Intel has just unveiled its new Gaudi2 AI chip, which is twice as powerful as the first generation. The firm wants to compete with Nvidia and AMD on all fronts and is building new GPUs. Find out what you need to know.
The field of artificial intelligence is evolving. In 2022, according to IDC, costs on this technology are expected to increase by 20% to reach $ 433 billion. The The market is growingAnd computer makers don’t want to miss the AI train.
In this context, Intel has just unveiled its new chip During the May 10, 2022 Vision Conference on Artificial Intelligence. The purpose of the firm in particular Restore market share from NvidiaAMD and other competitors.
Gaudi2 vs. Gaudi1
This is the new Gaudi2 chip Designed by Habana Lab, Located in Israel and acquired by Intel at the end of 2019 for 2 billion. She Twice as fast Than the first version, and should be integrated into the server by the end of 2022.
LawGEMM front-end architecture (Common matrix multiplication) was back-end by Gaudi1 10 Tensor Processor Core (TPC), But only eight of them have been exposed to users. This chip was significantly applied to TSMC’s 16 nanometer process, and offers 24 MB of on-chip SRAM, four banks of HBM2 memory for 32 GB capacity, and 1 TB of bandwidth per second.
For now, Intel did not release details Gaudi2 relates to architecture. However, we know that it will be based on the 7-nanometer process of TSMC and it will be possible to increase it to 24 TPC instead of 10. New 8-bit FP8 data format Supported, such as the Hopper GH100 GPU launched by Nvidia in March 2022.
This New data format Allows you to keep both low resolution inference data and high definition training data in the same format without having to convert the models when going from training to inference.
Embed the Gaudi2 chip 48MB SRAM. It comes with an HBM2e memory strip that provides 2.45 TB / sec bandwidth. Their numbers were not disclosed.
It is equipped with 24 Ethernet ports per 100GB / sec, Or one for each TPC. The component should be plugged into a PCI-Express 5.0 port and will use 650 watts.
This new chip should therefore be offered Performance multiplied by 2.5 Comparison with Gaudi1. However, it is unknown at this time what he will do after leaving the post.
Allows chips like the Gaud range Speed up mathematical calculations Specific to artificial intelligence. As an alternative, we can cite the Nvidia H100 designed with the AI revolution in mind.
They simplify and reduce Cost of AI model trainingWhich learns through complex real-world data processing to find patterns.
These elements specifically allow Improve voice recognition Or autopilot system of autonomous vehicles. Intel’s automated arm, Mobileye, trains its AI systems at Gaud.
A Third generation Gaudi3 chip Is already under development and will bring enhanced performance, more memory and better network capacity.
GPU vs. AI chips
With Gaudi2 and its new GPUs, Intel’s goal Restore his leadership position Computer market. Over the past two decades, the firm has gradually lost this status.
Because, The CPU that made it famous No more spotlights. Now, GPUs are used for artificial intelligence and the main manufacturer of these graphics processing units is Nvidia. That’s why Nvidia’s market cap is estimated at $ 424 billion, more than double Intel’s $ 181 billion.
Many manufacturers make specific AI chips, but prefer to continue to focus on the Nvidia GPU. These components can also be used for supercomputers and HPCs. This flexibility is Nvidia’s main selling point.
The Business prefers the versatility of GPUs, Because it allows them to remain productive in all situations regardless of the evolution of the AI model. General Motors’ autonomous vehicle business, Cruise, for example, leases Nvidia GPUs in the Google Cloud infrastructure to take advantage of their more mature AI software and extreme flexibility.
Similarly, GPUs and their software can speed up Drive the AI industry to adapt to constant change. They can adapt, for example, to new architectures, new types of layers, or fusion of AI models.
AI chip warfare
In addition to Intel, Lots of startups Work on special AI accelerators. We can mention Graphcore, SambaNova system, Tenstorrent or Cerebras. According to the next CEO, GPUs were more suitable for AI than CPUs, but remained Very limited compared to dedicated chips.
A War looms in this new market For the next five years. To get out of the game, Intel may adopt an aggressive pricing strategy.
LawAI is no longer the preserve of giants Such as Amazon and Google, and low-cost new applications such as fraud detection, crop monitoring, or medical image analysis.
AI chips and GPU: Intel fights on all fronts
However, Intel is betting Versatile GPU and AI accelerator both Specialists. The GPU Ponte Vechio At the Argonne National Laboratory, Aurora delivers its power to the supercomputer, which is expected to go live by the end of 2022.
Then, in 2023, Intel will start selling Ponte Vecchio in a wide market. The company is also planning development Successor to this GPUCheap and in large quantities.
The The GPU branch is headed by Raja KoduriWho built the GPU for AMD and Apple before joining Intel in 2017. He also led New range of Arc GPUs Dedicated to video games.
The first product in this range, bearing Code name AlchemistAlready on sale and new products for laptops and gaming PCs will be released later this year. Batmeze and Celestial successor Creating an expanded roadmap until 2025.
In conclusion, Intel fights Nvidia and AMD on all fronts. American firm wants to position itself as the third player in this lucrative market and meet the needs of all potential customers …