With the fifth iteration of its Zen architecture, AMD is focusing on speed and efficiency in a similar footprint. The architecture will initially be used in laptops, but will soon permeate AMD’s entire CPU portfolio. What are the main differences from Zen 4?

At Computex, AMD announced Ryzen 9000 for the desktop and Ryzen AI 300 for the laptop. The laptop chips have hit the market, and the desktop processors are scheduled for mid-August after a delay. New Epyc chips will follow later. All three CPUs are equipped with processing cores based on a new architecture: Zen 5.

(Sometimes) 3 nm

Zen 5 follows Zen 4 and brings innovations across the board. The classic Zen 5 cores are baked on TSMC’s N4X production line, a 4/5nm node that brings some minor improvements compared to N4P, on which the Zen 4 cores are manufactured.

AMD also plans to build Zen 5c cores for cloud-native processors for servers where the number of compute cores is an absolute priority. These cores will be baked on TSMC’s N3 (3 nm) process. The I/O for Zen 5 processors will remain a 6 nm chiplet, as with Zen 4.

AMD says that individual transistors in Zen 5 are smaller and therefore more efficient. This allows for faster chips that generate less heat at higher clock speeds. The production process for Zen 5 will mainly result in modest efficiency gains for the chips, with the exception of Zen 5c, where the 3nm node will theoretically have a larger positive impact. The biggest progress AMD needs to make with Zen 5 itself, and that will be through architectural improvements.

Further instructions

The processor specialist does that too. After all, the IPC (instructions per clock cycle) has improved significantly. In the same cycle, a Zen 5 core can process an average of sixteen percent more instructions than its predecessor. A higher IPC means a faster chip, at least for most applications.

Benchmarks from AMD itself show eleven percent more IPC for Handbrake (video conversion) in the lower range. For Puget Bench (Adobe Première Pro), the IPC gain has already increased to an average of sixteen percent. GeekBench 5.4 (AES-XTS) shows an outlier with a 35 percent higher IPC.

AMD has specifically focused on stronger hardware support for AES-XTS encryption workloads. Machine learning is also said to be significantly better on a Zen 5 processor, with up to 32 percent more performance.

Bandwidth and prediction

The improvements are the result of several innovations. AMD has significantly accelerated the execution and decryption processes. Smaller but relevant steps can be found in the fetching and prediction of instructions and the available bandwidth on the CPU.

To fetch instructions faster, Zen 5 cores use a dual fetch pipe with more accurate branch prediction. In other words, the chip can more quickly predict what the next instruction to be executed will be, even before the previous one has completed.

Latency to the cache has also decreased, while bandwidth has increased. Towards the L1 cache, AMD has doubled the bandwidth compared to Zen 4.

AMD remains true to its chiplet approach. This means that the Zen 5’s processing cores are packaged in different configurations to create end products for a specific target group. The core architecture of the laptop processors is therefore the same as that of the desktop and server.

Strong together

For the new chip generation, not only are the Zen 5 cores being upgraded, the associated graphics chiplet is also being improved. AMD is introducing the RDNA 3.5 architecture for the integrated GPUs. RDNA 3.5 is said to offer significantly more performance per watt. For 15-watt components, AMD measures an improvement in graphics utilization of nineteen to 32 percent. These upgrades also play a role in the overall experience of new chips on desktops and laptops.

Specifically for servers, AMD plans to announce the fifth generation of Epyc with Zen 5 later this year. The top part will be a chip with 192 Zen 5c cores.

In the laptop, AMD combines the Zen 5 cores with an NPU. It can accelerate AI-related workloads and is powerful enough to win Microsoft’s Copilot+ label. In practice, there aren’t many workloads that make the most of the NPU, but those who do a lot of video calling will still notice better video, better sound, and less battery consumption.

AMD will use Zen 5 as the basis for new chips in the coming months and years, although work is also underway on Zen 6. The Zen 5 architecture is a major evolution of Zen 4 that does not break drastically with the previous generation. For example, the Zen 5 desktop processors will still fit easily into motherboards with AM5 sockets.

Why not always small?

It is noticeable that AMD is using 3 nm for the Zen 5c cores, but not for the broader Zen 5 series. This makes it more difficult to achieve large generational leaps, although AMD claims to be successful with this. The availability of the N3 node from TSMC probably plays a role in this choice. The combination with the also renewed NPU and GPU helps to bring a noticeably more powerful overall package to market.

The Zen architecture does not harm AMD. Since the introduction of Zen, the manufacturer has been delivering competitive processors. This fifth generation shows that Zen still has a lot to offer and offers room for innovation, regardless of the production nodes used.