AMD's Ryzen AI Halo puts 128GB of unified memory on your desk — and runs headfirst into a bandwidth wall

AMD is shipping the Ryzen AI Halo Developer Platform, a turnkey mini-PC aimed squarely at engineers who want to run and prototype large local models without renting cloud GPUs. It went up for pre-order in June exclusively through Micro Center and hits shelves July 10 at $3,999 for the 128 GB / 2 TB configuration, per StorageReview. The pitch is explicitly comparative: Tom's Hardware titled its review "AMD builds a DGX Spark of its own," and AMD undercuts Nvidia's DGX Spark — recently repriced to roughly $4,699 amid the LPDDR5X/NAND crunch — by about $700, while adding a Windows 11 option the Linux-only Spark doesn't offer.

The most rigorous teardown comes from LTT Labs, and most of the review pack builds on it. Using llama-bench (from llama.cpp) with the default pp512/tg128 profile, LTT Labs ran Qwen 3.6 35B, Gemma 4 31B and GLM 4.7 Flash. Their load-bearing finding: on the dense Gemma 4 model, Apple's Mac Studio generated text 2–3× faster — attributed to bandwidth, since the Mac's unified memory runs at ~800 GB/s versus the Halo's 256 GB/s.

The pattern holds across the wider data: Mixture-of-Experts models fly (community labs measured Qwen3 30B-A3B MoE at ~86 t/s, GPT-OSS 120B at ~53 t/s decode) because only ~3 B parameters activate per token, while same-size dense models get throttled by the 256 GB/s wall. StorageReview's vLLM serving tests had the Halo trailing the DGX Spark 2×–4× at higher concurrency — stretching to 8.8× on prefill-heavy GPT-OSS 120B work, the Halo's weakest, compute-bound axis. The "up to 200 billion parameters" line is a capacity claim, not a throughput promise: it fits; it won't fly.

Against the Spark, the Halo is cheaper and dual-boots Windows 11 or AMD's Linux image, and XDA notes it does much of what the Spark does at a fraction of the power — but raw inference "trails DGX Spark and GB10 boxes." Against a Mac Studio, the Mac wins dense-model generation on bandwidth while the Halo undercuts it on price with an open ROCm/Linux stack. And against renting: a single H100 hour buys ~20× the bandwidth, so the Halo is for fitting 70B–120B-class models locally and iterating cheaply — not serving them at scale.