ToolScore
Local GPU AI 6 min read

Unified Memory, Honest Numbers: 284B on a Laptop, 26 t/s, and the Bandwidth Ceiling

~26 t/s — 284B on a laptop the number that matters here

By Stan K.

Apple M4 Max, M5 Ultra, DGX Spark, Strix Halo and the R9700 compared on verified speeds and prices — and why memory bandwidth, not GB, decides what unified memory can actually run.

A discontinued Apple laptop with 128GB of unified memory runs a 284-billion-parameter DeepSeek at ~26 tokens per second — slightly faster than the same machine runs a 27B dense model. Capacity decides what fits on the machine. Bandwidth decides how fast it thinks. Most buyers shop the first number and get burned by the second.

The one spec that decides everything

Generating a token means streaming the model’s active weights across the memory bus, every token. The people who aggregate these benchmarks say it flatly: generation speed is limited by memory bandwidth, not by memory size. The M4 Max offers 546 GB/s. That is the ceiling for every model it will ever load.

That ceiling produces counterintuitive results. On an M4 Max 128GB (used — Apple pulled it from the lineup, the M5 Max replaced it):

  • A 27B dense model decodes at 22–24 t/s, sagging to 14–15 t/s at 256K context.
  • DeepSeek V4-Flash 284B — a huge MoE whose ~75GB of weights fit comfortably in 128GB — decodes at ~26 t/s, because only a slice of those weights is active per token.

More parameters did not mean slower. Fewer active bytes per token meant faster. That single comparison is the whole unified-memory market in miniature.

The five machines that matter, honestly

Apple M4 Max 128GB (used/refurb). 546 GB/s; the numbers above. Discontinued — Apple pulled it from the lineup — but refurb 128GB units sit at about $6,000 (human-verified 2026-11-10), roughly $47 per GB of unified memory.

Apple M5 Ultra (Mac Studio, current line). 96/256/512GB configurations, all on a 1.2 TB/s bus, listing at $6,000 (96GB), $12,299 (256GB) and $22,299 (512GB). On the 256GB machine, GLM 5.3 Flash (a 320B MoE) runs at 68.8 t/s in MLX (oQ4e quant with MTP), and an 8B Q4 model decodes at 125 t/s — the bus ceiling, visible.

NVIDIA DGX Spark. Real and in stock at $6,950 for 128GB of LPDDR5x — roughly double its $3,999 launch price (verified 2026-11-10) — but on a 273 GB/s bus, half the M4 Max’s bandwidth with identical capacity. NVIDIA officially positions one unit for roughly 200B-parameter inference (fine-tuning to 70B). Pairs and quads scale: two Sparks pushed 2,600+ t/s prefill on DeepSeek V4.1, and four Sparks run a 511B model at ~83 t/s.

Strix Halo (Ryzen AI Max+ 395, 128GB). The value darling of the local-LLM forums: 27B at 64 t/s, a 125B MoE at 44–59 t/s (with ~1,400 t/s prefill), and a 35B-A3B MoE at 190 t/s (Q6). The easiest verified route: the GMKtec EVO-X2 mini-PC with the same chip, $3,650 on Amazon (checked 2026-11-10) — about $29 per GB of unified memory.

Radeon AI PRO R9700 (32GB). $1,900 new (B&H, in stock; human-verified 2026-11-10). Two of them run a 27B at Q8 at 37–50 t/s on ROCm 10 — “works, no wild skills required.” It is a discrete 32GB card rather than unified memory, but it competes for the same desk and budget.

The math, concretely

Compare the buses, not the gigabytes:

  • M5 Ultra: 1,200 GB/s
  • M4 Max 128GB: 546 GB/s
  • DGX Spark: 273 GB/s

The M5 Ultra’s bus is about 2.2x the M4 Max’s and about 4.4x the Spark’s — and single-stream decode scales with roughly that ratio for the same model and quant. The Spark’s 128GB holds the same weights as the M4 Max’s 128GB, then serves them at half the rate.

Unified memory bandwidth — the number that sets decode speed
M5 Ultra (96–512GB) 1,200 GB/s
M4 Max 128GB (used) 546 GB/s
DGX Spark 128GB 273 GB/s

~26 t/sDeepSeek V4-Flash 284B on M4 Max 128GB

A 284-billion-parameter model, on a laptop, at reading speed — because only its active experts cross the 546 GB/s bus each token. The same machine decodes a 27B dense model at 22–24 t/s.

AMD’s software tax, stated honestly

The R9700 and Strix Halo numbers above are real, but the AMD stack has two very different maturity levels. LLM inference on ROCm works, with effort: the community report behind the 37–50 t/s figure needed no patching, yet the official llama.cpp ROCm build has been caught not touching the GPU at all on some hosts — verify your setup actually uses the card. Image and video generation is rough: SDXL artifacts on RDNA4 cards, segmentation faults, and custom nodes that quietly drag in CUDA builds of torch over your ROCm one — the standard fix is pinning the whole ROCm runtime in Docker. ROCm 10 removed the NaN-era problems. Short version: LLMs okay with effort, ComfyUI currently ugly.

There is no M4 Ultra

People write “M4 Ultra” constantly. No such chip was ever sold — the benchmark aggregators carry M4, M4 Pro and M4 Max records only, and Apple’s Studio line jumped straight to the M5 Ultra. If a listing or a forum post says M4 Ultra, the rest of its claims deserve the same skepticism.

The takeaway

Prices on these machines will move; the physics won’t. Shop bandwidth per dollar for decode speed, gigabytes for capacity, and never buy either without a real report of your exact model and quant on that exact machine. If a 27B at 22 t/s is enough, a refurb M4 Max or a Strix Halo box is a quiet, power-sipping answer. If you need 500B-class weights resident, the M5 Ultra’s 1.2 TB/s bus is the only consumer-grade game in town — at $22,299 for the 512GB build, it had better be.

For the discrete-GPU side of this market — used 3060s, 3090s and the 24GB tier where most people should actually start — see our budget build under $800, the $800+ used-GPU guide, and the 2x 3060 vs 3090 duel. Everything local-AI lands in the Local GPU AI hub.

Prices are sold/retail figures human-verified 2026-11-10 (online ask/BIN runs higher on scalped lines); speeds are from community benchmark reports. No verified report means no number printed.

Numbers change — the math doesn't
Token prices and plan limits shift monthly; the structural arguments in this guide age slowly. For current per-tool numbers see the live rankings.