Advanced Micro Devices Inc. today detailed the Ryzen 9 9950X, a central processing unit for consumer-grade desktops that the company touts as the fastest chip in its category.

The processor debuted early today at the Computex hardware event taking place in Taiwan. Alongside the Ryzen 9 9950X, AMD debuted two laptop chips optimized for artificial intelligence applications. The data center AI market was also a major theme: the company detailed three upcoming machine learning accelerators that are expected to bring significant performance upgrades.

Faster desktops

AMD’s new flagship consumer desktop processor, the Ryzen 9 9950X, operates at a 4.3-gigahertz base speed that it can boost to 5.7 GHz if necessary. The chip promises to significantly outperform comparable Intel Corp. silicon. According to AMD, the Ryzen 9 9950X completed the popular Blender benchmark test 56% faster than Intel’s Core i9-14900K desktop chip in an internal test. 

The new CPU comprises two silicon dies with eight cores each. A third die holds the circuitry that manages data input and output operations, as well as the attached memory. That memory takes the form of a 80-megabyte cache pool.

According to AMD, the Ryzen 9 9950X is a four-nanometer implementation of a new core design known as Zen 5. The architecture can perform 16% more instructions per cycle, a core measure of CPU speed, than its predecessor. It also provides twice the AI performance.

CPUs use a technique called branch prediction to speed up applications. The technique allows a chip to estimate what calculations will be performed by a program in the near future and complete them in advance. Those predictions are sometimes incorrect: A program may not need the results of the calculations that the CPU completed ahead of time, in which case those results are simply discarded without providing a speed boost.

One reason AMD’s new Zen 5 architecture is faster than its predecessor is that it features a more accurate branch prediction mechanism. As a result, the results of calculations that are performed in advance have to be discarded less often, which improves application performance. Zen 5 also features a larger out-of-order instruction window size, which means that it can perform more calculations in parallel.

The Ryzen 9 9950X is part of a new desktop CPU family, the Ryzen 9000, that also includes three other chips. All three are based on the Zen 5 architecture. Their core counts range from 6 to 12, while top speeds peak at 5.6 GHz.

AI-optimized laptops

The Ryzen 9000 product line made its Computex debut alongside a second new AMD chip series, the Ryzen AI 300 Series, that comprises two processors. Both target the laptop market and feature an onboard neural processing unit, or NPU, designed to speed up AI-powered applications.

The faster of the two chips is known as the Ryzen AI 9 HX. Its onboard NPU has a top speed of 50 TOPs, or 50 trillion computations per second. To help applications the most out of that AI-optimized circuitry, AMD has equipped the CPU with support for an AI optimization technology called Block FP16.

Machine learning models represent knowledge using units of data called floating point values. Those units of data comprise two elements known as the significand and exponent. Typically, each floating point value has one significand and one exponent. 

According to AMD, its Block FP16 technology reuses the same exponent across multiple floating point values instead of assigning a dedicated exponent to each one. This lowers the total number of exponents in the dataset that an AI model processes, which reduces the dataset’s overall size. The end result is that the chip on which the AI model runs has to process less data, which speeds up calculations.

Alongside its NPU, the Ryzen AI 9 HX features an integrated graphics processing unit and a dozen Zen 5 CPU cores. Those cores can more than double their 2 GHz base speed to 5.1 GHz when operating at full capacity. The second laptop chip that AMD detailed today, the AI 9 365, features an identical NPU and a slightly slower onboard CPU with 10 cores.

Server speedup

AMD will use Zen 5 as the basis of not only its consumer processors but also an upcoming server CPU line known as Epyc Turin. Set to roll out in the second half of the year, the latter product line will comprise two processor collections. The first collection features CPUs with up to 128 cores, while the other offers a maximum core count of 192.

Both types of Epyc Turin chips will be made using a three-nanometer manufacturing process. According to AMD, some of the processors will comprise more than a dozen individual dies. The Epyc Turin line promises to provide up to 5.4 times the AI performance of comparable Xeon server chips from Intel. 

New AI accelerator roadmap

AMD’s Computex presentation also covered other topics besides CPUs. The company previewed the Instinct MI325X, an upgraded iteration of its existing MI300X machine learning accelerator that will arrive in the fourth quarter. AMD modified the original design by adding a new, faster type of memory known as HBM3E.

The upgrade boosts the MI325X’s cache capacity to 288 gigabytes from 192 gigabytes before. Its memory bandwidth, a metric that influences how fast AI models can shuffle data in and out of cache, is now 6 terabits per second instead of 5.3 terabits. Memory bandwidth is an important performance metric because neural networks move data to and from memory more frequently than many other types of workloads.

AMD detailed that it plans to start releasing a new AI accelerator every year after the MI325X. The chip’s successor, the MI350, will arrive in 2025 and is expected to run AI inference workloads up to 35 times faster. AMD plans to launch an even faster accelerator known as the MI400 in 2026.

Image: AMD