Sunday, June 2nd 2024
AMD Zen 5 Storms into Gaming Desktops with Ryzen 9000 "Granite Ridge" Processors
AMD today announced its much awaited Ryzen 9000 series desktop processors. Built in the Socket AM5 package, and drop-in compatible with all current AM5 motherboards with a BIOS update, the processors are based on the new "Zen 5" CPU microarchitecture. The operational part of the processor, the CPU complex dies (CCDs), are built on the 4 nm process, wired to a 6 nm I/O die. AMD didn't get down into the nuts and bolts of the microarchitecture, but briefly mentioned an impressive 16% IPC increase over "Zen 4." Coupled with the fact that the first wave of processors lack 3D V-cache and can sustain higher boost frequencies and TDP, processors in the series should beat the Ryzen 7 7800X3D in gaming performance, which also means that AMD has beaten the 14th Gen Core "Raptor Lake Refresh" processor series by a significant margin.
The 16% IPC increase over "Zen 4" is backed by branch prediction improvements, wider pipelines and vectors, and deeper window sizes across the core design, for more parallelism. The core also features doubling in instruction bandwidth for front-end instructions, FPU to L1, and L1 to L2 data bandwidth, and a redesigned FPU to double AI performance and AVX512 throughput. The company hasn't put out a block design for "Zen 5," and we'll learn more about it in the run-up to the market availability of these chips some time in July 2024.
The "Granite Ridge" processor features a chiplet-based design, just like the Ryzen 7000 "Raphael," Ryzen 5000 "Vermeer," and Ryzen 3000 "Matisse." CPU core counts range from 6-core/12-thread to 16-core/32-thread. The cores are located in CPU complex dies (CCDs). The 6-core and 8-core models feature a single CCD design, while the 12-core and 16-core ones come with a dual-CCD design. Each CCD contains a single CCX (CPU core complex) with 8 "Zen 5" CPU cores. Each core has 1 MB of dedicated L2 cache, and the eight cores share a 32 MB of L3 cache.
The client I/O die (cIOD) appears unchanged from the previous generation. It's built on the 6 nm process, packs a basic iGPU based on the RDNA 2 architecture with 2 compute units; a dual-channel DDR5 memory controller, and a 28-lane PCI-Express Gen 5 root complex.
AMD has planned four processor models for the first wave of Ryzen 9000 series desktop processors. These are led by the Ryzen 9 9950X, a 16-core/32-thread chip with a maximum boost frequency of 5.70 GHz, and a TDP of 170 W. This is followed by the Ryzen 9 9900X, a 12-core/24-thread part that boosts up to 5.60 GHz, with an interesting TDP number of 120 W. If you recall, its predecessor, the Ryzen 9 7900X, had the same 170 W TDP as the 16-core 7950X. So the TDP has generationally lowered.
The Ryzen 7 9700X is the 8-core/16-thread part from the series, with a maximum boost frequency of 5.50 GHz, and an impressive 65 W TDP. The most affordable part from the series will be the Ryzen 5 9600X. This 6-core/12-thread chip boosts up to 5.40 GHz, and has 65 W TDP.
If you're wondering whether the maximum boost frequencies and TDP have gone down generationally, it's because AMD has switched over to the slightly more efficient 4 nm foundry node for the CCDs, besides the "Zen 5" microarchitecture providing the 16% IPC gain.
AMD only put out performance numbers for the top Ryzen 9 9950X processor, which it compared with the Intel Core i9-14900K. In productivity and content creation workloads, the company is claiming performance leadership ranging anywhere between +7% in the UL Procyon benchmark, to +56% in Blender. The real story, though, is gaming performance, where the 9950X beats the i9-14900K by anywhere between +4% to +23%. Mistral LLM is shown running 20% faster on the 9950X compared to the i9-14900K.
The company didn't talk about pricing, we'll learn more about it as we near the July 2024 launch.
AMD is launching the AMD X870E and X870 desktop chipsets with these processors, heralding a new wave of Socket AM5 motherboards, which we'll see at Computex. These two chipsets standardize USB4 connectivity on all motherboards, as well as PCI-Express 5.0 x16 on all boards, even for the X870. The two chipsets also enable higher AMD EXPO memory frequencies. Again, the processors are perfectly compatible with AMD 600-series chipset motherboards with a BIOS update, and older Ryzen 7000 and Ryzen 8000 series processors should work on AMD 800-series chipset motherboards. AMD is promising longevity for AM5 going beyond even 2027.
The 16% IPC increase over "Zen 4" is backed by branch prediction improvements, wider pipelines and vectors, and deeper window sizes across the core design, for more parallelism. The core also features doubling in instruction bandwidth for front-end instructions, FPU to L1, and L1 to L2 data bandwidth, and a redesigned FPU to double AI performance and AVX512 throughput. The company hasn't put out a block design for "Zen 5," and we'll learn more about it in the run-up to the market availability of these chips some time in July 2024.
