AMD Ryzen 9000 Deep Dive – What does Zen 5 deliver?

AMD’s Ryzen 9000 series of Zen 5 processors is launching this month – Here’s what you need to know

The next generation of AMD Ryzen CPUs is launching this month! Ryzen 9000 is coming, boasting new Zen 5 cores, major architectural improvements, and leadership efficiency levels. Ahead of launch, AMD invited us to their 2024 tech day to talk everything Zen, be it Ryzen 9000, Ryzen AI, XDNA, or RDNA. Today, we are starting our coverage with Zen 5 and Ryzen 9000.

On paper, AMD’s Ryzen 9000 CPU lineup looks a lot like their current-gen lineup. We have the same CPU core/thread counts as before and a familiar naming scheme. Thankfully, a lot has changed architecturally to make AMD’s Ryzen 9000 series special. Two words sum up AMD’s primary changes with Zen 5. Those words are performance, and efficiency.

AMD Tech Day 2024

AMD’s tech day gave us the opportunity to discuss the companies latest products with their executives and engineers. No matter what question we had about Zen 5, AMD were ready to answer it. Yes, trade secrets are obviously off the table here, but AMD were willing to discuss Zen 5 at a deep level. We’ll try to summarise what we head there as best we can.

Zen 5 – A new foundation for AMD

AMD described Zen 5 as their new “Foundation” architecture. AMD has re-architected a lot of their core “Zen” design, and Zen 5 will act as a new baseline from which future AMD CPU architectures will be derived. As always, AMD has teams working on future Zen CPU designs, and it is clear to us that AMD has big changes in store for Zen 6/7.

One of the primary goals of AMD’s current-generation Zen 4 CPU architecture was to boost CPU clock speeds. With Zen 5, AMD’s goals were to boost IPC (instructions per cycle), deliver a wider core design, and to better feed their CPU cores.

Zen 5 Deep Dive

During their tech day, AMD gave us a lot more in-depth information than they did at Computex. At Computex we got a glimpse of where Zen 5’s performance gains come from, but at AMD’s tech day they gave us a lot more detail.

With Zen 5, a lot of AMD’s changes focus on latency and bandwidth. AMD’s new branch predictor promises to be less latent and more accurate than before. Furthermore, it also promises to deliver more throughput. All of these aspects are important. Lowered latency allows the rest of AMD’s CPU to access and start working on the branch predictor’s data faster. Increased accuracy results in fewer misses, leading to fewer wasted CPU resources. AMD’s wider Zen 5 core design needs to be fed with data, making AMD’s increased branch predictor throughput important. This is also why AMD’s additional decode pipeline is important.

With Zen 5, AMD now has an 8-wide dispatch. Previously, AMD Zen CPU cores had a 6-wide dispatch. This widening allows AMD’s Zen 5 CPU cores to do more work, but only if fed enough data to work effectively. This change increases the parallelism of AMD’s Zen 5 CPU cores.

This widened design for Zen 5 is likely why AMD sees this as a new “foundational” CPU architecture. After all, AMD’s Zen CPUs have had a 6-wide dispatch since Zen 1. I expect a lot of the changes in AMD’s future Zen architectures to focus on getting the most out of this new broadened core design.

With Zen 5’s wider pipelines, AMD needed to make changes that would keep these new pipelines fed with data. After all, a wider data path would be useless if you couldn’t keep it active. With Zen 5, AMD has managed to greatly increase L1 and L2 cache bandwidth. Additionally, AMD has managed to increase their L1 cache size by 50% without increasing cache latency. Typically, larger caches come with latency penalties, but not with Zen 5.

AVX-512 improvements

Zen 4 added AVX-512 support to AMD’s Ryzen lineup. This supported was added by double-pumping AMD’s 256-bit data path. With Zen 5, AMD now has a full 512-bit data path, allowing AVX-512 to run at full speed. Beyond that, AVX-512 computation is less latent. It now takes 2 cycles of latency instead of three on Zen 4.

With Zen 5, AMD has dramatically increased their AVX-512 performance. This is great for any workload that can make use of this feature. AMD has claimed that AVX-512 is important for AI workloads, making this improvement very important for the company’s broader AI ambitions.

Zen 5/Ryzen 9000 IPC gains

On average, AMD has claimed a 16% IPC uplift for Zen 5. Remember that IPC changes are workload dependent, as some workloads will benefit more from specific architectural changes than others. On the low-end, AMD delivers a 10% IPC increase in Far Cry 6, and on the high-end Zen 5 delivers a 35% increase in AES-XTS workloads.

Overall, AMD’s IPC gains come from a variety of sources. No single change is responsible for AMD’s performance gains with Zen 5. This is part of the reason why AMD has delivered performance gains across such a broad range of applications. AMD’s 16% average IPC gains come from a range of changes, so it make sense for these changes to deliver performance gains across a wide range or workloads.

