At the same time, pointing at Nvidia and TSMC, Intel can't sit still.
Text丨Zhijia.com Wang Xin
Editor丨Langlang Mountain and Mingzhi Mountain
On April 9, local time in the United States, Intel held the Intel ON Industry Innovation Conference for customers and partners. Intel CEO Pat Gelsinger unveiled the latest AI chip, the Gaudi 3.
The Gaudi series is Intel's chip brand specially launched for AI application scenarios, and the AI computing performance of Gaudi 3 is 2 to 4 times that of the previous generation Gaudi 2.
Although a few days ago, Nvidia released the latest Blackwell architecture GPU B100 and B200 chips at GTC, Intel's Gaudi 3 still chose Nvidia's earlier main product H100 as a competitor in terms of specific parameters.
The most immediate upgrade to the Gaudi 3 is in terms of performance and cost.
Performance: According to the official introduction, compared with NVIDIA H100, the model training speed and inference speed of the Intel Gaudi 3 AI chip can be increased by 40% and 50% respectively. The chip is also capable of supporting a variety of large models, including Llama, Stable Diffusion for Bunsheng Diagram, Whisper for speech recognition, and more.
Network connectivity is also a key point in GPUs.
Intel claims that the Gaudi 3 AI accelerator can connect up to tens of thousands of accelerators via Ethernet's common standards. Intel Gaudi 3 will bring 4x more BF16 AI computing power and 1.5x more memory bandwidth than the previous generation, with FP8 throughput doubling to 1835 TFLOPS.
Unlike Nvidia Blackwell, which uses the latest TSMC 3nm process, Gaudi 3 is manufactured using TSMC's 5nm process. At the same time, the number of tensor cores has been upgraded from 24 to 32.
In addition, Intel also announced that a large number of companies such as Naver (South Korean Internet giant), Bosch, IBM, Ola and others have become customers and partners of Intel's Gaudi accelerator. In the second quarter of this year, the Gaudi 3 will be shipped to OEMs, including Dell, HP, Lenovo and Supermicro, and will be officially launched in the third quarter.
The cost upgrade is reflected in a quote from Intel – "The Gaudi 3 costs a fraction of the cost of the Nvidia H100." ”
Das Kamhout, vice president of Intel, said bluntly at the meeting that compared with NVIDIA H100, Intel Gaudi 3 is expected to shorten the training time of large models by 50% on the 7 billion and 13 billion parameter Llama2 models and 175 billion parameter GPT-3 models, while the energy consumption and cost performance are better.
Pat Kissinger also added: "The performance of the Gaudi 3 will be on par with the Nvidia H200, and even better in some areas." However, this statement is not supported by comparative data, and he does not elaborate.
At the end of last year, Pat Kissinger blasted Nvidia's CUDA ecosystem as "shallow and narrow", and claimed that inference technology will be more important than training for artificial intelligence.
By all indications, Intel wants to prove that the inference market is the focus of competition, and the Gaudi 3 is exactly what Intel is using to break into the AI inference market.
In addition to the Gaudi 3 chip, Intel also announced six generations of Xeon processors at the Vision conference.
On the same day, Intel shared the latest information on next-generation products and services in various segments of enterprise AI. Intel has released the Intel Xeon 6, the next generation of processors for the data center, cloud, and edge. Intel Xeon 6 processors with E-cores will be available in Q2 2024, followed by Intel Xeon 6 processors with P-cores.
Now, Intel wants to build a broad AI ecosystem alliance to drive innovation in the field of AI.
Intel outlined a strategy for open, scalable AI systems that includes hardware, software, frameworks, and tools. On the same day, Intel and a number of companies announced that they would create an open platform to help enterprises drive AI innovation. The initiative aims to develop open, multi-vendor generative AI systems that deliver ease of deployment, performance, and value through RAG (Retrieval Enhanced Generation) technology. RAG accelerates the adoption of generative AI in the enterprise by enabling enterprises to augment the large number of existing proprietary data sources running on standard cloud infrastructure with open large language model (LLM) capabilities.
At the same time, Intel is also making all-round efforts in AI chip design and foundry, pointing at the two giants of Nvidia and TSMC.
In February this year, Intel launched its first system-level foundry for the AI era, and is willing to OEM chips for all customers, including Nvidia, Qualcomm, Google, Microsoft and AMD.
Intel aims to become the world's second-largest foundry by 2030 and hopes to produce 50% of the world's semiconductors in the United States and Europe within a decade. However, at present, this proportion is only 20%, and most of the world's production is concentrated in Asia.
As an old giant that adheres to the IDM model, Intel is encountering crazy attacks from different opponents in various segments.
According to TrendForce data, Intel ranked among the world's top 10 fabs in the third quarter of 2023, but fell out of the top 10 in the fourth quarter. In the fourth quarter, TSMC's share of the foundry market increased further from the previous quarter, exceeding 60%.
In 2023, Intel's overall revenue will be $54.2 billion, down 14% year-over-year. Intel's Data Center and Artificial Intelligence Division (DCAI) revenue was $15.5 billion, down 20% year-on-year, and the chip foundry business suffered an operating loss of $7 billion, an increase of about $1.8 billion from 2022. Intel expects its foundry business to break even by around 2027.
Since April, Intel's stock price has fallen by more than 13%. Foreign media pointed out that it is worth paying attention to whether the new artificial intelligence chip can help Intel find the momentum for the stock price to rise, and on April 9, Intel fell 1.7% at one point. But after the release of the new chip, Intel's stock price began to rise. As of the close of the U.S. stock market on the 9th, Intel closed up 0.92%, with a market value of $163.2 billion.
Will Gaudi 3 be the antidote to Intel's struggles?
After all, AMD also launched the MI300 series of AI chip products in early December 2023, which said that the memory density of the MI300X chip is 2.4 times that of Nvidia H100, and the memory bandwidth is 1.6 times that of Nvidia H100, with better inference performance.
After the release of Intel Gaudi 3, the current AI chip market has begun to show a three-legged situation - NVIDIA B200, AMD MI300 series and Intel Gaudi 3.
In addition, it was also revealed at the meeting that the code name of the next generation of Nvidia Gaudi chips will be Falcon Shores.
It can be seen that the fission of the chip market has begun to intensify differentiation.