DGX Spark vs Ryzen AI Max+ 395: which 128 GB AI box should you buy?
We earn a commission when you buy through retailer links on this page, at no cost to you. It never decides what we recommend. As an Amazon Associate I earn from qualifying purchases. How we make money
Quick answer
They write answers at about the same speed, because their memory bandwidth is nearly the same (273 versus 256 GB/s) and both have 128 GB. Choose the NVIDIA DGX Spark ($4,699 list price) if you need NVIDIA's CUDA software, faster prompt reading or its 200 Gbps networking. Choose a Ryzen AI Max+ 395 mini PC such as the GMKtec EVO-X2 ($3,499.99 with 128 GB at GMKtec) to save more than $1,000 and get an ordinary Windows or Linux PC.
These are the two ways to get 128 GB of fast memory into a box you can hold in one hand. NVIDIA’s DGX Spark uses its own GB10 chip; AMD’s Ryzen AI Max+ 395 (the chip family called Strix Halo) powers mini PCs from GMKtec, Minisforum, Beelink, Framework and others. On paper they look very alike. The differences are in software, prompt speed and price.
Side by side
| NVIDIA DGX Spark | Ryzen AI Max+ 395 (GMKtec EVO-X2) | |
|---|---|---|
| Memory | 128 GB unified LPDDR5X | 128 GB unified LPDDR5X-8000 |
| Memory bandwidth | 273 GB/s | 256 GB/s |
| Processor | 20 Arm cores | 16 x86 cores / 32 threads |
| Graphics | Blackwell GPU (CUDA) | Radeon 8060S, 40 compute units |
| Operating system | DGX OS (Linux) | Windows 11 Pro or Linux |
| Networking | 10 GbE + ConnectX-7 (200 Gbps) | 2.5 GbE (10 GbE on some rival models) |
| Storage | 4 TB | 1 or 2 TB, two M.2 slots |
| Price | $4,699 list price | $3,499.99 with 128 GB / 1 TB at GMKtec |
Prices checked September 25, 2026. The DGX Spark often sells above its list price.
Writing speed: a near tie
Writing speed is set by memory bandwidth, and 273 GB/s versus 256 GB/s is only a 7% difference. Published llama.cpp results for the same 7B reference model we use show the GB10 at 58.5 tokens per second and the Radeon 8060S at 48. Both sit close to the line our bandwidth study predicts.
| Machine | gpt-oss 20B12.11 GB file | Qwen3.6 35B-A3B22.13 GB file | Llama 3.3 70B42.52 GB file | gpt-oss 120B63.39 GB file |
|---|---|---|---|---|
| NVIDIA DGX Spark 128 GB, 273 GB/s | 61-79 | 60-77 | 4-5 | 43-56 |
| GMKtec EVO-X2 128 GB, 256 GB/s | 57-74 | 56-73 | 4-5 | 41-53 |
For everyday chat with the models most people run, you will not notice the difference.
Reading long prompts: the DGX Spark pulls ahead
Reading a prompt is limited by computing power, not memory, and here NVIDIA’s Blackwell GPU is much stronger. In the published results, the GB10 read prompts several times faster than the Radeon 8060S. If you feed the model long documents, codebases or retrieval results, the DGX Spark starts answering sooner.
Software: CUDA versus everything else
This is the real dividing line. Most AI research code, fine-tuning tools and serving frameworks are written for NVIDIA’s CUDA first. The DGX Spark runs the same software stack as NVIDIA’s data-center systems, so code developed on it moves to a server unchanged. AMD’s chip works well with llama.cpp and the apps built on it, such as LM Studio, through Vulkan and ROCm, but training and less common tools can take more effort.
Price and flexibility: the Ryzen machines win
A 128 GB Ryzen AI Max+ 395 mini PC can cost more than $1,000 less (the EVO-X2 is $1,199 below the DGX Spark’s list price), and it is an ordinary PC: it runs Windows or Linux, plays games, and doubles as a fast workstation. The DGX Spark is an AI appliance running NVIDIA’s Linux-based DGX OS.
