Nvidia's Project Digits is a 'personal AI supercomputer'

Finally a real ARM workstation from Nvidia! This will be much faster than Apple's offerings for AI work. And at $3000 it is much cheaper than any Mac with 128 GB RAM.

On the CPU size the Neoverse N2 doesn't compete particularly well with apple's M4, or the Zen5 for the matter.

Bit bit hard to tell what's on offer on the GPU side, I wouldn't be surprised if it was RTX 4070 to 5070 in that range.

If the price/perf is high enough $3k wouldn't be a bad deal, I suspect a Strix Halo (better CPU cores, 256GB/sec memory interface, likely slower GPU cores) will be better price/perf, same max ram for unified memory, and cheaper.

AI work happens predominantly on the GPU, not the CPU. This GPU with CUDA will run rings around M4 with MLX. And with much more RAM than you can get in a Mac for $3k.

A lot of people have been justifying their Mac Studio or Mac Pro purchases by the potential for running large AI models locally. Project Digits will be much better at that for cheaper. Maybe it won't run compile Chromium as fast, but that's not what it's for.

It’s actually “10 Arm Cortex-X925 and 10 Cortex-A725” [0]. These are much newer cores and have a reasonable chance of being competitive.

[0]: https://newsroom.arm.com/blog/arm-nvidia-project-digits-high...

The quotes I've seen mention the maximum config (128GB ram and 4TB of storage) and the minimum price. Nothing saying $3k for 128GB ram and 4TB of storage. I hope I'm wrong, but I'm betting the max price is at least twice the minimum price.

NVidia says 128GB ram at $3k[0], it looks like the 4TB storage might be variable (and possibly CPU or GPU cores?). This article says 128GB too.. but used up to twice in a row with different meanings which doesn't help.

[0]: https://nvidianews.nvidia.com/news/nvidia-puts-grace-blackwe...

It's GB200 in desktop compatible enclosure, the RAM is fixed, the SSDs are not, the network ports are fixed too.

It’s most definitely not a GB200 in a desktop enclosure.

The processor is using completely different cores, and the GPU is somewhere around a 5070 for TOPs.

The CPU is apparently the result of a "secret" project that wasn't on published roadmaps. It's called the GB110. So maybe they will offer differently binned CPU/GPUs with a different fraction of cores disabled and you can pick your SSD.

Remember: inference is memory bound.

https://www.databricks.com/blog/llm-inference-performance-en...

It's GB10 a much cut down version to fit the price point, and space, weight and power requirements.

This is NVIDIA, not Apple. They don't charge you a RAM tax (at least for this product). There is only one config for RAM: 128GB.

It's been far from clear what config options are going to be available, and the $3,000 price is the "starting at" price. Not sure what the options will be, but people have collectively found statements that imply all configs will have 128GB ram. Sounds good, I hope it's true.

Seems like the storage will have options, because it's "up to 4TB". Unsure if there will be differently binned CPUs (clock or number of cores). Or connectX optional or at different speeds.

Doubt it. I don't remember NVIDIA ever doing binning and having different SKUs for their Jetson Developer Kit line, which is similar to this Project Digits thing.

Well presumably there's some different configurations supported, otherwise they would say $2,999 instead of starting at $2,999.