128GB on Your Desk: Why Bandwidth Sets the Speed

Three kinds of desktop machine now put 128 GB of memory within reach of a home user: NVIDIA's DGX Spark, Apple's Mac Studio, and mini PCs built on AMD's Ryzen AI Max+ 395, the chip nicknamed Strix Halo. All three let the GPU work from one big shared memory pool, so a 70B model at 4-bit, about 44 GB by the llmhardware.io formula, fits on any of them. Fitting is the easy part. How fast the model answers depends mostly on memory bandwidth, and that number differs by more than four times across these machines.

What the spec sheets say

• DGX Spark: 64 GB or 128 GB of LPDDR5x on a 256-bit interface, 273 GB/s, per NVIDIA.
• Ryzen AI Max+ 395: up to 128 GB of LPDDR5x-8000 on a 256-bit bus, per AMD. That works out to about 256 GB/s.
• Mac Studio with M5 Max: 460 GB/s on the 32-core GPU chip or 614 GB/s on the 40-core GPU chip, with up to 128 GB, per Apple.
• Mac Studio with M5 Ultra: 1.2 TB/s, with 96 GB configurable to 256 GB or 512 GB, per Apple.

Why bandwidth sets the pace

When a dense model generates text, it reads essentially all of its weights from memory for every token it writes. A 70B model at 4-bit holds about 35 GB of weights, which is 70 billion parameters at half a byte each. Divide bandwidth by that and you get a hard ceiling on tokens per second, before any compute or software overhead:

• DGX Spark: 273 ÷ 35 ≈ 7.8 tokens per second.
• Strix Halo: 256 ÷ 35 ≈ 7.3.
• M5 Max (40-core GPU): 614 ÷ 35 ≈ 17.5.
• M5 Ultra: 1,200 ÷ 35 ≈ 34.

Measured speeds land below these ceilings, sometimes well below. They are arithmetic, not benchmarks. They still show the gap between machines: on a dense 70B, the M5 Ultra's ceiling is about four times the DGX Spark's.

Mixture-of-experts changes the math

A mixture-of-experts (MoE) model only reads the experts its router picks for each token. OpenAI's model card for gpt-oss-120b lists 117B parameters with 5.1B active. All 117B still have to fit in memory, which is why a 128 GB box suits it, but each token reads only a small slice of them. On bandwidth-limited hardware, that is the reason a large MoE model can feel quicker than a dense 70B that is smaller in total.

How to choose

Pick the DGX Spark if your work is CUDA-first: fine-tuning, or code you will later move to NVIDIA servers. Pick a Strix Halo mini PC if you want 128 GB in an x86 box running Windows or Linux and can accept the lowest bandwidth of the group. Pick a Mac Studio if you want the most bandwidth, or more than 128 GB, since the M5 Ultra goes to 512 GB.

If your models top out around 32B, a single 24 GB or 32 GB graphics card is the simpler buy. A 32B model at 4-bit, about 21.2 GB with overhead, fits on one card with no shared-memory tradeoffs.

The DGX Spark vs Mac Studio vs Strix Halo guide has the full spec table and which model sizes each memory tier holds.

Sources: NVIDIA DGX Spark specs, AMD Ryzen AI Max+ 395 product page, Apple Mac Studio technical specifications, gpt-oss-120b model card on Hugging Face. Token ceilings are bandwidth arithmetic, not measured speeds.