本文由半导体产业纵横（ID：ICVIEWS）编译自aimultiple
This article will introduce the top AI chip suppliers to help enterprises choose the right chip.
As you can see in the figure below, the number of parameters (i.e., width and depth) of the neural network as well as the size of the model are increasing. To build better deep learning models and powerful AI applications, organizations need to increase computing power and memory bandwidth.
Powerful general-purpose chips, such as CPUs, cannot support highly parallel deep learning models. As a result, there is a growing demand for AI chips that support parallel computing power, and McKinsey believes that this trend will continue.
However, even Intel, which has many world-class engineers and a strong research background, takes three years to develop its own AI chips. As a result, for most companies, buying chips from these vendors or renting capacity from cloud GPU providers is the only way to develop powerful deep learning models. This article will introduce the top AI chip suppliers to help enterprises choose the right chip.
What are the leading AI chip manufacturers?
1. Nvidia
Since the 90s of the 20th century, Nvidia has been producing graphics processing units (GPUs) for the gaming sector. Both PlayStation3 and Xbox use Nvidia graphics arrays. The company also produces artificial intelligence chips such as Volta, Xavier and Tesla. Thanks to the generative AI boom, Nvidia has achieved excellent results in 2023 with a valuation of trillions, cementing its position as the market leader in GPUs and AI hardware.
NVIDIA's chipsets are designed to solve business problems in various industries. For example, Xavier is the foundation for autonomous driving solutions, while Volta is aimed at data centers. The DGX A100 and H100 are NVIDIA's successful flagship AI chips, designed for AI training and inference in the data center. Nvidia released the H200, B200, and GB200 chips; HGX servers, such as the HGX H200 and HGX B200 that combine 8 of these chips; The NVL series and GB200 SuperPod combine more chips into large clusters.
Cloud GPUs
When it comes to AI workloads in the cloud, Nvidia has a near monopoly, with most cloud vendors only using Nvidia GPUs as cloud GPUs. Nvidia also introduced the DGX Cloud offering, which provides cloud GPU infrastructure directly to enterprises.
2.AMD
AMD is a chipmaker with CPU, GPU, and AI accelerator products. For example, AMD's Alveo U50 data center accelerator card has 50 billion transistors. The Accelerator can run 10 million embedded datasets and execute graph algorithms in milliseconds.
AMD introduced the MI300 for AI training workloads in June 2023 and will compete with NVIDIA for market share in the market. As ChatGPT has shown, the rise of generative AI, the rapidly increasing demand, has made Nvidia's AI hardware difficult to source, so there are startups, research institutes, enterprises, and tech giants adopting AMD hardware in 2023.
AMD has also partnered with machine learning companies like Hugging Face to enable data scientists to use their hardware more efficiently.
The software ecosystem is critical because hardware performance relies heavily on software optimization. For example, AMD and NVIDIA have a public disagreement on the H100 and MI300 benchmarks. The focus of the divergence is the packages and floats used in benchmarks. According to the latest benchmarks, the MI70 appears to be better or comparable to the H300 for inference at 10B LLMs.
3. Intel
Intel is the largest player in the CPU market and has a long history of semiconductor development. In 2017, Intel became the first AI chip company in the world to exceed the $1 billion mark in sales.
Intel's Xeon CPUs are suitable for a variety of jobs, including processing in the data center, and have had an impact on the company's commercial success.
Gaudi3 is Intel's latest AI acceleration processor. Since its public release in April 2024, there are currently limited benchmarks for its performance.
4.Alphabet/谷歌云平台
GoogleCloud TPUs are purpose-built machine learning accelerator chips that power Google products like Translate, Photos, Search, Assistant, and Gmail. It is also available through Google Cloud. Google released the TPU in 2016. The latest TPU is Trillium, the sixth-generation TPU.
Edge TPU is another accelerator chip from Google Alphabet that is smaller than a penny and is designed for edge devices such as smartphones, tablets, and IoT devices.
5.AWS
AWS manufactures the Tranium chip for model training and the Inferentia chip for inference. Although AWS is the leader in the public cloud market, it started building its own chips after Google.
6、IBM
IBM will release its latest deep learning chip, the Artificial Intelligence Unit (AIU), in 2022. IBM is considering using these chips to power its watson.x generative AI platform.
The AIU is built on the "IBM Telum processor" that powers the AI processing power of the IBM Z mainframe servers. Among the use cases that stood out when Telum processors were introduced included fraud detection.
IBM has also demonstrated that combining compute and memory can improve efficiency. These have been demonstrated in the NorthPole processor prototype.
7. Alibaba
Alibaba produces inference chips such as Hanguang 800.
What are the leading AI chip startups?
Here are some startups in the AI chip industry whose names may be heard a lot in the near future. Although these companies are only new, they have already raised millions of dollars.
Figure 2: Total financing for AI chip manufacturers, source: Statista
8.SambaNova system
SambaNova Systems was founded in 2017 with the goal of developing high-performance, high-precision hardware-software systems for high-volume generative AI workloads. The company developed the SN40L chip and raised more than $1.1 billion in funding.
Notably, SambaNova Systems also rents out its platform to businesses. SambaNova Systems' AI Platform-as-a-Service approach makes its systems easier to adopt and encourages hardware reuse for the circular economy.
9. Cerebras Systems
Cerebras Systems was founded in 2015. In April 2021, the company announced the launch of a new AI chip model, the Cerebras WSE-2, with 850,000 cores and 2.6 trillion transistors. There is no doubt that WSE-2 is a big improvement over WSE-1, which has 1.2 trillion transistors and 400,000 processing cores.
Celebra's system works with several pharmaceutical companies, including AstraZeneca and GlaxoSmithKline, because WSE-1's effective technology accelerates genetic and genomic research and shortens the time to drug discovery.
10. Groq
Groq was founded by former Google employees. The company stands for LPU, a new model of artificial intelligence chip architecture designed to make it easier for companies to adopt their systems. The startup has raised about $350 million and produced the first models, such as the GroqChip processor, the GroqCard accelerator, and more.
The company focuses on LLM inference and has released a benchmark for the Llama-2 70B.
In the first quarter of 2024, 70,000 developers registered on its cloud platform and built 19,000 new applications, the company said.
On March 1, 2022, Groq acquired Maxeler, which provides high-performance computing (HPC) solutions for financial services.
*Disclaimer: This article was created by the original author. The content of the article is his personal point of view, and our reprint is only for sharing and discussion, and does not mean that we agree or agree, if you have any objections, please contact the background.