• The Ryzen 9 9950X, AMD’s latest flagship consumer desktop processor, functions at a base speed of 4.3 gigahertz, with the ability to boost up to 5.7 GHz when required.
  • AMD’s improved Zen 5 architecture surpasses its predecessor in speed due in part to a more precise branch prediction mechanism.

Recently, Advanced Micro Devices Inc. (AMD) unveiled the Ryzen 9 9950X, a consumer-grade desktop CPU that the company claims is the fastest in its class.

The processor was launched recently at the Computex hardware event in Taiwan. In addition to the Ryzen 9 9950X, AMD introduced two laptop chips designed for artificial intelligence applications. The data center AI market was another key focus, with the company unveiling three forthcoming machine learning accelerators anticipated to deliver substantial performance enhancements.

Speedier Desktops

AMD’s latest flagship consumer desktop processor, the Ryzen 9 9950X, features a base speed of 4.3 gigahertz, with the capability to boost up to 5.7 GHz when needed. This chip is projected to surpass the performance of comparable Intel Corp. processors significantly. In an internal test, AMD reported that the Ryzen 9 9950X completed the popular Blender benchmark test 56% faster than Intel’s Core i9-14900K desktop chip.

The new CPU consists of two silicon dies, each containing eight cores. A third die houses the circuitry responsible for managing data input and output operations, along with the attached memory. This memory is organized into an 80-megabyte cache pool.

According to AMD, the Ryzen 9 9950X is a four-nanometer implementation of the new Zen 5 core design. This architecture can execute 16% more instructions per cycle, a key metric of CPU speed, compared to its predecessor. Additionally, it offers double the AI performance.

CPUs employ a technique called branch prediction to accelerate application performance. This method enables a chip to anticipate the calculations a program will soon require and execute them in advance. However, these predictions are not always accurate: if a program does not need the precomputed results, those results are discarded, offering no speed advantage.

One reason AMD’s new Zen 5 architecture outperforms its predecessor is its more accurate branch prediction mechanism. This accuracy reduces the need to discard precomputed results, thereby enhancing application performance. Additionally, Zen 5 features a larger out-of-order instruction window, allowing it to perform more calculations in parallel.

The Ryzen 9 9950X is part of the new Ryzen 9000 desktop CPU family, which includes three other chips. All of these processors are based on the Zen 5 architecture, with core counts ranging from 6 to 12 and top speeds reaching up to 5.6 GHz.

AI-tailored Laptops

The Ryzen 9000 product lineup premiered at Computex, introducing another new series from AMD called the Ryzen AI 300 Series. This series includes two processors tailored for the laptop market, equipped with an onboard neural processing unit (NPU) aimed at accelerating AI-driven applications.

The swifter of the duo is dubbed the Ryzen AI 9 HX. Its onboard NPU boasts a peak speed of 50 TOPs, equivalent to 50 trillion computations per second. To maximize the potential of this AI-optimized circuitry, AMD has outfitted the CPU with support for a technology known as Block FP16, aimed at enhancing AI optimization for applications.

Machine learning models encode knowledge using data units referred to as floating-point values. These values consist of two components: the significand and exponent. Normally, each floating-point value consists of one significand and one exponent.

As per AMD, its Block FP16 technology utilizes a single exponent across multiple floating-point values instead of assigning a separate exponent to each. This decreases the total number of exponents in the dataset processed by an AI model, thus reducing the dataset’s size. Consequently, the chip executing the AI model has to handle less data, resulting in faster calculations.

In addition to its NPU, the Ryzen AI 9 HX incorporates an integrated graphics processing unit and twelve Zen 5 CPU cores. These cores can surpass their 2 GHz base speed to reach 5.1 GHz when fully utilized. The other laptop chip unveiled by AMD recently, the AI 9 365, shares the same NPU but comes with a slightly slower onboard CPU featuring ten cores.

Enhanced Server Performance

AMD plans to leverage Zen 5 not only for its consumer processors but also for an upcoming server CPU line called Epyc Turin. Expected to launch in the latter half of the year, this product line will consist of two processor series. The first series includes CPUs with a maximum of 128 cores, while the second series offers up to 192 cores.

Both variants of Epyc Turin chips will be manufactured using a three-nanometer process. AMD has indicated that certain processors will incorporate over a dozen individual dies. The Epyc Turin series is expected to deliver up to 5.4 times the AI performance of equivalent Xeon server chips from Intel.

Updated AI Accelerator Plan

In addition to CPUs, AMD’s presentation at Computex touched on various other subjects. The company provided a glimpse of the Instinct MI325X, an enhanced version of its current MI300X machine learning accelerator slated for release in the fourth quarter. AMD enhanced the original design by incorporating a faster memory type called HBM3E.

The enhancement increases the MI325X’s cache capacity from 192 gigabytes to 288 gigabytes. Its memory bandwidth, which affects the speed at which AI models can transfer data to and from the cache, has been elevated to 6 terabits per second from 5.3 terabits. Memory bandwidth is a critical performance metric as neural networks often shuttle data between memory and processing units more frequently than many other types of workloads.

AMD announced its intention to introduce a new AI accelerator annually following the MI325X. The next iteration, the MI350, is slated for release in 2025 and is projected to enhance AI inference workloads by up to 35 times. Subsequently, AMD aims to unveil an even swifter accelerator, the MI400, in 2026.