Below, AMD has charted the impact of Zen 5’s major changes. Overall, AMD’s changes to data bandwidth, execution/retire, and decode/opcache deliver very balanced performance impacts. Fetch/Branch Prediction have a lesser impact, though they are responsible for significant performance gains.

It’s good at math

One of the most impactful changes to AMD’s Zen 5 architecture is its improved Math Acceleration Unit. This is what is responsible for AMD’s largest performance gains with Zen 5. AMD notes and up to 32% improvement in Single-Core Machine Learning performance, and up to 35% gains in AES-XTS workloads.

This change will make AMD’s Ryzen 9000 series CPUs a huge upgrade for some users. If you workload benefits from this change, Zen 5 will be a must have upgrade for you. If not, well… Zen 5 still delivers significant IPC gains.

Zen 5/Ryzen 9000 Performance Rundown

When discussing the performance of their Ryzen 9000 series CPUs, we will note that AMD did not directly compare their Ryzen 9 9950X to Intel’s i9-14900K. AMD only compared Intel’s i9-14900K to their Ryzen 9 9900X. In AMD’s eyes, their Ryzen 9 9950X is in a class of its own, and they only need their 12-core model to beat Intel’s best Raptor Lake CPU.

Below, AMD claims that their Ryzen 9 9900X is 2-41% faster than an i9-14900K in productivity workloads, and 4-22% faster in games. Intel’s i9-14900K may have 24 total cores (8 P-Cores + 16 E-Cores), but AMD claims to beat them with twelve Zen 5 cores.

Similarly, AMD claims performance leadership over Intel’s 20-core i7-14700K with their 8-core Ryzen 7 9700X. With these results, AMD is showing that fewer, stronger, cores can be better than having more cores. Beyond that, AMD is showcasing higher efficiency levels, as their Ryzen 7 9700X has a 65W TDP.

With their Ryzen 5 9600X, AMD is also claiming performance leadership over Intel’s i5-14600K. With just six cores, AMD claims to best Intel’s 14-core (6 P-cores + 8 E-cores) 14600K. Another strong result for AMD.

Better than V-Cache?

When walking into AMD’s event, one question on our mind was “what about V-Cache”. Yes, AMD has plans to release X3D series Zen 5 CPUs, but even without a 3D V-Cache upgrade, AMD has their older X3D CPU beat.

Below, we can see how AMD’s Ryzen 7 9700X performs against their Ryzen 7 5800X3D. On average, AMD’s new chip is 12% faster. That said, there are still workloads where AMD’s old X3D chip comes out on top.

What about the Ryzen 7 7800X3D? On average, AMD claims that their Ryzen 7 9700X is 2% faster. That means that the 7800X3D will be faster in some games and the 9700X will be faster in others. The 9700X is not a good upgrade over the 7800X3D, but it is faster in games on average. AMD almost certainly has a Ryzen 9 9800X3D in the works, making it the obvious upgrade for 7800X3D users (when it eventually comes to market).

Efficiency, efficiency, efficiency

Despite having lowered TDPs across most of its lineup, AMD’s Ryzen 9000 series handily outperform its last-generation counterparts. AMD has not just boosted the performance of their new CPUs, they have also made them more efficient. This is in stark contrast to Intel in the desktop CPU market, who have delivered much of their recent performance gains with hikes in power consumption.

AMD’s making overclocking fun with Zen 5 and Ryzen 9000

With Zen 5, AMD hopes to make overclocking more rewarding. For starters, AMD’s lowered TDPs give overclockers more power/thermal headroom to utilise. Beyond that, AMD now supports DDR5-8000 memory with their latest AGESA updates. AMD has also added a new “Curve Shaper” to add more tweakability to AMD’s Curve optimiser feature.

Precision Boost Overdrive with Ryzen 9000/Zen 5

Thanks to AMD’s lowered TDPs with Zen 5, Ryzen 9000 users have the ability to boost their TDPs to deliver higher performance levels. For AMD’s Ryzen 7 9700X, up to 15% performance gains can be achieved in multi-threaded workloads. If you have a strong enough cooling solution, this will be a great way for AMD users to get more from their Zen 5 CPUs.

Overall, AMD has delivered some solid improvements with Zen 5. Heightened power efficiency and performance are great things to have in tandem. Lowered TDPs for most models make CPU cooling a simple process, and a wide range of architectural enhancements promise performance gains across all workloads despite lowered power use for most models. AMD has done a great job with Zen 5, and we are excited to see what some next from AMD’s Zen teams.

You can join the discussion on AMD’s Ryzen 9000 series of Zen 5 CPUs on the OC3D Forums.