NVIDIA Blackwell GPU Jetson Thor 128GB workstation
Die NVIDIA Blackwell GPU mit 2560 CUDA-Kernen und 96 Tensor Cores der 5. Generation erreicht 2070 TFLOPS für KI, ein seltenes Leistungsniveau auf einer Grafikkarte. Mit 128 GB Speicher und einem PassMark G3D-Score von 20041 glänzt sie bei KI- und Workstation-Aufgaben, ist aber für Gaming schwach (27,1/100). Dieses Developer Kit richtet sich an Entwickler und Forscher, die LLMs und eingebettete KI-Anwendungen ausführen, nicht an Gamer.
Überblick
The 30-Second Version
The Jetson Thor is an AI and robotics powerhouse with 2070 TFLOPS and 128GB of memory, topping our AI charts. But it's a developer kit, not a gaming card, and its raster performance is mediocre at best. Worth every penny for AI work, but skip it if you want to play games.
Pros & Cons
Vorteile
- Top-tier AI performance with 2070 TFLOPS 100th
- 128GB of memory handles huge models locally 100th
- Compact form factor for edge AI and robotics
- 30% faster than DGX Spark for Flux2 workloads
Nachteile
- Gaming performance is rough, especially at 1080p
- Efficiency and thermals are near the bottom of our database
- No OS pre-installed, setup requires some tinkering
- Not optimized for Omniverse or IsaacLab
Was Besitzer sagen
The Word on the Street
Die Fakten
Performance
For AI and LLM work, the Thor is best-in-class. 2070 TFLOPS and 128GB of memory mean you can run massive models locally without breaking a sweat. One reviewer measured it 30% faster than a DGX Spark for Flux2, which tracks with the 100th percentile AI score. The 96 fifth-gen Tensor Cores do the heavy lifting here. But raster performance is just average, sitting at the 48th percentile, and ray tracing is a weak spot at the 36th percentile. Compute and efficiency scores are disappointing too, landing in the bottom quarter. This is a specialized tool, not an all-rounder.
Specifications
Full Specifications
Chipset
| GPU Chipset | NVIDIA Blackwell GPU |
| Brand | NVIDIA |
| Architecture | Blackwell |
| Class | workstation |
Memory
| VRAM | 128 GB |
Compute
| CUDA Cores | 2560 |
| Tensor Cores | 96 |
| Base Clock | 1570 |
| Boost Clock | 1570 |
Performance
| PassMark G3D | 20041 |
| GPU Compute | 8596 |
Connectivity
| DisplayPort | 1 |
| HDMI | 1 |
Physical
| Weight | 2.9 kg / 6.5 lbs |
vs Competition
Compared to the ASUS TUF RTX 5090 or Gigabyte AERO RTX 5080, the Thor gets smoked in gaming and raster workloads. Those cards are built for frame rates, and the Thor's 48th percentile raster score shows it. But flip the script to AI and LLM work, and the Thor leaves them in the dust. The MSI Ventus RTX 5070 Ti and ZOTAC APOCALYPSE are cheaper and better for gaming, but they can't touch 128GB of memory or 2070 TFLOPS. The Sapphire NITRO+ is a different beast entirely, aimed at AMD fans. If you're buying for AI, the Thor is the only real choice here. If you're buying for games, literally any of the others are better.
