At the Taipei Computer Show in early June, AMD officially released the Ryzen AI 300 series notebook processors and Ryzen 9000 series desktop processors based on the new generation Zen 5 architecture, both of which will be available in July.
At the time, AMD was relatively rare in announcing the specific model names and specifications (Lunar Lake next door only had architecture and technology), but it did not say much about the new CPU/GPU/NPU architecture, and only gave a rough idea of the new generation of motherboard chipsets on the desktop, leaving some suspense.
Last week, AMD held a Zen 5 Tech Day technology event in the United States, finally unveiling the mystery of the new architecture and announcing more performance data, technical features, and especially new processor and memory overclocking.
As for the exact launch time, the Ryzen AI 300 notebook will be launched on July 28, and the Ryzen 9000 series will be launched on July 31.
[Zen 5 architecture: overall width and size expansion, exclusive Intel secrets]
The Zen architecture was born in 2017. Compared with the previous Bulldozer architecture, the IPC has increased by 52%, far exceeding the original target of 40%. This is an extremely rare improvement in the entire history of microprocessors.
Over the past seven years, the Zen architecture has been continuously refined and has now evolved to the fifth generation. It is a considerable change, including increasing the number of instructions per clock cycle, expanding instruction dispatch and execution bandwidth, doubling cache data bandwidth, AI acceleration, and more.
CPU architecture design is an extremely complex project, even if it is an upgraded version.
Mike Clark has worked at AMD for 31 years. He is now an AMD academician and chip design engineer. He is also the soul figure in the research and development of Zen architecture over the years. He is known as the “Father of Zen” and “Zen Daddy”.
He expressed emotionally,People often don't realize how difficult CPU architecture design is and how long it takes. For example, the research and development of Zen 5 is the result of many years of hard work by a large number of design, verification, software and other teams in many parts of the world. It has even been integrated into everyone's blood. Many people think about it while eating or even dreaming, and it is quite a remarkable thing to finally see their hard work bear fruit.
In fact, modern CPU architectures have mature systems, which can be divided into two major parts: the front-end and the back-end. The subdivisions include different unit modules such as instruction prefetching and decoding, integer execution, floating-point execution, loading and storage, and cache.
Unless a completely disruptive computing system emerges, what CPU architecture designers have to do is to determine the specifications and scales of different unit modules based on preset goals, and then organically combine them into a whole to maximize efficiency without causing waste or bottlenecks.
Just like a canal or a water pipe, generally speaking the greater the water flow, the better. But you can't just keep widening and thickening it. There needs to be overall coordination from the source to the end. There shouldn't be insufficient water, and the water can't be blocked. What matters is a balance.
The overall idea of Zen 5 is to appropriately scale up, even doubling in many placesFor example, the front end has been changed to a dual pre-fetch and dual decoding pipeline.it can handle various loads more efficiently. For example, the sluice gate at the source is more open and can release more water flow.
at the same time,Branch prediction has also been greatly improved, with higher throughput, higher accuracy and lower latencyand the latency and bandwidth of the instruction cache have also been improved, just like the canal has become wider and will not overflow when facing more water flow.
Integer execution units widen the pipeline for dispatching and executing instructionsincluding dispatch and retirement increased to 8 widths, larger execution windows, and more integrated ALU schedulers, including six ALUs and four AGUs.
The most important change in the floating point and vector execution units isBased on the AVX-512 instruction set introduced in the Zen 4 architecture, it has been enhanced from supporting only 256-bit data width to supporting the full 512-bit data width.
The 256-bit version has greater flexibility because not all instructions require a 512-bit width. Now with the addition of the 512-bit full-blood version, it can be longer or shorter, which can greatly improve execution efficiency and performance while maintaining flexibility.
What is even more intriguing is that due to Intel's heterogeneous hybrid architecture design, the new generation of Lunar Lake and Arrow Lake do not support AVX-512, and it is uncertain when they will return.
Who would have thought that Intel's unique weapon back then has now become AMD's private dish.
In addition, the floating-point unit has 6 pipelines, and the delay of the FADD instruction is only 2 clock cycles.
The load storage unit part greatly improves the data bandwidth, includingThe L1 data cache capacity has been increased from 32KB to 48KB, and has been enhanced from 8-way to 12-way.
The maximum bandwidth to the first-level cache and floating-point unit has also doubled compared to the previous generation, and the corresponding data prefetching has been improved.
in addition,The performance of the math acceleration unit has made great progress, and the speed of single-core execution of math learning and AES-XTS instructions can be increased by up to 32% and 35% respectively.
What are the benefits? Of course, it can greatly accelerate the efficiency of AI computing and be more adaptable to the current environment, especially when paired with the Instinct accelerator in the EPYC processor, the effect is even more complementary.
With this series of improvements,The IPC improvement of the Zen 5 architecture is up to 16% on average (which can be understood as performance improvement at the same frequency), and in some scenarios the improvement is even as high as 35%.
The biggest contributions come from improvements in instruction execution and retirement, followed by improvements in data bandwidth, instruction decoding and OP cache, and finally changes in instruction prefetching and branch prediction.
According to official data from previous years, the average IPC improvements of Zen+, Zen 2, Zen 3, and Zen 4 compared to their previous generations were 3%, 15%, 19%, and 13%, respectively.
After five generations and six versions of evolution, the average IPC improvement of Zen 5 compared to the first generation Zen has reached about 85%!
Don’t forget that the frequency is also constantly increasing. The first generation had a maximum of only 4.0GHz, but now it has reached 5.7GHz, an increase of about 43%.
In terms of products, the Ryzen 9000 series for desktop computers will use pure Zen 5 or a combination of CCD+IOD.
The CCD process was upgraded from 5nm to 4nmeach with up to 8 cores, and a total of up to 16 cores.
The IOD part uses the Ryzen 7000 series, so it is still 6nm and integrates two RDNA 2 architecture GPU graphics cores.
The Ryzen AI 300 series for mobile devices are all combinations of Zen 5 and Zen 5c. As shown in the lower right corner of the above picture, the four orange ones at the bottom are Zen 5, and the eight purple-red ones at the top are Zen 5c.
Unlike Intel's heterogeneous architecture, Zen 5 and Zen 5c still have exactly the same architectural design, IPC performance, and ISA instruction set. The only difference is that the latter has a smaller cache and lower frequency (but higher energy efficiency).
The manufacturing process of the Ryzen AI 300 series is 4nmthe same as the previous generation Ryzen 7000/8000 series.
In the data center,The fifth-generation Turin EPYC will be released in the second half of this year, using an advanced 4nm and 3nm process combination. This is also the first time AMD has introduced 3nm.
Turin EPYC will be upgraded to up to192 cores and 384 threadsthe official specifically mentioned the new featuresAI encryption based on Trust IO functionwhich can undoubtedly better meet today's cloud AI deployment needs.
Next, we will see Zen 6 and Zen 6c. The subsequent Zen 7 is also under development. Guess what process will be used?












