AMD is working on its latest mobile processor lineup, the Strix Halo, under the Ryzen AI Max 300 series. These processors utilize a chiplet-based architecture, incorporating one or two Zen 5 Core Complex Dies (CCDs) similar to those found in the Granite Ridge desktop processors. The design integrates a substantial System on Chip (SoC) die featuring an enlarged integrated GPU (iGPU). This configuration allows AMD to equip the Strix Halo processors with up to 16 Zen 5 CPU cores and an iGPU boasting up to 40 RDNA 3.5 compute units, equivalent to 2,560 stream processors. Additionally, the processors support a 256-bit LPDDR5/x memory interface for Unified Memory Architecture (UMA), enhancing data throughput and overall performance.

The Strix Halo series is tailored for ultraportable gaming laptops and mobile workstations where minimizing the printed circuit board (PCB) footprint is crucial, and the inclusion of a discrete GPU is not feasible. For high-performance gaming notebooks that require discrete GPUs, AMD is developing the Fire Range processor. This processor is essentially a mobile Ball Grid Array (BGA) variant of the Granite Ridge and serves as the successor to the Ryzen 7045 series Dragon Range. The Ryzen AI Max series includes three models differentiated by their CPU and iGPU compute unit (CU) counts: the Ryzen AI Max 395+ featuring 16 cores and 32 threads with 40 CUs, the Ryzen AI Max 390 with 12 cores and 24 threads paired with 40 CUs, and the Ryzen AI Max 385 offering 8 cores and 16 threads alongside 32 CUs. An engineering sample of the Ryzen AI Max 390 has been identified in the Geekbench AI benchmark online database, indicating the processor’s specifications and performance metrics.

Benchmark analysis from Geekbench AI reveals that the engineering sample labeled "AMD Eng Sample: 100-000001421-50_Y," affiliated to the Ryzen AI Max 390, operates at a base frequency of 3.20 GHz and can boost up to 5.00 GHz. This sample was tested within a prototype HP ZBook Ultra 14 G1a mobile workstation equipped with 64 GB of memory. The processor achieved a single-precision score of 4,733 points, a half-precision score of 4,944 points, and a quantized score of 13,944 points. These results indicate a performance difference of approximately 60% compared to the desktop Ryzen 9 9900X processor. Several factors may contribute to this discrepancy, including the Balanced power plan used during testing and the utilization of 256-bit AVX2 SIMD instructions instead of the more advanced AVX512. The Zen 5 cores in the Strix Halo retain the full 512-bit floating-point (FP) data paths from Granite Ridge and EPYC Turin, unlike those in the Strix Point monolithic silicon, which are limited to dual-pumped 256-bit FP data paths even when executing AVX512 or Vector Neural Network Instructions (VNNI). Consequently, AI benchmarks leveraging AVX512/VNNI may produce varying outcomes. Additionally, as the current data pertains to an engineering sample, there is a possibility that AMD may intentionally limit its performance before the final retail release.

Sources: Geekbench Browser