This article provides a technical overview and performance analysis of AMD's Strix Halo, a high-end, chiplet-based APU designed for premium mobile devices. The architecture combines up to two 8-core Zen 5 CCDs (for 16 cores total) with a large SoC die featuring a 40 CU RDNA 3.5 integrated GPU, 32MB of Infinity Cache, and a 256-bit LPDDR5X-8000 memory interface.
Performance benchmarks reveal that Strix Halo's CPU is highly competitive, approaching the performance of AMD's flagship desktop CPUs despite a lower clock speed. The integrated GPU is the standout feature, delivering performance that challenges and, in some scenarios, surpasses mid-range dedicated mobile GPUs like the NVIDIA RTX 5070M, particularly in gaming workloads on battery power. The analysis also details the memory subsystem, noting the impressive shared bandwidth of up to 256 GB/s and the GPU's effective use of the large Infinity Cache. The article concludes that Strix Halo is a class-leading, all-in-one processor that enables high-end gaming and compute in small form factors, though its machine learning capabilities could not be tested due to delayed software support.
AMD's Strix Halo: A Chiplet APU Deep Dive
0:00 Introduction to Strix Halo: Strix Halo is AMD's first chiplet-based APU for the consumer market, targeting high-performance mobile devices with a TDP of 55W to 120W. It aims to deliver high-end CPU and GPU performance without requiring a separate dedicated graphics card.
0:30 Core Specifications: The APU features up to 16 Zen 5 CPU cores with a 5.1 GHz boost clock, a powerful RDNA 3.5 integrated GPU with 40 Compute Units (CUs) and a 2.9 GHz boost clock, and 32MB of on-die Infinity Cache.
1:07 High-Bandwidth Memory: To feed its powerful components, Strix Halo is equipped with a 256-bit LPDDR5X-8000 memory bus, providing up to 256 GB/s of shared bandwidth, significantly higher than any other APU tested.
2:02 CPU Memory Performance: From the CPU's perspective, memory latency is around 123ns, which is good for a mobile chip but higher than desktop counterparts. The CPU cores can pull over 175 GB/s of bandwidth from memory, but a single CCD is limited to ~64 GB/s read, similar to desktop chiplet designs. The CPU does not have access to the 32MB Infinity Cache.
3:41 Desktop-Class CPU Performance: In performance tests, Strix Halo's 16-core CPU can match a last-generation desktop flagship (Ryzen 9 7950X) in integer tasks and nearly matches the current flagship (Ryzen 9 9950X) in floating-point tasks, demonstrating impressive efficiency and power.
5:05 GPU Cache & Memory Advantage: The integrated GPU benefits greatly from the 32MB Infinity Cache, which delivers over 40% higher bandwidth than the L2 cache on a comparable dedicated RTX 5070 Mobile GPU. Memory latency for the GPU is also ~35% lower than the dedicated card.
7:01 Competing with Dedicated GPUs: In gaming benchmarks like Cyberpunk 2077, Strix Halo's iGPU is in a class of its own compared to other integrated solutions. It directly competes with the RTX 5070M, even outperforming it by 7.5% at 1080p when both are limited by battery power. On wall power, it remains competitive, trading blows depending on the resolution and settings.
9:15 Conclusion & Limitations: Strix Halo is a versatile and class-leading chip that successfully brings desktop-level CPU and mid-range dedicated GPU performance to an integrated, efficient package. Its primary limitation at the time of review was the lack of mature ROCm software support, preventing an analysis of its machine learning performance.