Zen 7 aka “Grimlock” at a glance, big leaps at AMD

With Zen 7 aka “Grimlock”, an epochal generational leap in AMD’s CPU architecture is on the horizon. Based on numerous leaks and internal AMD documents, a picture is emerging that is more than just an iterative upgrade. Zen 7 not only brings higher clock speeds and more cores, but also a radically increased cache architecture, a significantly increased efficiency profile and a clear strategic thrust against Intel’s upcoming generations such as Nova Lake and Raptor Lake successors.

In the desktop segment, AMD is planning up to 32 cores and 64 threads, realized by two 16-core chips (codenamed “Silverton”) based on the TSMC A14 node, which presumably corresponds to a 1.4 nm process. These CCDs should not only accommodate more cores, but also have 2 MB L2 cache and 4 MB L3 cache per core. In total, this results in a massive on-die cache of 96 MB per CCD. The L3 cache can be expanded to up to 224 MB per chiplet with additional 3D V-cache layers with 10 MB per core. This means that up to 448 MB L3 cache is possible in dual CCD configurations, an unprecedented value on the desktop market.

This cache monstrosity is not an end in itself: benchmarks from internal AMD documents speak of an IPC increase of 15 to 25 % compared to Zen 6. In SPECint 2017, depending on the configuration, even up to 25 % more performance was measured. In Cinebench and other applications outside of gaming, performance increases of 16 to 20 % are mentioned. Cache optimization alone is said to deliver an 8% IPC gain. Combined with an expected increase in clock speed (up to 7 GHz is envisaged for certain variants), this results in a single-thread performance increase of 15 to 30 %, and a jump of 50 to 67 % in multi-thread compared to Zen 6.

Zen 7 continues to use the AM5 platform and can therefore run on existing motherboards, provided the BIOS is updated accordingly. This underlines AMD’s strategy of combining platform loyalty with maximum scalability. The IODs will largely remain on the familiar 6 nm node, which saves costs and simplifies integration. There are two variants in the mobile sector: Grimlock Point (basic APU with integrated cores) and Grimlock Halo (flagship with additional chiplet support). Grimlock Point offers up to 12 cores (4x Zen 7 8x Zen 7C) in a monolithic die, supplemented by possible low-power cores. Grimlock Halo, on the other hand, combines 8 Zen 7 cores with 12 Zen 7C cores and supports two additional Silver King CCDs, enabling up to 36 cores in a mobile SoC – a first in the laptop segment.

The energy efficiency is particularly impressive: at just 3 watts per core, performance increases by 30 to 36% compared to Medusa (Zen 6). This means: the same TDP, more performance or the same performance with significantly lower power consumption. In typical 25-50 watt TDP laptops, a performance increase of over 30 % is therefore realistic. The efficiency ratio compared to Zen 5 doubles in some scenarios. In terms of price, however, Zen 7 is no bargain. The large silicon area (up to 351 mm² without V-cache) and the expensive 1.4 nm manufacturing process suggest that a 32-core top model with V-cache will cost well over USD 999. Nevertheless, the price-performance ratio remains attractive: A hypothetical Ryzen 11 970X3D2 would be two to three times as fast as a Ryzen 9 9950X3D2, at perhaps a 50% premium. At the same time, smaller eight-cores with Silver King CCDs are to be offered for less than 300 USD.

Conclusion: Zen 7 is not a stopover, but a technological one-two punch, comparable to the evolution from Zen 2 to Zen 3. If Zen 6 is the basis, then Zen 7 is the weapon. AMD plans not only to take the technological lead, but to shift the performance spectrum towards HEDT on the mainstream AM5 socket. The competition, especially Intel, will have to dress warmly or risk falling behind for good in 2028

Source: Moore’s Law Is Dead via Youtube