ASUS Ascent GX10, Cortex, 128GB, 1TB SSD
The NVIDIA GB10 Grace Blackwell Superchip delivers 1 petaFLOP of AI performance and 128GB of memory, enabling fine-tuning of 200B parameter models on a desktop. Its stackable chassis with NVLink-C2C and ConnectX-7 networking allows dual-system scaling, while the compact 1.48kg build and advanced cooling sustain long-running agentic workflows. This is for AI developers who need secure, private on-device inference and sandboxed execution for large model development.
Aucun prix actuel chez les marchands
Aperçu
The 30-Second Version
The ASUS Ascent GX10 is a tiny AI powerhouse with 128GB of memory and a petaFLOP of performance, but it's a one-trick pony. The CPU is weak, thermals limit sustained training, and ports are sparse. Worth it for serious local AI work, not for much else.
Pros & Cons
Points forts
- 128GB of unified memory is best-in-class for local AI work. 99th
- The GB10 Superchip delivers a genuine petaFLOP of AI performance. 84th
- Tiny 1.48kg chassis that actually fits on a desk. 82nd
- NVLink-C2C and ConnectX-7 let you stack two units for more headroom. 69th
Points faibles
- The Cortex CPU is a weak spot, 20th percentile for this category.
- Thermal limits reportedly choke moderate training workloads.
- Only 1TB of storage, which fills up fast with big models and datasets.
- Port selection is sparse, 15th percentile.
L'avis des propriétaires
The Word on the Street
Les preuves
Performance
The headline number is the 1 petaFLOP of AI performance from the GB10 Superchip. That's the whole point of this machine, and it delivers. The 128GB of DDR5 RAM sits in the 99th percentile, which is what lets you load and fine-tune models that would choke a normal workstation. The GPU with 128GB of VRAM is also strong, 84th percentile. But the CPU is a real weak spot at 20th percentile, and the thermal design seems to hold things back under sustained load. One reviewer mentioned thermal limits preventing moderate training workloads, like fine-tuning a 7B Mistral or 8B Llama. So the peak numbers are impressive, but sustained performance is a different story.
Specifications
Full Specifications
Processor
| CPU | Cortex |
| Frequency | 3.8 GHz |
Graphics
| GPU | Blackwell GPU |
| Type | Discrete |
| VRAM | 128 GB |
| VRAM Type | LPDDR5X |
Memory & Storage
| RAM | 128 GB |
| RAM Generation | DDR5 |
| Storage | 1 TB |
| Storage Type | NVMe SSD |
Build
| Form Factor | mini |
| Weight | 1.5 kg / 3.3 lbs |
Connectivity
| Wi-Fi | Wi-Fi 7 |
| Bluetooth | Bluetooth 5.4 |
System
| OS | NVIDIA DGX OS |
vs Competition
The Mac Studio M5 Max is the obvious alternative for creators and general developers. It's more versatile, has better ports, and Apple's silicon is no slouch. But it can't touch 128GB of unified memory for AI workloads. The Lenovo Legion Tower 5a and HP OMEN GT16 are gaming machines first, with better CPUs and GPUs for gaming but nowhere near the AI memory capacity. The Dell Area-51 is a wild card, big and flashy but not built for this kind of work. If you need local AI fine-tuning, the GX10 is in a class of its own. If you need a general desktop, all of these competitors are better rounded.
