Recently, Mark Zuckerberg of Meta Company posted an article on Instagram stating that it plans to purchase 350,000 H100 GPU chips from chip designer Nvidia before the end of this year. The memory of the Internet is: In 2023, Meta Company plans to develop its own v1 chip for artificial intelligence.
Meta’s third-quarter financial report showed that total expenses in fiscal 2024 will be between US$94 billion and US$99 billion, in part due to expansion in computing power.
In 2023, Meta disclosed the progress of its self-developed chips for the first time, stating that it began to plan to develop an in-house chip for training artificial intelligence models. They named this new chip the Meta Training and Inference Accelerator, or MTIA (MTIAv1) for short, and classified it as a "chip family" that accelerates artificial intelligence training and inference workloads, and plans to launch it in 2025. This custom chip uses the open source chip architecture RISC-V, which is an ASIC in type.
Since the popularity of the OpenAI large model in 2023, Internet giants' demand for AI chips has exploded overnight, and it was hard to find a card for a while. In order to avoid being controlled by others and save costs, major giants have announced the development of self-developed AI chips.
Amazon seems to have taken the lead and already has two AI-specific chips - the training chip Trainium and the inference chip Inferentia; Google has the fourth-generation tensor processing unit (TPU). In contrast, Microsoft, Meta and others still rely heavily on off-the-shelf or custom hardware from chip manufacturers such as Nvidia, AMD and Intel.
According to reports, Microsoft has been secretly developing its own AI chip, codenamed Athena. The chip is manufactured by TSMC and uses a 5nm advanced process. It is planned to be launched as early as 2024. MTIA's v1 also has very strong literal performance. It uses TSMC's 7nm process technology, runs at 800MHz, has a TDP of only 25W, an INT8 integer computing capability of 102.4TOPS, and an FP16 floating point computing capability of 51.2TFLOPS.
According to OpenAI's calculations, since 2012, the amount of computing used in global AI training has grown exponentially, doubling every 3.43 months on average. In 2023, the rush for AI chips by major giants caused Nvidia H100 to continue to increase in price and become out of stock.
Why did Meta spend huge sums of money to buy Nvidia chips regardless of cost? Zuckerberg also said in Thursday's post that he will link artificial intelligence investments to his vision of an AR/VR-driven metaverse and plans to launch the next version of the Llama large-scale language model.
For global Internet giants and technology companies, the competition in 2024 has begun. Will a new round of computing power competition begin?