AMD is allegedly planning to release its next-gen Ryzen 8000 processors in the first half of 2024. Based on the 4nm (TSMC N4) Zen 5 cores, these chips are supposed to focus on single-threaded performance, leaving the core counts and Last Level Cache untouched. As per RedGamingTech, the IPC gains for the Zen 5 core will range from 20-30%, but to be honest, another above 20% is extremely unlikely.
Apparently, the core clocks will take a hit, regressing by 200-300MHz compared to existing Ryzen 7000 parts. This is quite strange as the N4 process is simply a fine-tuned N5, not something new. Internally, the Zen 5 core is expected to feature a wider frontend (instruction Q, decoders, uop Q, etc.), a larger Reorder buffer, and improvements to the lower-level cache.
The Zen 5 CCD core-to-core interconnect will be upgraded to a “Ladder L3 Fabric“, originally leaked by AdoredTV. In addition to this, the load-store units should be beefed up along with the L2 bandwidth. AMD has also promised support for integrated AI and Machine Learning optimizations. This probably just means low-precision modes such as FP16 AVX512.
The L1 cache is expected to get bigger, while the L2 and L3 cache on the Ryzen 8000 CPUs will likely stay unchanged. The I/O will also be mostly untouched, with minor improvements to memory speeds and timings. AMD performs AVX512 instructions on its Zen 4 designs using AVX256 units, but this takes multiple cycles, twice as much as a native AVX512 unit.
Seeing how Intel plans to bring back AVX512 to client PCs with its AVX10 package, AMD may decide to place some extra execution units for wider vector compute. Last but not least, the Infinity Fabric definitely needs a higher frequency limit. An increase from 3000MHz to 3,500MHz or 3,600MHz should be sufficient.