AMD defines Strix Halo APU: the 16nuclear Zen 5 CPU + 40nuclear RDNA 3.5 GPU.

In multiple patches on the backend of Linux and AMDGPU LLVM, RDNA 3.5's "GFX1150" and "GFX1151" IDs have appeared multiple times, but the highlight of this patch is that Strix Point and Strix Point Halo are mentioned together with GPU IDs for the first time.

According to previously exposed details, the Strix Point is the mainstream chip for next-generation laptops, while the Strix Halo is a monster level APU aimed at competing with Apple's M-series chips.

Both of these APUs adopt the next-generation Zen 5 microarchitecture, upgraded to RDNA 3.5 for nuclear display, and equipped with XDNA 2 AI accelerator, improving AI performance to between 45 TOPS and 50 TOPS.

AMD's Strix Point Halo APU will once again use the GFX1151 IP, while Strix Point 1 APU will use the GFX-1150 IP. Compared to the existing GFX110X IP, the upcoming GFX115X IP has been upgraded to have higher clock frequencies and more cores, as well as an RNDA 3.5 GPU architecture upgrade.

Sarlak/Strix Halo:

• Also known as Sarlak, it is a multi chiplet derivative of Strix Point, with 16 Zen5 cores and 40 CU RDNA3.5 cores.
• Sarlak offers performance of 45-50 TOPS, which is approximately 5 times higher than the current Phoenix Point product.
• Zen 5 chip design, up to 16 cores, 64MB shared three-level cache, 90W power consumption, CPU speed 25% faster than 16 core Dragon Range.
• Integrate 40 RDNA 3+computing units, equivalent to RTX 4070 Max-Q (90W).
• 256 bit LPDDR5X memory controller, integrated with XDNA engine, 40 TOPS artificial intelligence engine.

Strix Point/Strix Monolithic:

• Strix Monolithic is the true successor to Phoenix and will be launched in mid-2024.
• Strix Point has 4xZen5 and 8xZen5C cores, based on RDNA3.5 iGPU, providing 45-50 INT8 TOPS.
• Zen 5 (4nm) single-chip design, up to 12 hybrid configuration cores (Zen 5+Zen 5C), 32MB shared L3 cache, 50W power consumption, CPU speed 35% faster than Phoenix.
• Integrate 16 RDNA3+computing units, equivalent to RTX 3050 Max-Q.
• 128 bit LPDDR5X memory controller, integrated with XDNA engine, 20 TOPS artificial intelligence engine.
• Expected to be released in the second or third quarter of 2024.