Intel unveiled its technological advancements for the future at Intel Architecture Day 2021. In a conference led by Raja Koduri, senior vice president and general manager of the Accelerated Computing Systems and Graphics (AXG) Group, the manufacturer showcased its extensive lineup of new products for all the fields in which Intel operates.
They have shown new architecture x86 CPUs, their first domestic GPUs, the novelties that the Alder Lake series of CPUs and chipsets will bring us and more.
Table of Contents
Intel CPUs with Efficient x86 Core and Performance x86 Core microarchitecture
Looking ahead, your CPUs will have two types of cores. The first model of cores will be used for processes that require efficiency as background processes, while the second model will focus on performance for more complex processes.
Intel Efficient X86 Core Features
- 5000 input branch target cache resulting in more accurate branch prediction.
- 64 kilobyte instruction cache to keep helpful instructions close by without wasting memory
- subsystem power.
- Intel's first on-demand instruction length decoder to generate pre-decode information.
- Intel clustered messy decoder that allows up to six instructions to be decoded per cycle while maintaining power efficiency.
- A spacious back-end with five-width allocation and eight-width retreat, 256 out-of-order input window, and 17 execution ports.
- Strong security features that support Virtualization Technology Redirection Intel Control Flow Application Technology.
- The implementation of AVX ISA, along with new extensions to support Artificial Intelligence operations.
Intel Performance X86 Core Features
- A 19% Geomean improvement across a wide range of workloads over Intel's current 11th Gen.
- CoreTM (Cypress Cove) processor architecture at ISO frequency for general purpose performance.
- Exposure to greater parallelism and an increase in execution parallelism.
- Inte Advanced Matrix Extensions, the next generation embedded AI acceleration breakthrough, for deep learning inference and training performance. Includes dedicated and new hardware.
- Instruction set architecture to perform matrix multiplication operations significantly faster.
- Reduced latency and better compatibility with large data and code applications.
Intel Advanced Matrix Extensions for Data Centers (Intel AMX)
In this section of Intel Architecture Day 2021, they presented new features for their AMX architectures for data centers, which consist of two parts: the blocks and TMUL. The blocks have an expandable 2D log file with eight new iK registers each. They are capable of logging basic data operators.
The blocks declare the state and are managed by the operating system thanks to the XSAVE architecture. The TMUL is a group with a multiplication matrix of instructions that calculates the data using three register blocks. Thus they manage to generate more work per instructions and save power to allocate them to other processes.
Intel Thread Director
Intel Architecture Day 2021 was the occasion for the unveiling of the new Intel Thread Director system . With a precise branch prediction system, it will improve the efficiency of repetitive processes by using predictive techniques to reduce CPU queries and enhance speed.
It has been designed together with Microsoft to be more efficient in Windows 11, so we will see its potential with the new version of Microsoft's operating system that will arrive this Christmas.
Intel Alder Lake Architecture
Looking ahead to the upcoming release of the Alder Lake chipsets, on Intel Architecture Day 2021, they spent some time talking about them. Among its greatest novelties is that there will be models with up to 16 cores (8 for efficiency and 8 for performance), up to 24 threads and up to 30MB of non-inclusive LL cache. They will be the first Intel chipsets to support the upcoming DDR5 RAM technology alongside DDR4, LP5 and LP4. They will have dynamic voltage frequency settings and improved support for overclocking.
Some chipsets will support up to 16 PCIe Gen5 ports and 4 PCIe Gen4 ports. I7O Fabric supports up to 64GB / s, compute fabric up to 1000GB / s, and Memory Fabric up to 204GB / s. They have focused on minimizing electricity consumption, which has achieved a 40% improvement in efficiency compared to Sky Lake. Their interconnectors are designed to make the most of connections with other devices.
Intel XPG GPUs
Intel will enter the home GPU arena starting next year to rival NVIDIA and AMD. Its proposal is the XPG series of GPUs, whose first Alchemist model will arrive in early 2022 and will be manufactured by TSMC. It will be followed by the Battlemage, Clestial and Druid GPUs, the latter being a model with a new Xe architecture.
