AI HW: AMD Ryzen AI Max+ 395 vs Nvidia DGX Spark (GB10)

it does not seem to be really available for $1500, but for $3400...

AMD’s CEO Destroyed NVIDIA's Most Expensive Supercomputers With a $1,500 Lunch-Box PC! - YouTube

• AMD Ryzen AI Max+ 395 (Strix Halo): The dedicated XDNA 2 NPU on this chip is rated for approximately 50 TOPS (INT8). When accounting for the CPU and the integrated Radeon 8060S GPU, the total platform capability is often cited as reaching up to 126 TOPS.
• Nvidia DGX Spark (GB10): The "1000" figure you mentioned refers to the device's marketing performance at FP4 precision (often cited as up to 1 PetaFLOP of compute). Comparing the two is complex because they operate on different architectures and precision formats.

Amazon.com: GMKtec AI Mini PC Ryzen Al Max+ 395 (up to 5.1GHz) 2TB PCIe 4.0 SSD 64GB LPDDR5X 8000MHz (8GB*8) Quad Screen 8K Display/WiFi 7/ USB4/ SD Card Reader 4.0 EVO-X2 : Electronics

 AMD Destroyed Nvidia’s $4,000 AI Box! 😱 - YouTube

The video is explaining the Ryzen AI Max+ 395 (codenamed "Strix Halo") APU from AMD, showcased inside a compact, lunchbox-sized mini PC (such as the GMKtec EVO-X2).
The Technology
• The Processor: The system is powered by AMD's flagship Ryzen AI Max+ 395 processor, built with 16 Zen 5 CPU cores and a massive Radeon 8060S integrated graphics chip (RDNA 3.5).   

• The Unified Memory Trick: Unlike traditional PC setups where the graphics card (VRAM) and system memory (RAM) are completely separate, this APU uses up to 128GB of high-speed unified memory. On Linux, it can allocate up to 110GB of that memory directly to the GPU.   
The Nvidia "Killer" Context
• The VRAM Problem: Large Language Models (like the 235-billion parameter DeepSeek-R1) require massive amounts of video memory (VRAM) to run locally. To do this on Nvidia hardware, you would typically need multiple expensive graphics cards or dedicated AI enterprise hardware (like the Nvidia DGX Spark) costing $4,000+.   

• The Price Advantage: Because AMD's "lunchbox" mini PC can use its unified system memory as VRAM, it can hold these massive AI models entirely local for a starting price of around $1,499. AMD claims this allows the mini PC to outperform a single Nvidia GeForce RTX 5080 by up to 3x in specific DeepSeek-R1 inference workloads due to Nvidia's strict VRAM limits.
Ryzen AI Max+ 395: Run a 235B Model Locally on a Mini PC 
XDNA 2, over 50 TOPS (about 126 TOPS combining CPU, GPU and NPU)