Unified‑memory rigs compared
A unified‑memory rig roundup puts AMD’s Strix Halo (Ryzen AI Max+ 395, ~$2,500) out front for local AI + AAA gaming, posting 9–12 tps decode and beating pricier M3 Ultra and DGX Spark configs. The comparison highlights that integrated memory designs can hit strong AI decode speeds while retaining full AAA gaming support. (x.com)
The published comparisons were run by multiple outlets testing “unified‑memory” mini workstations side‑by‑side, including The Register’s hands‑on and an updated DGX Spark review that rebenchmarked Spark against desktop APU boxes. (theregister.com)) AMD’s Strix Halo platform centers on the Ryzen AI Max+ 395 APU with 16 Zen‑5 CPU cores, a 40‑CU Radeon 8060S iGPU and a 50‑TOPS XDNA 2 NPU, and ships in systems using up to 128 GB of LPDDR5x‑8000 unified memory. (tomshardware.com)) NVIDIA’s DGX Spark is built around the GB10 “Grace Blackwell” superchip, offers 128 GB of coherent unified LPDDR5x system memory and peak petaflop‑class FP4 performance, and its MSRP was adjusted to $4,699 following a $700 increase announced by NVIDIA. (nvidia.com)) Apple’s M3/M3 Ultra systems use a much higher‑bandwidth unified memory fabric (Apple lists M3 Ultra configurations with up to 819 GB/s of memory bandwidth on Mac Studio spec pages), a factor reviewers flagged when comparing throughput on very large models. (support.apple.com)) Reviewers and community testers noted software and tooling differences: AMD’s Strix Halo devices are leveraging ROCm/HIP stacks with an explicit migration path toward datacenter ROCm, while community projects—like the strix‑halo‑ai‑stack and independent llama.cpp test harnesses—have emerged to optimize those APUs for local LLM inference. (theregister.com)) Benchmarks and spec sheets show the practical trade-offs: Strix Halo configurations hit theoretical memory bandwidths around ~256 GB/s for LPDDR5x‑8000 setups, DGX Spark lists roughly 273 GB/s of shared bandwidth on GB10, and Apple’s Ultra parts deliver much higher raw bandwidth (hundreds of GB/s higher), a spread reviewers point to when explaining why Strix Halo can be cost‑efficient for many models but still trails the highest‑bandwidth Apple slabs on the largest workloads. (hardware-corner.net))
