Heidelberg AI News

Video · Alex Ziskind ·

NVIDIA, It’s Happening Again

In short

The video compares the $7,000 GMKtec EVO X5 Pro with NVIDIA’s DGX Spark for local AI, focusing on model capacity, coding agents, inference, image and video generation, and operating-system tradeoffs. The EVO X5 Pro fits larger models and longer contexts, while the Spark is generally faster when models fit within its 128 GB GPU allocation and is better suited to serving workloads.

Key points

  • The EVO X5 Pro has 192 GB unified memory versus the Spark’s 128 GB, enabling models such as GLM 5.3 at 157 GB and up to 262,000-token contexts.
  • With Linux and a 160 GB GPU allocation, the EVO X5 Pro runs GLM 5.3 at 14.2 tokens per second, DeepSeek V4 Flash at 15.9, and Qwen 235B at 15.8.
  • The Spark processes prompts roughly 1.5 to 3 times faster and generates images and video faster, including a 5-second LTX 2.3 clip in 40 seconds versus 196 seconds on the EVO X5 Pro before Linux tuning.
  • The EVO X5 Pro completed a four-model agent workload at 0.63 tasks per minute versus 0.54 on the Spark because the Spark had to swap its fourth model.
  • Allocating 160 GB to the GPU leaves about 32 GB for system RAM, limiting the EVO X5 Pro to roughly two clean Docker agents versus eight on the Spark.
  • WSL cannot access the EVO X5 Pro’s AMD GPU, so serious Linux use requires dual booting; a one-line Linux fix improved video generation from 316 to 126 seconds.
  • The recommendation is the EVO X5 Pro for 200-billion-class models and long context, and the Spark for faster inference, media generation, or multi-user serving when models fit in 128 GB.

Written by an AI model from the video’s captions and description.

Watch on YouTube The opening of the article at youtube.com.

More from Alex Ziskind

All stories

VideoAlex Ziskind

I Had 192 GB of VRAM… My AI Agents Still Failed

The video tests whether system RAM or GPU VRAM limits local AI agents on a Threadripper workstation. It finds that RAM determines model and agent capacity, while VRAM mainly determines parallelism and speed, making more RAM valuable for large agent workloads.

VideoAlex Ziskind

NVIDIA Didn't Want Me to Do This… Linux on RTX Spark

The video tests Linux on MSI's RTX Spark N1X laptop against Windows, WSL, and DGX Spark, focusing on hardware support, memory allocation, CUDA, and AI workloads. Linux mostly works, but an unavailable NVIDIA power-management controller severely limits plugged-in GPU performance, making WSL the practical option for developers today.

VideoAlex Ziskind

I Tested Two $699 Laptops for Coding… It Got Weird

The video compares the $699 Dell XPS 13 and Apple MacBook Neo for coding, local AI, thermals, and battery life. The XPS is generally faster and more upgradeable, while the Neo offers better efficiency and ecosystem integration, but 8 GB of RAM limits larger local AI models on both.