The client I/O die (cIOD) appears unchanged from the previous generation. It's built on the 6 nm process, packs a basic iGPU based on the RDNA 2 architecture with 2 compute units; a dual-channel DDR5 memory controller, and a 28-lane PCI-Express Gen 5 root complex.
The Ryzen 7 9700X is the 8-core/16-thread part from the series, with a maximum boost frequency of 5.50 GHz, and an impressive 65 W TDP. The most affordable part from the series will be the Ryzen 5 9600X. This 6-core/12-thread chip boosts up to 5.40 GHz, and has 65 W TDP.
If you're wondering whether the maximum boost frequencies and TDP have gone down generationally, it's because AMD has switched over to the slightly more efficient 4 nm foundry node for the CCDs, besides the "Zen 5" microarchitecture providing the 16% IPC gain.
The company didn't talk about pricing, we'll learn more about it as we near the July 2024 launch.
100 Comments on AMD Zen 5 Storms into Gaming Desktops with Ryzen 9000 "Granite Ridge" Processors
I guess better motherboard traces connecting the CPU to the memory slots might have an effect on memory speeds but I doubt it.
My guess is 9000 series will be capped at 6400-6800 at best if you want to get the lowest latency.
We will have to wait for reviews to know testing shown was done with 6000 mem and a 7900XTX so it isn't even all that clear what actual perfomance is like vs 14th gen with better memory.
AMD has a very special formula to calculate TDP (Thermal Design Power) from actual package power consumption (PPT: PackagePowerTracking)
A formula with very specific conditions that include environment and hardware thermal(temp) conditions and also a cooler with specific thermal resistance.
All and all AMD is saying with this formula (and eventually TDP) the bare minimum cooler required under specific ambient (PC case temp/CPU cooler fan inlet air temp) and CPU Tcase to run the CPU within default/rated specifications.
A CPU that "produces" 142W of heat needs a cooler that can dissipate at least 105W under those specific conditions. The idea is that not all heat coming from any chip is "moving" towards the surface the cooler is installed.
Some of the heat inevitably will be conducted through the socket to board. How much that will be is determined upon all the above thermal conditions of the setup.
-------------------------------------------
2027+ support? Just support or possibly 1-2 additional Gen upgrades on AM5 along the way?
1. Because they will not hurry to release AM5+.
2. Because DDR6 is not coming, there is no need for it. The vast majority of people still use old SATA HDDs & SSDs, PCIe 3.0 and 4.0 drives are modern as we speak. PCIe 5.0 is still terrible and a no-go.
When Zen first launched in 2017 they said at least until 2020, then the first 3D model came 2 years after that in 2022.
Nothing is certain or confirmed though. Things may change down the road.
Frequencies above 6400 were largely not an advantage for single CCD zen 4 parts, whereas dual CCD zen 4 parts could better take advantage of the bandwidth increase (not as relevant in game performance as it is for synthetics and productivity software).
From personal experience, most CPU limited games I’ve tested, my 7900X3D/6400 DDR5 provides better minimums on average than my 14700k/8000 DDR5. Memory and appropriate IF tuning have provided better results on AM5 than memory tuning on Z790, albeit limited due to lack of free time.
The question though is was it the CPU or MB upgrade that improved support. My 1700X could not do 3200 Mhz but the 3600 could easily support 3600 and 5000 chips can do 4000 mhz. Those were always the best speeds anyway to fit the IMC so maybe it is the CPU that improves memory speed and support. Interesting question. I look forward to hearing from people that buy X870E boards in the 7000 CPU owners club. At least one of us will pull the trigger.Don't mention the 7900X3D in that regard you may get attacked for having a chip that is not good at anything. I play at 4K and made sure I had 4K when I had my 5800X3D and the 7900X3D feeds the GPU 3-5 more GB/s in Gaming. All that does is make 4K 144Hz with a 7900XT awesome. I am only running at 5200 too. I am not getting one of these but when X3D parts launch day one for me. Maybe I will look for some Fast Expo RAM on Black Friday.
With regard to IPC, I expected somewhere in the vicinity of 20% but I think with all the front end improvements in both this and Zen 4, they're somewhat bottlenecked on the back end. Let's see what the diagram is in a few weeks time.
On another note, Turin! 192 cores/384 threads on 13 chiplets in one socket is pretty wild.Man that chip was being given away at stupid low prices. I did two builds and it's a great chip if you can get it for a good price. All that drama about having two CCD's with 6 cores each was blown way, way out of proportion.
I mean at some point it was almost the same price as a 5800X3D but the 7900X3D is way faster.