Intel has taken advantage of the unique scenario of this year’s Hot Chips to give us some first glimpses of the future of the Xeon family, and also of the important commitment that the company has made by specializationnot only through the accelerators built into that family, but also through the use of high-performance P cores and high-efficiency E cores.
The chip giant has been able to adapt its strategy based on three big keys: the current market situation, the needs of your customers and the reality of different workloads. In this sense, it is important to highlight that the workloads have become highly specialized, and that we can differentiate them into general, intensive and AI-focused, and also into high-density and scale-out tasks, that is, of increasing intensity. .
Intel has known how to face the challenge represented by these three keys, and has done so based on a simple but clear approach, offer powerful solutions and optimized for high performance computing and solutions focused on efficiency for loads of high density and increasing intensity.
The new Intel Xeon family uses a modular SoC type design, in which the computing blocks and the I/O block are in chiplets different with their own encapsulations, although in the end all of them are integrated into the same package. This chiplet-centric approach facilitates both design and execution on the wafer, improves success rate, reduces costs, and at the same time drives the scalability and flexibility of the Xeon series. It is a totally winning bet that also allows the integration of different computing elements, such as specialized accelerators, in the same package.
Intel has also been able to improve its Xeon platform by offering configurations of between one and eight sockets in its range of CPUs with P cores, and will offer between one and two sockets with its line of CPUs with E cores. Other important innovations that we must highlight are the support for memory configurations of up to 12 channels, the 136 PCIe Gen5 lines, the support of the CXL 2.0 standard and up to 6 UPI links (144 lines.
A look at the Intel Xeon Granite Rapids and Sierra Forest
The chip giant gave us an interesting preview of what to expect from the new Xeon Granite Rapids and Sierra Fores processors. The first will use P nucleiwhile the seconds will mount E nuclei.
Intel Xeon Granite Rapids will be a family of CPUs with high performance cores that will maximize performance per core, offer high IPC, have efficiency improvements, and will be built on the 3nm node. These new processors will have a AMX deep learning and AI acceleration engine that it will be able to work in FP16 (single precision), it will have improvements at the encryption level and other changes at the microarchitecture level that will significantly improve performance compared to the previous generation.
Intel Sierra Forest processors will use high efficiency cores, which means that they will be optimized to offer good performance and a high density of cores and threads with low consumption. They will be built on the Intel 3 node, they will offer software B16 and FP16 conversion, they will be able to work with a wide variety of instructions and with INT8 models. These new high-efficiency cores have also received microarchitecture improvements, with each block of two or four cores featuring 4MB of L2 cache.
According to Intel, the performance improvement in tasks related to AI that the P cores of the Xeon Granite Rapids will offer will be up to triple compared to the previous generation, the memory bandwidth will improve by 2.8 timeswe will have higher core scaling, higher efficiency thanks to Intel node 3 enhancements, and major advances in AI and inference thanks to FP16 support.
With Sierra Forest you will increase density per rack by 250% and performance per watt by 240%. The performance and consumption values will translate into high efficiency, and these CPUs will give their best in environments where the really important thing is the high degree of parallelismthat is, having a high number of cores and threads, more than the raw performance per core.
Both processors will arrive next year, and will share a common platform both at the hardware and software level. The first to arrive are expected to be the Intel Xeon Sierra Forest in the first half of 2024, with the Xeon Granite Rapids arriving soon after, which will be the successor to the current Sapphire Rapids. In the last image we can see the full intel roadmap for the next two years, and other important products appear in it within its range of GPU solutions, hardware specialized in AI and FPGAs.