Hyper Racetrack | Meta Platforms' Self-developed Inference Chip: Deployment to be Completed This Year

Author: Zhou Yuan / Wall Street See

While Meta Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms set a record for the highest single-day increase in the history of the US stock market, an internal document of the company was exposed: Meta Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms plans to deploy a new version of its self-developed custom chip (ASIC) in the company's data centers this year to support the further development of its AI business.

In the past 2023, the company has made "great progress" in promoting AI and the metaverse, said Mark Zuckerberg, CEO of Meta Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms, at its earnings conference.

A spokesperson for Meta Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms confirmed the plan and stated that its self-developed chip will work in synergy with the NVIDIA chips purchased by Meta Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms to enhance AI computing power and solidify the company's AI infrastructure capabilities.

The spokesperson said that the self-developed chip by Meta Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms will be put into production in 2024, reducing the cost of purchasing AI accelerator cards and reducing reliance on NVIDIA.

Public information shows that Meta Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms' ASIC is internally referred to as "Artemis" and its main performance is focused on the inference field. The technical development is based on the second-generation internal chip product line announced in 2023. On February 1st, Meta Platforms announced its financial results for the fourth quarter of the fiscal year 2023, ending on December 31st. The company's financial data exceeded market expectations by a large margin. At the same time, Meta Platforms also made an optimistic estimate for its first-quarter operating performance this year, which led to a record-breaking single-day increase in its stock price on February 2nd (Beijing time).

Self-developed ASIC to reduce costs

Since the release of ChatGPT-3.5 by OpenAI on December 22, 2022, the cost of AI chips, infrastructure capabilities, and energy consumption required for the application of GenAI (Genetic Artificial Intelligence) technology has become a "money drain" for technology companies, to some extent offsetting the explicit or implicit benefits associated with this technology.

Therefore, major US technology giants, including Microsoft, Amazon, Google, Intel, AMD, Qualcomm, and others, have all joined the army of self-developed AI chips, and Meta Platforms is also one of them.

The price of NVIDIA H100 has soared to $25,000-$30,000, which means that the cost of a single query for ChatGPT will increase to about $0.04. Even just maintaining the basic operation of ChatGPT would require an annual cost of about $16 billion.

Meta Platforms' "Artemis" chip, like its predecessors, can only perform "inference" workload tasks. The models are required to use their algorithms to make ranking judgments and respond to user prompts. The company shared its first-generation Meta Platforms training and inference accelerator (MTIA) in 2023. The details of the project have not been disclosed since then.

In a statement, a spokesperson for Meta Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms said, "We believe that the internally developed AI accelerator is highly complementary to commercial GPUs and can provide the best combination of performance and efficiency for specific workloads on Meta Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms."

According to Mark Zuckerberg's video message released in January this year, Meta Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms plans to have approximately 350,000 flagship AI chips H100 from NVIDIA by the end of 2024. The H100 is currently the most popular server GPU for AI workloads developed by NVIDIA. Zuckerberg emphasized that, combined with the self-developed new AI chip and AI chips from other potential suppliers, Meta Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms will accumulate computing power equivalent to 600,000 H100 AI GPUs.

In the video, Zuckerberg revealed the updated roadmap for Meta Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms' AI plan: Meta Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms Platforms will build a new AI roadmap around the upcoming Llama 3, which is currently undergoing AI training. Llama 3 will compete with Google's recently released Gemini model, OpenAI's GPT-4, and the upcoming GPT-5 model. The earliest rumors about Llama 3 appeared in August 2023, but Meta Platforms has never officially acknowledged the technological iteration of this new GenAI product until the video released by Mark Zuckerberg in January this year.

On January 30th of this year, Zuckerberg directly mentioned Llama 3 in a tweet. The industry believes that Meta Platforms' disclosure of the deployment of the self-developed "Artemis" AI inference chip in 2024 is related to Llama 3.

Zuckerberg hinted that Llama 3 is likely to be Meta Platforms' first multimodal model that supports multimedia and voice input. Llama 2, on the other hand, is just a chatbot where users can only ask questions and write stories.

Currently, Meta Platforms has not officially revealed the release date of Llama 3, perhaps due to the timing of when Artemis can truly be deployed in the company's data centers.

