AMD deep-dives Zen 5 architecture — Ryzen 9000 and AI 300 benchmarks, RDNA 3.5 GPU, XDNA 2, and more

AMD revealed the deep-dive details of its Zen 5 Ryzen 9000 ‘Granite Ridge’ and Ryzen AI 300 series ‘Strix Point’ chips at its Zen 5 Tech Day, digging into the company’s next-generation silicon that will vie for a spot on our list of best CPUs for gaming. AMD has slowly teased the details of its Zen 5 processor, but today, we can share more in-depth AMD benchmarks of desktop and mobile Ryzen processors against competing Intel, Apple, and Qualcomm processors.

AMD also unveiled a new series of motherboards and gave deep dives into the Zen 5 CPU, RDNA 3.5 GPU, and XDNA 2 NPU microarchitectures. We’ll cover these in full on the following pages as AMD speeds toward an on-time launch of its desktop and mobile processors this month.

Image 1 of 3

AMD’s Zen 5-powered chips come to market bearing a new microarchitecture bristling with improvements that yield an average 16% increase in instructions per cycle (IPC) throughput. As you can see in the roadmap above, the Zen 5 architecture will span both the 4nm (N4P) and 3nm process nodes over the course of its life. This means that, much like the company’s other generations of microarchitectures, we can expect second-gen models with even more performance to arrive on a smaller process node. Let’s dive into the details of the processors and benchmarks, then move on to the architectures.

AMD Zen 5 Ryzen 9000 ‘Granite Ridge’ Benchmarks and Specs

We’ve already covered the Ryzen 9000 series product stack, which you can see here. As a reminder, all models except the flagship have lower TDPs than their predecessors yet still deliver impressive generational performance gains. AMD still hasn’t shared pricing, but we’re told an announcement will come before Ryzen 9000’s July 31 launch date.

All the Ryzen 9000 models drop into existing AM5 platforms and also support the upcoming 800-series chipset family, which we’ll cover below. As such, these chips support many of the same features as before, including PCIe 5.0 and DDR5-5600 memory. As with all vendor-provided benchmarks, take these with a grain of salt (test notes at the end of the album).

Image 1 of 8

AMD has previously shared benchmarks showing the flagship 16-core 32-thread Ryzen 9 9950X against Intel’s Core 9-14900K with an average of a 21% advantage in productivity and 11% in gaming. Notably, a portion of the productivity gains come from Ryzen 9000’s doubled performance in AVX-512 workloads, but the uplift in other types of workloads is also impressive across the board.

AMD’s 12-core 24-thread Ryzen 9 9900X is equally impressive in its comparison with the 14900K, notching anywhere from a 2% to 41% lead in a range of productivity applications. It also takes from a 4% to 22% lead over the 14900K in a range of game titles. All told, these figures work out to a ~10% lead in both productivity and gaming over Intel’s flagship (geomean).

The Ryzen 9 9900X’s win is particularly noteworthy given that this 120/162W chip (a 68W reduction in peak from prior gen) faces a 125/253W Intel chip known for easily reaching 350W with power limits removed. Of course, Intel’s Arrow Lake chips will arrive later this year to challenge this power/performance lead.

The eight-core 16-thread Ryzen 7 9700X faces off with the Core i7-14700K in the benchmarks. AMD’s advantage in productivity work becomes more significant despite the fact we’re again looking at a power-sipping AMD 65/88W chip versus a 125/253W Intel model. Overall, AMD claims a ~13% lead in productivity apps and a 10% lead in gaming.

The six-core 12-thread Ryzen 5 9600X contends with the Core i5-14600K (88W vs 181W peak power draw, respectively) and takes a 22% lead in productivity performance (15% without HandBrake outlier) and an 11% lead in gaming. That represents a strong performance gain at the bottom of the Zen 5 CPU stack.

AMD also included AI benchmarks that highlight the advantages of its support for VNNI (an AI-optimized instruction set) for running the Llama and Mistral Large Language Models (LLMs) on Zen 5 processors. As you can see, the Ryzen 9 9900X easily beats the 14900K, but we don’t think this is a very pertinent comparison — most would simply run these models on a GPU. AMD also showcased its PCIe lane advantage over Intel’s current chips, which results in more performance when running an AI model on multiple GPUs while maintaining the full PCIe 5.0 speed for the attached SSD. AMD wins in these AI workloads, but again we don’t see these as relevant workloads for the majority of desktop PC users.

AMD underscored Zen 5’s leap forward in gaming performance by comparing the 9700X to the fastest Zen 3 gaming chip, the Ryzen 7 5800X3D. Long after its release for the AM4 platform, the Zen 3-powered 5800X3D has maintained a significant lead over AMD’s newer standard Zen 4 processors — to this day, it beats the fastest standard Zen 4 chip by roughly 8% in gaming. Now, you’d have to step up to a pricier Zen 5 X3D model to beat the 5800X3D in gaming, but AMD claims Zen 5 Ryzen 7 9700X is roughly 12% faster in 1080p gaming than the 5800X3D, despite its 40W lower power draw.

Notably, AMD avoided comparing its new chips to its own previous-gen Zen 4 models but says the 9700X would beat the Ryzen 7 7800X3D, currently the best gaming CPU on the market, by a “couple percentage points.” These results imply the Ryzen 7 9700X delivers a roughly ~20% gain in gaming performance over the previous-gen Ryzen 7 7700X.

The individual gaming results still warrant a look as well. There remain a few games in AMD's tests where the 9700X failed to beat the 5800X3D, never mind the newer 7800X3D. For those games that really like the large L3 cache, the IPC and other improvements still won't close the gap and we'll need to wait for the inevitable Zen 5 X3D parts to arrive later this year.

The TSMC N4P node and Zen 5 IPC

AMD’s engineers confirmed our suspicion that the Zen 5 processors use TSMC’s N4P node, a marked improvement over Zen 4’s 5nm node. This is important because TSMC’s standard 4nm node (N4) offers much less of an improvement over 5nm, whereas N4P delivers the some of the best generational improvements of the 4nm lineup (TSMC’s 4nm node is officially part of its 5nm family).

TSMC says the N4P node offers 11% more performance, 22% higher power efficiency, and 6% higher transistor density (optical shrink) than the N5 (5nm) node. Additionally, TSMC says this process uses 6% fewer masks for production due to an increased number of EUV layers, implying that it’s more cost-effective. AMD has also designed the Zen 5 process for the TSMC 3nm node, but the company hasn’t given any details on the node or the launch date for those chips.

The N4P process node obviously helps deliver faster and more power-efficient processor performance. Paired with the 16% increase in IPC, AMD's Zen 5 is exceptionally promising. We'll dive deeper into the IPC gain on the Zen 5 architecture page. First, let's take a look at power efficiency, overclocking, motherboards, and Ryzen AI 300.