Intel Xe SS Technology
As part of their strategy for the GPU market, they have created a graphic image rescaling technology that reduces load times by 25%. It is compatible with the Unreal Engine, Vukan, Unity and XII engines, and supports Raytracing as standard. They have worked with APEC so that right now it works with Unreal engine 5. Xe SS technology improves the quality from 1080p to 4k so that next to a native 4K the difference is almost imperceptible. But Intel Xe SS is not something exclusive to Intel GPUs, it will also work on GPus from other brands.
Intel Sapphire Rapids
Sapphire Rapids is Intel's new data center technology, arriving in early 2022. The entire architecture and software have been built from the ground up. It's optimized for microservices, AI with advanced memory, and IoT workloads. It utilizes the Xeon architecture and AMX Acceleration. It features improved microarchitecture and OPC with enhanced support for large data and code footprints.
Some of the improvements to its efficiency core include the use of Intel Advanced Matrix Extensions ( AMX ) for Artificial Intelligence, Accelerator Interface Architecture, FP16, and CLDEMOTE for cache management. Sapphire Rapids will use integrated core accelerators to improve the efficiency of all cores by up to 39 percent.
Sapphire Rapids is comprised of a scalable architecture that enables workload balancing and works as a single unit. Each thread and nucleus has access to all resources to have low latency and reduced energy consumption. Its Intel Quick Assist Technologyy compresses and decompresses operations at a rate of 160 Gb7s, reducing workloads by up to 98%.
Intel Infrastructure Processing Unit
Intel will begin manufacturing processing hardware designed for cloud service infrastructures that they have developed from the ground up. These units will be called Infrastructure Processing Units (IPUs ). They are designed for cloud services, allowing the separation of infrastructure and the Root of Trust from the CPU. Their accelerators help process tasks efficiently and simplify the architecture of any data center, eliminating the need for disks. For example, providing this IPU support to Facebook improved its efficiency by between 31 and 81 percent.
Intel is manufacturing ASIC IPUs with improved optimization and performance, and improved networkind security. Alongside this, it will drive FPGA-based acceleration with software such as Intel Agilex, Intel Stratix, Intel Xeon, and Intel Ethernet.
Oak Springs Canyon is another of its IPUs, designed for high-level networking and storage acceleration for cloud service providers. On the other hand, the Arrow Creek IPU with Intel Agilex FPGA and Ethernet E810 will have two QSFP56 with support for two 1 GbE, 16 PCIe Gen4, 16 Gb of DDR4 memory with 1 GB of DDR4 destined for the integrated CPU.
Mount Evans will be its high-end 200G IPU with a subsystem network and compute complex. The former provides support for four 200 GB/s Xeons, ROCEv2, NVMe, a programmable package with QoS and telemetry capabilities, and intraline IPSec. Its compute complex supports 16 Arm neoverse cores, has a triple-channel cache, crypto and compression systems, and a dedicated management processor.
Intel Ponte Vecchio
The last of Intel's announcements at Intel Architecture Day 2021 was Ponte Vecchio , the codename for a high-end GPU designed for AI professionals, with no plans for a consumer market at this time. It comprises over 100.000.000.000 billion transistors, 47 active blocks, and five processing nodes. It features a System-on-a-Chip (SoC) with a collection of interconnected chips that function as a unified whole. Each of these blocks contains eight Xe cores. It will be manufactured by TSMC using their 5-nanometer technology.
It is designed to withstand high workloads even with Artificial Intelligence. It reaches 3400 images per second. It features Advanced Raytracing technology to take visual ability to new levels. All of Intel's own technologies are embedded.
Intel one API
oneAPI is a unified API protocol so developers don't have to tie themselves to a single resource library. This reaches low-level libraries, languages, middleware, frameworks and communication with CPUs and GPUs. More than 200.000 developers have already signed up to use it.