Which to buy
- Buy the DGX Spark if you develop or fine-tune AI models, need CUDA, work with long prompts, or plan to link two units over its 200 Gbps port.
- Buy a Ryzen AI Max+ 395 mini PC if you mainly want to run models, want Windows, or want the most memory for the money. Our mini PC guide compares the models.
NVIDIA DGX Spark
128 GB / 4 TB. 20-core Arm (10 Cortex-X925 + 10 Cortex-A725)
- Memory
- 128 GB unified
- Bandwidth
- 273 GB/s
- Price
- $4,699
NVIDIA's own desktop AI computer, built on the GB10 chip with 128 GB of unified memory. It writes at about the same speed as a Ryzen AI Max+ 395 machine, but runs NVIDIA's CUDA software stack, which most AI development tools support first.
What we like
- Full NVIDIA CUDA software stack
- 200 Gbps ConnectX-7 to link two units for bigger models
- Small for its power
What we don't
- Writes no faster than cheaper Ryzen AI Max+ 395 machines
- Linux only (DGX OS), not a general-purpose Windows PC
- Often sells above list price
Where to buy
- Check price on Amazon Sold by third-party sellers on Amazon
list price $4,699 (NVIDIA list price since February 23, 2026 (previously $3,999); checked Feb 23, 2026). Prices change often; the buttons show today's price at each store.
GMKtec EVO-X2
128 GB / 1 TB (also sold with 64 GB). AMD Ryzen AI Max+ 395 (16 cores / 32 threads, up to 5.1 GHz)
- Memory
- 128 GB unified
- Bandwidth
- 256 GB/s
- Price
- $3,499.99
A Ryzen AI Max+ 395 mini PC with 128 GB of fast unified memory, and the lowest-priced 128 GB machine among those we compared in September 2026. It holds gpt-oss 120B and 70B-class models.
What we like
- 128 GB of 256 GB/s unified memory for less than rivals
- Two M.2 slots
- Also sold with 64 GB for $1,300 less
What we don't
- Memory is soldered; choose the size you need up front
- Only 2.5 Gigabit Ethernet, where some rivals have 10 Gigabit
Where to buy
- Check price at GMKtec 64 GB or 128 GB
- Check price on Amazon 128 GB / 2 TB
$3,499.99 at GMKtec for 128 GB / 1 TB on Sep 25, 2026; list price $3,999.99 (GMKtec store regular price for 128 GB / 1 TB; checked Sep 25, 2026). Prices change often; the buttons show today's price at each store.
Questions people ask
Can the DGX Spark run Windows?
No. It ships with NVIDIA DGX OS, a Linux-based system, and its processor uses Arm cores. Ryzen AI Max+ 395 mini PCs are standard x86 PCs that run Windows or Linux.
Which is faster for local AI?
For writing answers they are close: in published llama.cpp results on the same reference model, NVIDIA's GB10 wrote 58.5 tokens per second and AMD's Radeon 8060S 48. For reading long prompts the GB10 was several times faster in those results.
Can either run a 120B model?
Yes. Both have 128 GB of unified memory, enough for gpt-oss 120B (a 63 GB file) with room for long conversations. We estimate 43 to 56 tokens per second on the DGX Spark and 41 to 53 on the EVO-X2.
Are there cheaper GB10 machines?
ASUS, Dell, HP, Lenovo, MSI, Acer and Gigabyte sell machines with the same GB10 chip. They are sometimes cheaper, often with smaller drives, but prices vary widely, so compare before buying.
Sources
- NVIDIA DGX Spark specifications, checked Sep 25, 2026
- AMD Ryzen AI Max+ 395 specifications, checked Sep 25, 2026
- llama.cpp discussion #10879: Vulkan results (NVIDIA GB10), checked Sep 25, 2026
- llama.cpp discussion #15021: ROCm results (Radeon 8060S), checked Sep 25, 2026
- GMKtec EVO-X2 product page (price), checked Sep 25, 2026
What changed
- Sep 25, 2026: First published.