ASUS Ascent GX10 vs GMKtec EVO-X2 128GB: Recheck the Price Before Buying
TL;DR: Choose the GX10 for a verified CUDA requirement and the EVO-X2 for a general-purpose x86 system. Do not buy on this article’s former claim that the CUDA box is cheaper. On November 5, the GX10 1TB listing was $5,999 and out of stock; a matching current EVO-X2 128GB checkout quote remains unverified.
| ASUS Ascent GX10 (1TB) | GMKtec EVO-X2 (128GB) | Used RTX 3090 tower | |
|---|---|---|---|
| Best for | CUDA development | x86 desktop plus local inference | Workloads that fit in 24GB |
| Purchase evidence, Nov 5 | $5,999 displayed; out of stock | Obtain a selected-variant quote | Obtain a complete-system quote |
| The catch | Arm/Linux compatibility and availability | Verify 128GB variant, runtime, and stock | Card cost is not tower cost |
Honest take: Make CUDA compatibility the reason to choose a GX10; its old low-price argument no longer supports a purchase recommendation.
The ASUS Ascent GX10 and GMKtec EVO-X2 serve different software needs despite their shared 128GB capacity option. Run your target workload through the VRAM calculator before paying for capacity you may not need.
Price correction, November 5, 2026
NVIDIA’s GX10 1TB listing displays $5,999 and Out Of Stock. The ASUS US store page for part 90MS0371-M000A0 also marks it out of stock. A displayed price is not an available checkout quote, and taxes and shipping are not established by this check.
The GMKtec product page initially selects 64GB/1TB at $2,199.99. That is not the 128GB configuration in this comparison. We did not verify a selected 128GB checkout quote, so the old $3,499–$3,649 range is withdrawn as a current price. Do not subtract the 64GB starting price from the GX10 price to create a same-capacity savings claim.
The previous $3,099–$3,400 GX10 range and the resulting “$100–$550 cheaper” conclusion are also withdrawn. The current NVIDIA DGX Spark landing page directs buyers to partners; it does not establish a checkout price in the page we inspected. Earlier Spark prices are not used in the recommendation below.
Scope of this update: the purchasing conclusions and associated arithmetic were corrected. The remaining performance figures are historical third-party reports carried over from the September article, not new measurements or a controlled current head-to-head test. Match the model, quantization, backend, context length, and power settings before relying on those figures.
Same capacity, different silicon
| ASUS Ascent GX10 | GMKtec EVO-X2 (128GB) | |
|---|---|---|
| Chip | NVIDIA GB10: 20-core Arm CPU + Blackwell GPU | AMD Ryzen AI Max+ 395: 16 Zen 5 cores + Radeon 8060S iGPU |
| Memory | 128GB LPDDR5X @ 273 GB/s | 128GB LPDDR5X-8000 @ 256 GB/s |
| Compute | Up to 1 petaFLOP sparse FP4 | 40 RDNA 3.5 CUs + 50 TOPS XDNA 2 NPU |
| OS | DGX OS (Ubuntu-based, Arm) | Windows 11 or Linux, ordinary x86 |
| Software stack | Full CUDA, TensorRT, NIM | ROCm / Vulkan / llama.cpp |
| Rated draw | 240W box | 147–160W measured under dense-70B load |
| Storage | 1TB or 4TB (soldered-adjacent; CID warranty limits swaps) | User-accessible M.2 slots |
Two numbers in that table decide almost everything. The first is memory bandwidth: 273 vs 256 GB/s, a 7% difference. LLM decode is bandwidth-bound, but bandwidth alone does not establish equal measured throughput; software and workload still matter. The second is FP4 compute: the GB10’s petaFLOP-class tensor hardware can matter differently during prefill and decode; peak compute is not an application benchmark.
Inference: a decode tie, a 5× prefill gap
The Register’s head-to-head testing of the Strix Halo platform against the GB10 platform put gpt-oss 120B decode at 34.13 tok/s (AMD) vs 38.55 tok/s (NVIDIA) — a 13% gap that tracks the 7% bandwidth difference plus software polish. On the EVO-X2 specifically, independent testing lands around 31 tok/s on the same model (ServeTheHome measured the identical-chip Beelink GTR9 Pro at 31.41 tok/s and 125–128W; same chip, same speed). StorageReview’s vLLM testing confirmed the GX10 performs identically to other GB10 systems, but a result from one configuration should not be treated as a guarantee for another.
