> Nvidia didn't ship a 256gb system at sub-500gb/s transfer rate
DGX Spark has 128 GB and only 273 GB/s BW. Are we lucky that NVIDIA did ship something even worse than what you specified? I'm confused.
People have been complaining [1] about how little VRAM NVIDIA ships with their GPUs for decades. Their whole game has been "oh, you want more VRAM? Buy more or pay us 50x for server grade with 10x as much VRAM. The more you buy, the more you save."
Apple did everyone a solid by shipping something way out of that distribution. We now know more than we did before! We know that a 284B parameter model with 13B active params (or 35B with 3B active, or 671B with 37B active) can outperform a 2T model and draw a fraction as much power. How can you think that's a bad thing?
You could point out that Apple didn't invent the idea of MoE. Everyone knows that. But other than Macs, there simply were no machines with >100GB VRAM directly coupled to ~50 TFLOP/s of compute until the DGX Spark last Dec. If you wanted to run a model with more than 32 GB of weights, you had to either pay up for dozens of GPUs idling at hundreds of watts or really pay up for some $50,000 server GPUs idling at... also 100-200W each.
I feel lucky to have a $3k machine on my shelf that can run DS4-Flash with 1M context at 20t/s while drawing ~150W and making very little noise. The best part? It idles at 30W with DS4 loaded, dropping to 6W after a reboot. There isn't a single GPU on the market that can match that in the same shoebox volume.
The DGX Spark is also a niche, arbitrarily limited machine that will not displace serious datacenter workloads. It's targeted directly at the homelab LARPers and arguably a waste of money versus similarly priced GPU clusters. A 256gb Spark at LPDDRX5 transfer rates would be a genuine travesty.
You can try to weasel out any sort of edge justification you want - these are not industry-grade machines. They are slow, expensive, bandwidth-constrained SOCs that don't hold a candle to either datacenter GPUs or even decade-old gaming GPUs. It's worth criticizing when Apple does it, and also worth criticism when Nvidia does it. The only difference being that Nvidia has natural datacenter buy-in, while Apple can't even justify their own hardware in the face of TPU inference costs: https://9to5mac.com/2026/03/02/some-apple-ai-servers-are-rep...
Rust is already well past 1.0. At best an LLM could discover a vulnerability (and the human using it can file a patch) or can help a human improve ergonomics.
At a certain point the level of abstraction / genericization necessary for a big flexible project (like llama.cpp or Linux) blows things up into a huge number of files. Something newer and smaller can move faster.
When I was at the fruit company I became aware of multiple phones starting house fires. Their lawyers would immediately gag the victims and offer $$$$. IIRC it was close to 1 every fortnight I was there.
I have never had an issue with fire with one of my machines, but I have had 3 or 4 of them fail because they detected inconsistencies with their thermistor, immediately removing voltage from heating elements until they can be convinced that the problem has been resolved.
IIRC Some enders shipped with a faulty marlin firmware that had faulty thermal shut off code, and you used to receive a pamphlet with the printer to update the firmware immediately. I suspect that a lot of people simply did not.
Dont know what this bloke did, but between anycubic and the kind of person who carries a flaming 3d printer in his arms, something went very wrong.
Anycubic is a poor quality brand that doesn't really go through the engineering effort to design good quality stuff. These types of printers definitely still exist, no doubt.
But, it's not really straight propaganda that the well designed machines (Bambu, Prusa, and many other vendors) don't have these issues.
You can find equally alarming statements about all sorts of other poor quality goods.
