DGX Spark, Ryzen AI Max+ 395, Mac Studio, RTX tower: four kinds of machine now run big models locally, and spec sheets make them all look alike. They are not. Start from the job, and the choice gets short.
Two numbers do most of the sorting. Memory size decides which model fits at all. Memory bandwidth, how fast the chip reads its own memory, decides how fast the answer is written (tok/s, tokens per second, roughly words per second). A third matters for documents: prompt processing, how fast the machine reads what you give it before it starts answering.
The decision table
| Job | Best fit | Why | Watch out |
|---|---|---|---|
| One person chatting with the biggest MoE model | A 128 GB unified-memory box, or a Mac with high bandwidth | MoE (mixture of experts) models use only a slice of their weights per word, so big memory plus decent bandwidth is enough. For one person, a Mac is great | Macs have no CUDA, and we do not sell them; this is a comparison |
| CUDA development and fine-tuning experiments on a desk | NVIDIA DGX Spark or ASUS Ascent GX10 | Full CUDA on DGX OS, 128 GB, and a ConnectX-7 port to link two units | Dense models (every weight used per word) run slowly on it |
| 128 GB of unified memory in a small box, without CUDA, one to five people | Ryzen AI Max+ 395 box | Linux can give the GPU about 120 GB, Windows about 96 GB. Writing speed is in the same class as a Spark | Prompt reading is slower, and CUDA tools do not run |
| A team chatting with long documents | RTX PRO GPU tower or server | Fast prompt reading and room for many sessions | VRAM (graphics card memory) is the limit: size it per user |
| Image and video generation | NVIDIA GPU, sized per card | The tools are built for CUDA and need fast compute | Unified boxes work but are slow for video |
| Files plus AI in one box, office of 10 to 100 | QNAP QAI-h1290FX | NAS storage and an RTX PRO card in one chassis | Only two cards are validated: 32 GB or 96 GB |
Evidence 1: memory bandwidth
Writing speed is bounded by how much data the chip can read per second, so bandwidth predicts tok/s for large models better than any other spec. The chart uses manufacturer figures.
Memory bandwidth
GB/s. Higher is better. Manufacturer spec pages.
Source: Manufacturer spec pages (AMD, NVIDIA, Apple). The Ryzen figure follows from LPDDR5X-8000 on a 256-bit bus.
The two 128 GB boxes read memory at almost the same speed. A discrete GPU reads three to seven times faster, but holds far less per card. That is the whole trade: size against speed.
Evidence 2: what the estimates say
| Model | GX10 (128 GB) | Ryzen AI Max+ 395 | GPU |
|---|---|---|---|
| Qwen3.8 27B (dense) | about 10 | about 10 | about 63 on RTX 5090 |
| gpt-oss-120b (MoE) | about 42 | about 40 | about 176 on one RTX PRO 6000 |
The dense 27B model, which uses every weight for every word, crawls on a unified box and flies on a GPU. The big MoE model is comfortable on both, which is why a 128 GB box makes sense for it. The 120B model does not fit in a 32 GB card at all.
Evidence 3: reading prompts
RAG (answering from your documents) and agents send long prompts, so reading speed counts. Public llama.cpp results on a 7B Q4_0 model show prompt processing (pp512, 512 tokens) and writing (tg128, 128 tokens):
| Hardware | Reading (pp512) | Writing (tg128) |
|---|---|---|
| Ryzen AI Max+ 395 (ROCm) | 911 to 1,004 | about 50 |
| DGX Spark | 3,062 to 3,661 | about 57 |
| Apple M5 Max (40-core GPU) | 3,220 | about 120 |
| Apple M5 Ultra (80-core GPU) | 4,945 | about 179 |
| RTX PRO 4500 Blackwell | 7,252 | about 169 |
| RTX 5090 | 14,073 to 14,970 | about 290 to 300 |
Read it fairly. The Apple chips read as fast as or faster than the DGX Spark here. The big gap is graphics cards (roughly 14,000 and up for the top cards) against every unified-memory machine (about 900 to 4,900). The Ryzen box is the slowest reader of the group.
A myth worth killing: "50 TOPS"
The "TOPS" figure on an AI PC spec sheet usually describes the NPU, a small block for light background tasks. It does not make a chat model write faster. Memory bandwidth does.
What we would put on your desk
- One person or a small group, models up to about 30B: the Studio workstation with a 32 GB card.
- A department with a big model and long documents: the AI workstation, two RTX PRO 6000 Max-Q cards, 192 GB.
- An office that wants its files and AI together: the QNAP AI server.
- CUDA development, or 128 GB in a small box: an ASUS Ascent GX10 or a 128 GB Ryzen AI Max+ box. Ask us through the quote form and we will check what fits your work.
Clustering several small boxes lets you load a bigger model. It does not make that model faster.
Read more in this series
- Unified memory vs VRAM: what a 128 GB AI box really gives you
- Fits in memory, still slow: why bandwidth sets your tok/s
- Prompt processing: the hidden spec for RAG and agents
- What a 128 GB AI box can run, model by model
- What fits in 32 GB of VRAM
- One dev or the whole office? Where an AI box stops and a GPU tower starts
- Sizing local AI for a team of 5, 15 or 50
The short version
- Pick by job. Memory size sets what runs, bandwidth sets writing speed, and prompt reading sets how documents feel.
- A 128 GB unified box suits big MoE models for one to five people. It is slow on dense models and on long prompts.
- The DGX Spark family is for CUDA work. The Ryzen box is the other way into 128 GB, with no CUDA.
- A team reading long documents needs a GPU tower, sized by memory per user.
- "50 TOPS" does not speed up chat. Bandwidth does.
Speeds in the second table are estimates. Try your own model in the model finder, or see which hardware runs which model on the AI hardware page.