| Workload | GX10 (GB10) | EVO-X2 (Strix Halo) |
|---|---|---|
| gpt-oss 120B decode | ~38 tok/s stock, up to ~60 tok/s (LMSYS tuned SGLang) | ~31–34 tok/s stock, mid-40s–low-50s with llama.cpp Linux tuning |
| gpt-oss 120B prefill | ~1,720 tok/s | ~340 tok/s |
| Dense Llama 3.3 70B decode | ~2.6 tok/s | ~5 tok/s (Q6_K, 147–160W) |
| Qwen3-235B-A22B | fits, bandwidth-bound | ~11 tok/s |
| Fine-tuning | LoRA Llama 3 8B at 53K peak tok/s (StorageReview) | works via ROCm, rougher and slower |
Read that table by row, not by column. Decode: effectively a tie — both feel comfortably interactive on MoE models (human reading speed is ~7–10 tok/s). Prefill: not a tie at all. If your workload is agentic — long system prompts, RAG contexts, coding assistants re-reading a repo every turn — the quoted prefill rates differ by about 5.06× (1,720 / 340). That ratio does not establish a 5× reduction in end-to-end time to first token, which also includes other work. That’s the same conclusion our Spark vs Strix Halo comparison reached, and it’s the workload split that should drive this purchase. If you’re pairing either box with a local coding stack, the prefill row is the one that matters — see the editor-side setups at aicoderscope.com for what those context sizes look like in practice.
Dense 70B models are bad on both (~2.6–5 tok/s). Nothing at 256–273 GB/s fixes that; a Mac Studio M5 Max at 614 GB/s roughly doubles it, and that’s still not fast.
Software: a DGX appliance vs a normal PC
The GX10 runs DGX OS on an Arm CPU. That buys you the full CUDA stack — day-one model support, TensorRT, NIM containers, PyTorch training that just works — with updates that flow from NVIDIA through ASUS validation. It also means the machine is an appliance: no Windows, no x86 binaries, no gaming, no second life as a family desktop. You’re buying it to do AI work, full stop. One more GB10 quirk to know about: the SSD sits under a warranty-void CID sticker, so buy the storage tier you’ll actually need.
The EVO-X2 is an ordinary x86 PC that happens to have 128GB of unified memory. It ships with Windows 11, dual-boots Ubuntu, runs Steam, and can be repurposed the day you upgrade. The AI cost of that flexibility: ROCm on the gfx1151 iGPU still has rough edges (llama.cpp’s Vulkan backend is often the pragmatic choice, and if rocminfo can’t see the GPU you’re in HSA_OVERRIDE territory), and the best decode numbers on Strix Halo come from community Linux tuning, not the out-of-box Windows experience.
Power is the EVO-X2’s quiet win. Reviewers measured 8–14W idle and 147–160W under sustained dense-70B inference; at an assumed $0.12/kWh and 30 days, 8–14W of continuous idle draw costs about $0.69–$1.21 (watts / 1,000 × 24 × 30 × rate). This excludes active workloads and is a calculation from the quoted historical idle range, not a fresh measurement. The GB10 box is rated at 240W and isn’t built around idle frugality. For a 24/7 home server, idle power dominates the annual bill — we ran that math in the power-bill breakdown.
The GX10 bug you should know about (and its fix)
The GB10 platform’s headline expansion trick — pooling two boxes over the 200 Gbps ConnectX-7 link — shipped broken on early GX10 firmware. Owners measured inter-box bandwidth capped at ~13 Gbps because the ConnectX-7 was being power-throttled at 27W on the PCIe bus, which NVIDIA’s forums traced to driver 580.126.09. The fix is a standard package upgrade to 580.142:
$ sudo apt update && sudo apt full-upgrade
# NVIDIA driver 580.126.09 → 580.142, then reboot
$ iperf3 -c 192.168.100.2
# before: ~13 Gbits/sec
# after: ~196 Gbits/sec
Two takeaways. First, if you’re buying a GX10 planning to stack a second one later for 256GB, run apt full-upgrade before benchmarking anything. Second, this is what the ASUS update path looks like in practice: the fix arrived, but through ASUS’s validated channel rather than NVIDIA-direct — the trade we detailed in the GX10 vs Spark piece.
The EVO-X2’s equivalent gotcha is smaller: out of the box on Windows, the default runtime allocates a fraction of the 128GB to the GPU; you set the GPU memory split in BIOS (up to 96GB) and use Linux + llama.cpp for the published-best numbers.
What neither box fixes
A 128GB capacity label does not tell you how quickly a particular workload will run. If your model fits in 24GB, also evaluate a used RTX 3090 system. Compare the complete host, power supply, cooling, storage, and card cost. The previous roughly $1,400 tower estimate conflicted with a similar card-only quote, so neither the “half the price” nor “one third the price” conclusion is retained.
