeezed by AMD and ARM

In 2023, Intel launched the "Four Years Five Nodes" strategy to regain its leading position in processors. However, the latest data shows that the implementation of the strategy is not going very smoothly.

According to a report by Citi, the latest data for 24Q1 shows that Intel is being squeezed by AMD and ARM in various areas such as servers, desktops, and laptops:

ARM: Market share is steadily increasing in all areas (including laptops, desktops, and servers), mainly due to the high performance and low power consumption characteristics of ARM CPUs, which have gained market recognition;

AMD: Market share is growing, especially in the server and desktop CPU markets, with only a slight decrease in laptop market share, mainly because its strategy to enter the "high-end chip" market has been successful;

Intel: Overall, market share in all areas is decreasing, indicating that Intel has not yet emerged from the shadow of previous strategic implementation errors;

Therefore, despite Intel's loud slogans and aggressive corporate strategies last year, as well as the launch of the 7nm process Ultra series processors and AI PC Lake series processors, the latest data indicates that the company has not yet experienced a turnaround.

ARM ——

AMD ——

Intel ——

4. Bank of America: Increasingly Fierce Competition in AI PC Processors

The report predicts growth in the AI PC market, estimating that by 2027, it will increase from the current 50 million units to 167 million units. In 2023, global PC shipments were approximately 247 million units, equivalent to nearly 67.6% of computer shipments in 2027 being AI PCs.

In addition, Bank of America has ranked the performance of AI PC processors currently available in the market: From the chart, it can be seen that Apple's newly released M4 chip performance is second only to Qualcomm Snapdragon X, but ahead of Intel's latest Lake series.

Finally, the report emphasizes the importance of memory in the advancement of AI, stating that larger AI models require larger, faster, and more energy-efficient memory, which will benefit companies like Micron