This AI Supercomputer can fit on your desk...

NetworkChuck Technology Careers Career & Money

About this video

Chuck reviews the NVIDIA DGX Spark, a palm-sized desktop AI machine NVIDIA sent him, and benchmarks it against ‘Terry,’ his home-built dual-RTX-4090 AI server. The Spark uses a GB10 Grace Blackwell superchip with a 20-core ARM processor and 128 GB of unified memory shared between CPU and GPU, can run models up to 200 billion parameters, and his 4 TB Founders Edition costs $3,999 (he says cheaper OEM variants, such as a 2 TB model around $3,000, are coming). On raw speed Terry wins decisively: on a small Qwen model Terry hit 132 tokens per second versus the Spark’s 36, image generation in ComfyUI ran about 11 iterations per second versus roughly 1, and training ran about 1 second per iteration versus 3. After a call with NVIDIA, Chuck reframes the Spark’s real advantages. First, capacity: Terry’s two 4090s offer only 48 GB of VRAM, while the Spark’s 128 GB lets it run multi-model agent setups (one demo used 89 GB across three models) and load large models Terry cannot even start training. Second, fine-tuning on your own desk instead of renting cloud GPUs he prices at about $30 an hour. Third, dedicated FP4 hardware: the Spark runs NVFP4 quantization natively at close to FP8 quality (he cites under 1% accuracy loss), enabling speculative decoding, where a small draft model and a large verifier model run simultaneously; his test used 77 GB and felt fast on a 70B model. Power is another win: the Spark draws 240 watts versus Terry’s 1,100, which he estimates at roughly $315 a year to run 24/7 versus about $1,400 for Terry. Two Sparks can also be linked over a QSFP connection at 200 Gbits per second. His verdict: not a supercomputer for high-speed inference hobbyists, but a genuinely new option for AI developers who fine-tune and need lots of memory in a tiny, quiet, easy-to-set-up box (NVIDIA Sync makes remote access from a laptop simple). A sponsor segment covers Twingate for remote access, and the video ends with a short prayer for viewers, a new closing habit he explains on camera.

Key takeaways

DGX Spark specs per the video: GB10 Grace Blackwell chip, 20-core ARM CPU, 128 GB unified memory, models up to 200B parameters, $3,999 for the 4 TB Founders Edition
His dual-4090 server beat the Spark on every speed test (132 vs 36 tokens/sec; ~11 vs ~1 image iterations/sec; ~3x faster training iterations)
The Spark’s edge is memory capacity, not speed: multi-model agent setups and large-model training that 48 GB of VRAM cannot hold
Native FP4 hardware enables speculative decoding (draft model plus verifier model at once) at near-FP8 quality
Running costs strongly favor the Spark: 240 W vs 1,100 W, which he estimates at about $315 vs $1,400 per year

Episode details: published October 14, 2025 · 23:59 long. Summary written by Self Growth Videos from the episode itself. Business results and any income figures mentioned are NetworkChuck’s own claims, shared for education — they are not typical or guaranteed.

← Previous: n8n Now Runs My ENTIRE Homelab · Next: You’ve Been Using AI the Hard Way (Use This Instead) →

All NetworkChuck videos · NetworkChuck profile · NetworkChuck on YouTube