Honest take: what to actually buy
| Your situation | First candidate | Purchase checkpoint |
|---|---|---|
| Your software requires CUDA and more memory than your existing GPU offers | GX10 | Available quote for the exact storage SKU; current listed 1TB offer is out of stock |
| You also need a general-purpose x86 desktop | EVO-X2 128GB | Selected 128GB variant price, stock, and tested runtime compatibility |
| Your workload fits in 24GB | Used discrete-GPU system | Complete-system total and a benchmark of your actual workload |
| You want NVIDIA’s own GB10 system | DGX Spark | Current authorized-partner quote, storage, warranty, and availability |
| You have not tested the model at all | Existing hardware or a small rented trial | Exact GPU capacity and total trial cost before accepting a quote |
If renting, verify the instance’s VRAM and the model’s memory requirement. A rental does not make an oversized model fit, and a headline hourly price does not include every possible storage or transfer charge. We have removed the unverified hourly starting price from this comparison.
FAQ
Is the ASUS Ascent GX10 really the same as a DGX Spark? Same GB10 chip, same 128GB at 273 GB/s, and StorageReview measured identical vLLM performance across GB10 systems. You’re choosing storage tier, chassis, and update path — we compared them in full here.
Which is faster for chat-style inference? Nearly a tie. gpt-oss 120B decodes at ~31–34 tok/s on the EVO-X2 and ~38 tok/s on the GB10 platform stock — both well past reading speed. Tuned stacks raise both (mid-40s–50s on Strix Halo Linux llama.cpp, ~60 tok/s on GB10 via LMSYS’s SGLang build).
Which is faster for agents, RAG, and coding assistants? The historical prefill figures favor the GB10 platform for that specific comparison: 1,720 / 340 is about 5.06. This is a throughput ratio, not a verified 5× improvement in application latency. Test your actual model, context, and backend before making that purchase claim.
Can the EVO-X2 run CUDA software? No. It’s ROCm/Vulkan territory. Most inference stacks (llama.cpp, LM Studio, Ollama) run fine, but CUDA-only tooling — much of the fine-tuning and TensorRT ecosystem — needs the NVIDIA box.
Why not just buy a used RTX 3090 instead? If your workload fits in 24GB, compare a complete used-GPU system before paying for 128GB unified memory. We have not established a current threefold speed advantage or one-third total price for your workload. Total model weights and runtime memory matter, not just the number of active MoE parameters.
Sources
- ASUS Ascent GX10 Review — StorageReview
- ASUS Ascent GX10 Review: A New NVIDIA GB10 Solution — ServeTheHome
- ASUS Ascent GX10 1TB Amazon price history — camelcamelcamel
- NVIDIA GB10 and ASUS Ascent GX10: Specs, Price and Launch Details — RottenWifi
- Nvidia DGX Spark gets $700 price hike to $4,699 — Tom’s Hardware
- Tested: AMD’s Strix Halo vs Nvidia’s DGX Spark — The Register
- Optimizing GPT-OSS on NVIDIA DGX Spark — LMSYS Org
- Performance of llama.cpp on NVIDIA DGX Spark — ggml-org GitHub discussion #16578
- Beelink GTR9 Pro Review (same Ryzen AI Max+ 395: 31.41 tok/s gpt-oss 120B) — ServeTheHome
- GMKtec EVO-X2 AI Mini PC review — Tom’s Hardware
- GMKtec EVO-X2 Review: an AMD Ryzen AI Max 395 Powerhouse — ServeTheHome
- GMKtec releases cheaper AMD Strix Halo desktop with 128GB RAM — Notebookcheck
- Mini PC with AMD’s fastest AI processor and 128GB RAM gets $800 discount — TechRadar
- ConnectX-7 Inter-Spark Link Capped at ~13 Gbps — PCIe Power Throttling (27W) — NVIDIA Developer Forums
- AMD Ryzen AI Halo Developer Platform — AMD
Price and purchase-conclusion update: November 5, 2026. Historical performance reports remain dated to the original September 30 article and were not rerun for this correction.
Recommended Gear
- ASUS Ascent GX10 (128GB, 1TB) — CUDA/Arm option; verify a stocked quote
- GMKtec EVO-X2 (Ryzen AI Max+ 395, 128GB) — select the 128GB variant before comparing prices
- Used RTX 3090 24GB — compare full-system cost and measured workload performance
- NVIDIA DGX Spark — obtain a current authorized-partner quote
Last updated November 5, 2026 for prices and purchasing conclusions. Verify the selected configuration and stock before buying; historical benchmark reports were not rerun.
Was this article helpful?
Thanks for the feedback — it helps improve future articles.
Need hands-on help?
I offer 1-on-1 technical consulting for local AI setup, GPU selection, and AI coding tool configuration — same topics covered on this site.
Book a session — $49 / hour →Get the numbers before you buy
New GPU and mini-PC benchmarks, VRAM thresholds, and price checks — sent when there's something worth acting on, not on a schedule. No spam, unsubscribe anytime.