| Spec | ASUS Ascent GX10 | Apple Mac Studio M3 Ultra | Lenovo T Series Towers Legion Tower 5a Gen 10 | Dell Desktop ECT1250 | HP OMEN TPC-W062-MT | CLX SET TGMSETRTU5204BM |
|---|---|---|---|---|---|---|
| CPU | Cortex | Apple M3 Ultra | AMD Ryzen 7 7800X3D | Intel Core Ultra 7 265 | Intel Core Ultra 9 285K | Intel Core i9 14900KF |
| RAM (GB) | 128 | 96 | 32 | 32 | 64 | 64 |
| Storage (GB) | 1024 | 1024 | 4096 | 2048 | 2048 | 8192 |
| GPU | NVIDIA Blackwell GPU | Apple M3 Ultra 60-core | NVIDIA GeForce RTX 5070 Ti | Intel UHD Graphics | NVIDIA GeForce RTX 5090 | NVIDIA GeForce RTX 5070 |
| Form Factor | mini | sff | mid-tower | sff | tower | mid-tower |
| Psu W | - | - | 850 | 180 | - | 850 |
| OS | NVIDIA DGX OS | macOS | Windows 11 Home | Windows 11 Home | Windows 11 Home | Windows 11 Home |
| Compare | Compare | Compare | Compare | Compare |
| Produit | CPU | GPU | RAM | Connectique | Stockage | Fiabilité | Preuve sociale |
|---|---|---|---|---|---|---|---|
| ASUS Ascent GX10 | 20.3 | 84.4 | 98.9 | 15.5 | 68.9 | 37.3 | 82.1 |
| Apple Mac Studio M3 Ultra Compare | 99.6 | 76.2 | 95.7 | 93.6 | 53.6 | 99.3 | 86.2 |
| Lenovo T Series Towers Legion Tower 5a Gen 10 Compare | 73.5 | 93.2 | 78 | 88.3 | 98.1 | 70 | 100 |
| Dell Desktop ECT1250 Compare | 88.3 | 58.7 | 81.7 | 74.5 | 82.8 | 70 | 99.1 |
| HP OMEN TPC-W062-MT Compare | 98 | 97.8 | 93.8 | 99.2 | 90.1 | 70 | 67.1 |
| CLX SET TGMSETRTU5204BM Compare | 94.8 | 89.4 | 96.5 | 84.1 | 99.1 | 11.6 | 94.7 |
Prix
Value & Pricing
At $6,995 USD, this is not a casual purchase. But for the right buyer, it's actually a bargain compared to renting cloud GPUs month after month. If you're fine-tuning models regularly, the math works out. If you just want to run inference or dabble, you're paying a lot for hardware you won't fully use. The 1TB drive and limited ports also mean you'll likely spend more on external storage and docks. Value depends entirely on whether you actually need petaFLOP-scale local AI.
En savoir plus
Overview
The ASUS Ascent GX10 is a weird one. It's a tiny box, about 1.48kg, that packs NVIDIA's GB10 Grace Blackwell Superchip and 128GB of unified memory. The pitch is simple: a petaFLOP of AI performance on your desk, no cloud bill required. For AI developers who want to fine-tune 200B parameter models locally, this is basically the only game in town at this size.
But here's the thing. The CPU is a Cortex running at 3.8GHz, which lands in the 20th percentile for this category. The storage is a single 1TB NVMe drive, which is fine but not exactly roomy for big datasets. And the port situation is rough, 15th percentile. This is a specialized tool, not a general-purpose desktop.
Common Questions
Q: Can this actually fine-tune a 200B parameter model?
Yes, the 128GB of unified memory is designed for exactly that. But keep in mind the thermal limits reported by some users, which may slow down longer training runs.
Q: Is this a good gaming PC?
No. The gaming score is 33.5 out of 100, and the CPU is a Cortex chip that's not built for gaming. This is an AI workstation, not a gaming rig.
Q: Can I upgrade the storage or RAM?
The RAM is unified and soldered, so no upgrades there. The 1TB NVMe SSD is the only internal storage, and with the sparse port selection, you'll likely want external drives for larger datasets.
Who Should Skip This
Skip this if you need a general-purpose desktop, a gaming machine, or a creator workstation. The weak CPU, limited ports, and 1TB of storage will frustrate you daily. If you're not actively fine-tuning large AI models, you're paying $6,995 for hardware you won't fully use. Look at the Mac Studio M5 Max or a solid gaming tower instead.
Verdict
Buy this if you're an AI developer or researcher who needs to fine-tune large models locally and wants to avoid cloud costs. The 128GB of memory and GB10 Superchip are the real deal. But if you're looking for a general-purpose workstation, a gaming rig, or even a creator machine, look elsewhere. This is a specialized tool with real compromises in CPU, storage, and ports.
Usage Scores
- Global23.8
- AI/LLM35.7
- Jeux33.5
- Portabilité23.4
- Création30.4
- Professionnel22.3
- Développement25.7
- Home Office25.5
- Workstation28.1