AMD Strix Halo (64GB)

AMD Strix Halo (64GB): AMD unveiled Strix Halo (retail name: Ryzen AI Max 300 series) at CES on January 6, 2025. Every vendor selling a 64GB configuration, ASUS's ROG Flow Z13 (GZ302), Framework Desktop, GMKtec's EVO-X2, pairs it with the full Ryzen AI Max+ 395: 16 Zen 5 cores, a 40-compute-unit Radeon 8060S iGPU (RDNA 3.5, 14.8 TFLOPS FP16), and a 50 TOPS NPU, unlike the 32GB tier where some vendors substitute a cut-down chip. Quad-channel LPDDR5X-8000 over a 256-bit bus gives a 256 GB/s theoretical ceiling, identical to every other Strix Halo tier this site tracks (independent testing puts roughly 215 GB/s, ~84%, of that actually achieved), so this 64GB configuration decodes exactly as fast as the pricier 96GB and 128GB tiers on any model that fits in all three: it's the cheapest way to that same speed. Framework Desktop's Ryzen AI Max+ 395 configuration at this capacity launched at $1,599.

This site's calculator puts Llama 3.1 8B at its recommended Q5_K_M (7.58 GB) at 24.6 tok/s and Qwen3 8B at its recommended Q5_K_M (7.73 GB) at 24.1 tok/s, the same numbers the 96GB and 128GB tiers get for the identical builds. Independent Vulkan/RADV community benchmarks report closer to 45 tok/s for a similar 7B Q4 build; this site's decode-efficiency constant was calibrated against a discrete RTX 4090, not this unified-memory architecture, so treat the figures here as conservative. 64GB is the smallest tier where 70B-class dense models genuinely fit rather than needing offload: Llama 3.3 70B and Llama 3.1 70B both reach Q4_K_M (50.75 GB, 3.7 tok/s), and Qwen 2.5 72B lands close behind (52.12 GB, 3.6 tok/s), all with real headroom in this tier's 56 GB of usable memory, something the 32GB tier can't do at any quantization. This site's own top-of-board picks for this capacity include Qwen 3.8 27B and Qwen 3.6 27B at Q8_0 (32.75 GB, 5.7 tok/s), and Qwen 3.5 122B-A10B, a 122B MoE release, squeezing in at the aggressive Q2_K (52.29 GB, 12.9 tok/s) despite its total parameter count dwarfing the dense models around it, only a fraction of a MoE model's weights activate per token. 62 of the 97 models this site tracks fit natively here, 10 more than the 32GB tier below it. None of those tok/s figures capture this platform's most-discussed real-world weakness, though: prompt processing. Strix Halo's iGPU is bandwidth-bound rather than compute-bound, so long prompts are genuinely slow to chew through before the first output token appears. Independent benchmarking (datahardware.ai) measured GPT-OSS 120B prefilling at only about 340 tok/s on this chip, roughly a fifth of the ~1,700 tok/s NVIDIA's compute-bound DGX Spark reaches on the identical model, and real document text prefills 24-33% slower still than the synthetic prompts most benchmarks use. The practical effect: a 12,000-token prompt needs roughly 35 seconds of processing before generation even starts, versus about 7 seconds on DGX Spark. Decode itself keeps sliding well past this page's 8k-context estimate too: one independent long-context benchmark found generation speed dropping by roughly two-thirds once the KV cache filled to around 76k tokens.

Vulkan (RADV or AMDVLK) via llama.cpp works everywhere; ROCm 6.4+ targets this chip directly as gfx1151 on Linux. Independent backend testing (soothill.io, kyuz0's toolboxes benchmarks) found no single fastest backend: ROCm usually wins prompt processing, Vulkan usually wins token generation, so which one to pick depends on whether long prompts or generation speed matter more for your workload. On Windows, AMD's Adrenalin driver exposes a Variable Graphics Memory setting to split this pool between system RAM and VRAM; on Linux the more common approach is a small BIOS UMA carve-out plus the amdgpu.gttsize kernel parameter.