| Spec | NVIDIA Blackwell GPU Jetson Thor | ASUS TUF Gaming TUF-RTX5090-32G-GAMING | Gigabyte Eagle GV-N507TEAGLEOC ICE-16GD | MSI Ventus RTX 5070 Ti 16G VENTUS 3X PZ OC | ZOTAC APOCALYPSE ZT-B50800C-10P | Sapphire NITRO+ 11322-01-40G |
|---|---|---|---|---|---|---|
| Chipset | NVIDIA Blackwell GPU | GeForce RTX 5090 | GeForce RTX 5070 Ti | GeForce RTX 5070 Ti | GeForce RTX 5080 | Radeon RX 7900 XTX |
| Vram | 128GB | 32GB GDDR7 | 16GB GDDR7 | 16GB GDDR7 | 16GB GDDR7 | 24GB GDDR6 |
| Boost Clock Mhz | 1570 | 2407 | 2542 | 2482 | 2730 | 2680 |
| Tdp W | - | 575 | 300 | 300 | 360 | 339 |
| Memory Bus Bits | - | 512 | 256 | 256 | 256 | 384 |
| Ray Tracing | - | true | true | true | true | true |
| GPU Class | workstation | consumer | consumer | consumer | consumer | consumer |
| Compare | Compare | Compare | Compare | Compare |
| Produkt | AI | Rt | Vram | Raster | Compute | Feature | Thermal | Efficiency | Connectivity | Nutzerresonanz | Affordability |
|---|---|---|---|---|---|---|---|---|---|---|---|
| NVIDIA Blackwell GPU Jetson Thor | 100 | 35.8 | 99.9 | 47.9 | 24.9 | 11.1 | 6.6 | 8 | 21.3 | 47 | 2.6 |
| ASUS TUF Gaming TUF-RTX5090-32G-GAMING Compare | 97 | 90.3 | 97.7 | 98.7 | 97.6 | 96.2 | 72.4 | 24.4 | 99.3 | 99.1 | 3.3 |
| Gigabyte Eagle GV-N507TEAGLEOC ICE-16GD Compare | 79.4 | 88.5 | 85.8 | 89 | 88.1 | 96.2 | 91.5 | 73 | 94.1 | 95.5 | 26.1 |
| MSI Ventus RTX 5070 Ti 16G VENTUS 3X PZ OC Compare | 79.4 | 92.3 | 85.8 | 89 | 88.1 | 96.2 | 91.5 | 73 | 94.1 | 88.8 | 21.2 |
| ZOTAC APOCALYPSE ZT-B50800C-10P Compare | 85.4 | 97.6 | 88.8 | 94.4 | 93.3 | 96.2 | 91.5 | 57.9 | 94.1 | 99.1 | 19.2 |
| Sapphire NITRO+ 11322-01-40G Compare | 94.9 | 96.2 | 91.7 | 83.9 | 95.6 | 83.8 | 91.5 | 52.8 | 98.8 | 78.3 | 17.2 |
Preis
Value & Pricing
At $5,249 USD, this is not a casual purchase. The affordability percentile is 3rd, meaning it's one of the most expensive products in our database. But if you need 128GB of unified memory and 2070 TFLOPS of AI compute, there's nothing else at this price point that comes close. For robotics companies and AI developers, it's a bargain. For anyone else, it's a very expensive paperweight.
Amazon 1 Angebot Ab 5.718 €
Price History
Mehr erfahren
Overview
The Jetson Thor Developer Kit is not a graphics card in the traditional sense. It's a 128GB AI and robotics workstation crammed into a small box, built around NVIDIA's Blackwell architecture with 2560 CUDA cores and 96 Tensor Cores. The headline number is 2070 TFLOPS of AI performance, which puts it at the absolute top of our database for AI workloads. If you're training models or running edge AI, this thing is a monster.
But it's also a developer kit, not a plug-and-play gaming GPU. The PassMark G3D score of 20041 lands it in the middle of the pack for raster performance, and gaming at 1080p is its weakest area by a wide margin. You're paying for AI muscle and 128GB of unified memory, not frame rates.
Common Questions
Q: Can I use this as a gaming GPU?
Technically yes, but it's not great at it. The 20041 PassMark G3D score puts it at the 48th percentile for raster performance, and 1080p gaming is its weakest area. You'd be much better off with a dedicated gaming card like the RTX 5070 Ti or 5080.
Q: Does it come with an operating system installed?
No, the OS is not pre-installed. You'll need to flash it yourself, which is standard for developer kits but can be a surprise if you're expecting a consumer product.
Q: How much memory does it have and why does that matter?
It has 128GB of unified memory, which is the highest we've seen in our database. That's crucial for running large AI models locally without offloading to the cloud or hitting memory limits.
Who Should Skip This
Skip this if you're a gamer or a general PC user. The raster performance is mediocre, efficiency is near the bottom, and you'll pay $5,249 for AI features you'll never use. If you need a card for gaming or creative work, grab an RTX 5080 or 5090 instead. Also skip it if you want a plug-and-play experience, since there's no OS pre-installed and setup requires some technical comfort.
Verdict
Buy the Jetson Thor if you're a robotics engineer, AI researcher, or edge computing developer who needs massive memory and Tensor Core performance in a small package. It's the best AI hardware we've tested, full stop. But if you're looking for a gaming GPU or a general-purpose workstation card, this is the wrong tool. The weak raster performance and poor efficiency make it a poor fit for anything outside its niche.
Usage Scores
- Gesamt44.6
- AI/LLM67.6
- Budget27.4
- Gaming 4k31.8
- Ray Tracing28.5
- Workstation50.6
- Gaming 1080p27
- Gaming 1440p30.3
- Content Creation38.4