NVIDIA DGX Spark Personal AI Desktop Supercomputer Desktop GB10 Grace Blackwell Chip Review
NVIDIA's tiny DGX Spark brings supercomputer AI power to your desk, but reliability reports and a wild price range mean only dedicated ML pros should bite.
The 30-Second Version
A genuine petaFLOP AI monster that fits on your desk, but reliability hiccups and a staggering price spread mean it's only for hardcore ML engineers. If you don't need to locally train 70B+ models, this isn't your machine.
Overview
The NVIDIA DGX Spark is a petaFLOP of AI compute crammed into a box the size of a thick book. If you're a data scientist or machine learning engineer who needs to fine-tune huge models locally, this is a game-changer. But it's not a regular PC. That 20-core Apple Cortex CPU is mediocre, and reliability reports make us nervous. Buy this because you need its specific AI muscle, not because you want a general-purpose desktop.
Performance
The Grace Blackwell GPU here is a beast, sitting in the 95th percentile of our database for raw compute. It churns through inference and training with 128GB of unified memory, so you can throw 200-billion-parameter models at it without breaking a sweat. But that CPU? 36th percentile means it's pretty weak by desktop standards. And the thermal design seems iffy. A few owners report that things get toasty under sustained load, which aligns with that 12th-percentile reliability score. For bursty AI workloads it's fine, but expect some fan noise and maybe a repaste if you're pushing it 24/7.
Pros & Cons
Pros
- 1 petaFLOP AI performance in a tiny package 99th
- 128GB unified memory handles massive models 97th
- Full NVIDIA software stack works out of the box 95th
- Surprisingly compact and energy-efficient at 240W 88th
Cons
- CPU is painfully average for a $4k+ machine 12th
- Reliability concerns and thermal throttling reports
- Gaming performance is basically nonexistent (20.9/100)
- Price varies wildly from $4,299 to nearly $590k
The Word on the Street
Specifications
Full Specifications
Processor
| CPU | 3.8 GHz cortex |
| Cores | 20 |
| Frequency | 3.8 GHz |
Graphics
| GPU | Graphics |
| Type | integrated |
| VRAM | 48 GB |
| VRAM Type | GDDR6 |
Memory & Storage
| RAM | 128 GB |
| RAM Generation | DDR5 |
| Storage | 4 TB |
| Storage Type | SSD |
Build
| Form Factor | Mini |
| PSU | 240 |
| Weight | 3.0 kg / 6.7 lbs |
Connectivity
| USB-C Ports | 4 |
| USB Ports | 4 |
| HDMI | 1x HDMI 2.1a |
| DisplayPort | 1x DisplayPort |
| Wi-Fi | Wi-Fi 7 |
| Bluetooth | Bluetooth 5.4 + LE |
| Ethernet | 10 GbE |
System
| OS | NVIDIA DGX OS |
Value & Pricing
There's a $585,691 gap between the lowest and highest price we found. The $4,299 listing on Amazon is the only one that makes sense. At that price, you're getting unique AI hardware that no traditional desktop can match. But if you see a listing for anything above $5,000, run. Some resellers are clearly inflating the price because it's rare. For the right buyer, the entry price is almost a steal. For everyone else, it's a confusing mess.
vs Competition
Next to something like an HP OMEN 45L or Lenovo Legion Tower 5i, the DGX Spark looks like a different species. Those machines crush games and general productivity with high-end consumer GPUs and fast CPUs, but they can't touch 1 petaFLOP of AI compute or run 200B-param models locally. The Spark is for niche AI work only. If you need a PC that does everything, grab a Legion. If you're training foundation models and value time-to-solution, the DGX Spark is in a league of its own—just don't expect it to play Fortnite.
| Spec | NVIDIA DGX Spark Personal AI Desktop Supercomputer Desktop GB10 Grace Blackwell Chip | HP OMEN 45L GT22-3080 | ASUS Republic of Gamers GM700TZ-BS978 | Lenovo Legion Tower 5i Legion Tower 5i Gen 10 | MSI EdgeXpert EdgeXpert-11SUS | Dell XPS EBT2250 |
|---|---|---|---|---|---|---|
| CPU | 3.8 GHz cortex | Intel Core Ultra 7 265K | AMD Ryzen 9 9950X | Intel Core Ultra 7 265F | ARM | Intel Core Ultra 7 265 |
| RAM (GB) | 128 | 32 | 64 | 32 | 128 | 32 |
| Storage (GB) | 4096 | 2048 | 2048 | 2048 | 4096 | 2048 |
| GPU | AMD Graphics | NVIDIA GeForce RTX 5080 | AMD Radeon RX 9070 XT | NVIDIA GeForce RTX 5070 | NVIDIA Blackwell GPU | NVIDIA GeForce RTX 5060 |
| Form Factor | Mini | mid-tower | mid-tower | mid-tower | mini | mid-tower |
| Psu W | 240 | 850 | 850 | 850 | 240 | 460 |
| OS | NVIDIA DGX OS | Windows 11 Pro | Windows 11 Home | Windows 11 Home | Windows 11 Pro | Windows 11 Pro |
| Compare | Compare | Compare | Compare | Compare |
| Product | Cpu | Gpu | Ram | Port | Storage | Reliability | Social Proof |
|---|---|---|---|---|---|---|---|
| NVIDIA DGX Spark Personal AI Desktop Supercomputer Desktop GB10 Grace Blackwell Chip | 36.2 | 95.4 | 98.9 | 88.1 | 97.3 | 12.3 | 68 |
| HP OMEN 45L GT22-3080 Compare | 95.9 | 88.3 | 78 | 93.8 | 91.1 | 71.6 | 84.8 |
| ASUS Republic of Gamers GM700TZ-BS978 Compare | 98.8 | 77.3 | 94.1 | 97.4 | 91.1 | 39.8 | 72.2 |
| Lenovo Legion Tower 5i Legion Tower 5i Gen 10 Compare | 86.5 | 81.3 | 82.1 | 90 | 91.1 | 71.6 | 95.4 |
| MSI EdgeXpert EdgeXpert-11SUS Compare | 99.6 | 95.4 | 98.9 | 88.1 | 97.3 | 39.8 | 83.6 |
| Dell XPS EBT2250 Compare | 88.8 | 69.4 | 78 | 79.6 | 83.8 | 71.6 | 99.7 |
Common Questions
Q: Can I use the DGX Spark for gaming?
No. It scored a pathetic 20.9 out of 100 in gaming tests. The integrated AMD GPU and CPU just aren't built for it. Stick to a gaming rig.
Q: What size AI models can it run?
Up to 200 billion parameters at FP4 precision, thanks to 128GB of unified memory. For FP16 or FP32, you'll hit memory limits sooner, but it crushes most open-source models.
Q: Why is the price so inconsistent?
It's a niche product with limited availability, so scalpers inflate prices. The $4,299 listing on Amazon is the real deal. Anything above $5k is a ripoff; don't pay it.
Who Should Skip This
If you're building a gaming PC, a video editing workstation, or just need a fast machine for web development, this isn't it. Go get an HP OMEN or Lenovo Legion with a beefy consumer GPU. The DGX Spark is a specialized AI tool that's overkill and underperforming for everyday tasks.
Verdict
For the right person, this is the most exciting desktop hardware in years. If you genuinely need to train or fine-tune massive models outside the cloud, the DGX Spark is unmatched. But if your work doesn't involve billion-parameter AI, skip it. Seriously. The middling CPU, iffy thermals, and zero gaming capability mean it's a terrible general-purpose PC. Buy it only if AI is your full-time job and you know exactly why you need a petaFLOP on your desk.