Gaudi2: all about Intel’s new super-powerful AI chip –

Intel has just unveiled its new Gaudi2 AI chip, twice as powerful as the first generation. The firm seeks to compete with Nvidia and AMD on all fronts, and is also developing new GPUs. Find out everything you need to know.

The domain artificial intelligence is booming. In 2022, according to IDC, spending on this technology is expected to increase by 20% to reach $433 billion. the market is growingand computer manufacturers do not want to miss the AI ​​train.

In this context, Intel just unveiled its new chip of artificial intelligence on May 10, 2022 during its Vision conference. The purpose of the firm is in particular to regain market share from NvidiaAMD and other competitors.

Gaudi2 vs. Gaudi1

This new Gaudi2 chip is designed by Habana Labsbased in Israel and acquired by Intel in late 2019 for $2 billion. She is twice as fast than the first version, and should be integrated into the servers by the end of 2022.

L’GEMM front-end architecture (general matrix multiply) by Gaudi1 was back-ended by 10 Tensor Processor Cores (TPCs), but only eight of them were exposed to users. This chip was notably implemented in TSMC’s 16 nanometer process, and offered 24 MB of on-chip SRAM, four banks of HBM2 memory for 32 GB of capacity and 1 TB per second of bandwidth.

For now, Intel did not reveal details concerning Gaudi2 architecture. We know, however, that it will be based on TSMC’s 7-nanometer process and will make it possible to increase to 24 TPC instead of 10. The new 8-bit FP8 data format is supported, such as on the Hopper GH100 GPUs launched in March 2022 by Nvidia.

This new data format allows to have 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.

The Gaudi2 chip embeds 48MB of SRAM. It comes with HBM2e memory strips offering 2.45 TB/sec bandwidth. Their number has not been revealed.

It is equipped with 24 Ethernet ports at 100GB/second, or one for each TPC. The component should plug into a PCI-Express 5.0 port, and will consume 650 watts.

This new chip should therefore offer performance multiplied by 2.5 in comparison with Gaudi1. However, it is not yet known whether any changes in architecture and cadence have been made.

Chips like those in the Gaudi range allow speed up mathematical calculations specific to artificial intelligence. As an alternative, one can also cite the Nvidia H100 designed to support the AI ​​revolution.

They simplify and reduce the cost of training AI modelswhich learn by processing complex real-world data to find patterns.

These components allow in particular improve voice recognition or the autopilot system autonomous vehicles. Intel’s automotive arm, Mobileye, trains its AI systems on Gaudi.

A third generation Gaudi3 chip is already in development and will bring increased performance, more memory and better network capabilities.

GPU vs AI chips

With Gaudi2 and its new GPUs, Intel’s goal is to regain its leadership position of the computer market. Over the past two decades, the firm has gradually lost this status.

Because, the CPUs that made it famous are no longer in the spotlight. Now, GPUs are exploited for artificial intelligence and the main manufacturer of these graphics processing units is Nvidia. This is why the market cap by Nvidia is estimated at $424 billion, more than double Intel’s at $181 billion.

Many manufacturers develop specific AI chips, but Nvidia prefers to continue to focus on GPUs. These components can also be used for supercomputers and HPC. This flexibility is Nvidia’s main selling point.

The businesses love the versatility of GPUs, because it allows them to remain productive in all circumstances and regardless of the evolution of an AI model. General Motors’ autonomous vehicle branch, Cruise, for example, rents Nvidia GPUs on Google Cloud infrastructure to take advantage of their more mature AI software and extreme flexibility.

Likewise, GPUs and their software can quickly adapt to the constant changes driving the AI ​​industry. They can for example adapt to new architectures, new types of layers or to the fusion of AI models.

The AI ​​chip war

Besides Intel, many startups work on specialized AI accelerators. We can mention Graphcore, SambaNova Systems, Tenstorrent or Cerebras. According to the CEO of the latter, GPUs were more suitable than CPUs for AI, but remain too limited compared to dedicated chips.

A battle looms on this new market for the next five years. To get out of the game, Intel could adopt an aggressive pricing strategy.

L’AI is no longer the preserve of giants such as Amazon and Google, and falling cost could open the door to new applications such as fraud detection, crop monitoring or medical image analysis.

AI chips and GPUs: Intel fights on all fronts

However, Intel is betting both on versatile GPUs and AI accelerators specialized. the GPU Ponte Vecchio provides its power to the Aurora supercomputer at Argonne National Laboratory which is expected to go live by the end of 2022.

Thereafter, in 2023, Intel will begin selling the Ponte Vecchio to a wider market. The firm also plans to develop successors to this GPUcheaper and in larger quantities.

The GPU branch is led by Raja Koduriwho previously developed GPUs for AMD and Apple before joining Intel in 2017. He also leads the new range of Arc GPUs dedicated to video games.

The first product in this range, carrying the code name Alchemistis already on sale and new products will be released later this year for laptops and gaming PCs. The Battemage and Celestial successors are being developed on the roadmap that extends to 2025.

In conclusion, Intel fights Nvidia and AMD on all fronts. The American firm wants to place itself as the third player in this lucrative market, and meet the needs of all potential customers…

We would like to say thanks to the author of this short article for this amazing material

Gaudi2: all about Intel’s new super-powerful AI chip –

Discover our social media profiles , as well as the other related pages