Public information shows that Meta Platforms is accelerating the construction of its data centers (IDCs) with a focus on GPU computing. The latest efforts in updating its IDC mainly involve building large clusters with thousands of accelerators. The core network of the IDC is organized in a grid form, with a bandwidth of 1 TB per second between accelerators. Meta Platforms has 21 data centers worldwide. But obviously, this is not enough. To achieve Zuckerberg's ultimate goal, more GPUs are needed.

"It is clear that the next generation of services needs to build comprehensive general intelligence, build the best AI assistant, create progress in various AI fields for enterprise creators, from reasoning to planning to coding to memory and other cognitive abilities," Zuckerberg said. "People also need new AI devices that combine AI and the metaverse, because over time, I think many of us will have frequent conversations with AI throughout the day."

VR/AR department achieves record-breaking revenue

Meta Platforms' remarkable performance was announced on February 1st.

On this day, Meta Platforms released its financial results for the fourth quarter of the fiscal year 2023, ending on December 31, 2023.

According to the earnings report, Meta Platforms achieved a revenue of $40.11 billion in the fourth quarter of 2023, a 25% increase compared to the same period last year, surpassing analysts' expectations of $39.01 billion. This is also the largest revenue increase for Meta Platforms since the third quarter of 2021. Net profit increased by 201% year-on-year to $14.017 billion, exceeding market expectations of $12.89 billion. Diluted earnings per share increased by 203% year-on-year to $5.33, higher than the market expectation of $4.95.

For the full year of 2023, Meta Platforms achieved a revenue of $134.902 billion, a 16% increase year-on-year, and net profit increased by 69% year-on-year to $39.098 billion.

Furthermore, Meta Platforms expects strong performance in the first quarter of this year. Meta Platforms announced a stock repurchase of $50 billion and will distribute dividends for the first time in its history in March this year, including Class A and Class B common stock, with a cash dividend of $0.50 per share.

Stimulated by multiple positive news, Meta Platforms' market value increased by approximately $200 billion on February 1st, marking the first time in the history of the US stock market.

At the close of trading, Meta Platforms recorded a huge increase of 20.32%, surpassing the previous single-day performance records set by Apple and Amazon. Apple's market value increased by $190.9 billion on November 10, 2022, and Amazon's market value increased by $190.8 billion on February 4 of the same year. NVIDIA's market value increased by $184.1 billion on May 25, 2023.

It is worth mentioning that Meta Platforms' VR/AR division, Reality Labs, is responsible for the development of Quest headsets, Ray-Ban smart glasses, the Horizon platform, as well as AR glasses and their neural wristband input devices.

According to Meta Platforms' latest financial report, Reality Labs' quarterly revenue reached $1.07 billion, setting a new record. Meta Platforms' CFO, Susan Li, stated that this record-breaking revenue was "driven by the sales of Quest 3 during the holiday season." Quest 3 was launched on October 10, 2023, and Meta Platforms' fourth-quarter revenue includes October, November, and December.

Apple began accepting pre-orders for the Vision Pro on January 19th of this year. As of February 1st, it is reported that over 200,000 units have been sold (at a price of $3,499 per unit). According to users at Apple stores in the United States on February 2nd, some people were willing to pay an additional $2,000 per unit. This has brought Apple approximately $700 million in revenue.

According to Meta Platforms' financial report, the cost of Reality Labs in the fourth quarter of 2023 reached a record high of $5.72 billion, resulting in a quarterly loss of $4.65 billion for the department.

However, the market seems to be very tolerant of this, considering XR headsets like Quest are still relatively early-stage technology and far from mature. Therefore, this loss is seen as a necessary investment in early-stage development. Currently, Meta Platforms has not yet launched AR glasses, but over 50% of Reality Labs' expenses are focused on the development of AR glasses.

According to data released by Valve in January 2024, the usage of VR devices on the Steam platform increased by 0.4% in January. Among them, Quest 2 ranked first with a usage share of 40.64% in January. Valve Index HMD ranked second with a usage share of 15%. Quest 3 ranked third, accounting for 14.05% of the overall usage share.