N1X, which is based on this chip, is increasingly looking DOA.

I mean, seriously, mediatek WiFi on device that costs this much? :/

Then there's the fact that it has high idle power consumption even with arm cores? It's almost like Nvidia didn't bother about power efficiency at all since they initially were aiming it at AI.

What does that tell us about the n1x which this is based on in terms of efficiency?

Then there's the fact that Nvidia were found recently attempting to hire proton and linux on arm vulkan Devs, literally weeks ago.

Like isn't it a little late for that? Shouldn't this have been done *years ago* if they were planning to release something this year in consumer market?

Man, what a mess. Feels like Nvidia were caught with their pants down again.

This is neither here, nor there:
128 GB RAM is DOA, the price doesn't even matter. Had it at least 192 GB RAM, then it could fit the very usable huggingface.co/unsloth/Qwen3.5-397B-A17B-GGUF 3-bit quant and many other usable quants of other LLMs that don't fit into 128 GB RAM.

Who wants to pay 4000, but at the same time doesn't have enough space to place a proper, much faster, system for cheaper (DOA again?):

QuoteThe centerpiece of the ZGX Nano G1n is the Nvidia GB10 chip, whose performance profile is roughly based on an Nvidia GeForce RTX 5070

Rhetorical question: Can one build a (much) better/faster PC for cheaper? A PC will also allow for upgrades.

Also, isn't the OS Linux-based, but still proprietary/closed (this is the least of its problems)?

QuoteFrom Expensive Toy to AI Supercomputer

This is a 5070 chip, connected to 128 GB, 256-bit memory bus (aka quad-channel (4*64-bit) if this were a desktop PC), RAM, but only 273 GB/s, and indeed:

Quotememory bandwidth noticeably slows down GPU performance

A normal desktop 5070 card has 672.0 GB/s (this one has only 273 GB/s).

The only advantage this has over a desktop PC is that a normal AM5 B850 midrange desktop PC motherboard has a 128-bit memory bus (aka dual-channel (2*64-bit)) and the memory of this runs at over 8533 MT/s (vs 4800 to 5600 MT/s in a desktop (when at 256 GB RAM)), but the advantage of the B850 is that it supports up to 256 GB RAM.

This (128 GB RAM): 273 GB/s = 256-bit * 8533 MT/s / 1000 / 8.
Desktop PC (258 GB RAM) (Zen 4 CPU): 76.8 GB/s = 128-bit * 4800 MT/s / 1000 / 8.
Desktop PC (258 GB RAM) (Zen 5 CPU): 89.6 GB/s = 128-bit * 5600 MT/s / 1000 / 8.

These systems kinda need to double in RAM, GPU prompt processing performance, and hopefully not in price, to become interesting.

Utterly ridiculous claim. No GPU comes with 512GB and even 512GB is not sufficient for full precision training of very large models, so why do you need 512GB? 128GB for a GPU on a sub-£4000 workstation is unmatched by any other device - it is in a league of its own.

shouldve made it 512GB no matter the cost..128GB is a joke in LLM scene