The digitization race to dominate the market and provide the best resources in terms of artificial intelligence has led to Intel to announce the launch of Falcon Shores. It’s about your new 288 GB memory chip able to admit 8-bit floating point computing. In this way it will be possible to provide power and performance to artificial intelligence applications and services such as ChatGPT.
Intel does not want to be left behind compared to its main competitors, which is why it estimates that by 2025 your new AI chip will have burst into the market with force and posed an iron threat to NVIDIA, which currently leads the market for AI chips. To this is added Advance Micro Devices (AMD)which has recently released a new AI-focused chip dubbed MI300.
But once again, and just as it happened with the chip Ponte Vecchio, it is believed that Intel will be late in the race and will be left with little share in the AI market, which continues to evolve at a frantic pace. When the Falcon Shores is available to users, it is possible that NVIDIA has already developed another chip (Grace Hopper), which will again mean a slowdown for Intel in this regard. AMD is also expected to have released by then Instinct MI 300.
In fact, NVIDIA currently leads the production of chips for the AI industry with the A100, used to train extensive language models (LLM). There are companies like Satability AI that use up to 5,400 A100 for Stable Diffusion.
Nevertheless, Jeff McVeighcorporate vice president of Intel’s supercomputing group, says Intel has taken its time to rework the chip and transform the initial strategy combining graphics processing units (GPUs) and central processing units (CPUs).
All this is added to the fact that the expansion of generative AI is drastically changing the requirements demanded, as well as the methods of combining both cores. Intel’s goal is none other than maximize return on investmentmaking Ponte Vecchio the only GPU for high-performance computers until 2025.
A few days ago, Intel announced that it had completed shipments of its supercomputer Aurora from Argonne National Laboratorywhich is based on Ponte Vecchio and has better performance than NVIDIA’s latest AI chip.
The Keys to Falcon Shores
The Falcon Shores, presented in the framework of the ISC conference of high-performance computers held in Germany, is characterized by having 288 GB memory HBM3with bandwidth up to 9.8 TB/s and for being able to admit such varied types of data that they oscillate from FP64 to BFLOAT16 or FP8. It also supports Compute Express Link (CXL)which allows sharing resources for higher and more efficient performance.
Falcon Shores is based on a tiled modular architecture, allowing you to incorporate and enhance discrete CPU and GPU configurations. Likewise, it will be part of Maxa family of high-performance computer (HPC) chips.
Now Intel outlines 2025 as the launch date, although Falcon Shores will be the successor to Rialto Bridgea GPU with which it intended to rival NVIDIA, the successor to Ponte Vecchio and which was expected to be launched in mid-2023. However, the strategy has changed and Falcon will arrive before Rialto.
Falcon Shores, without CPU, will keep the chiplet disaggregated design like its predecessors and will incorporate CXL interfaces. This involves launching a SMX version for systems that use another CPU and also implement a PCI Express version.