Hacker News Reader

I can't wait until Nvidia finally releases a Tegra based lineup for consumer/prosumer devices.

These days I have a Jetson Agx Orin sitting on my desk. It's mostly for testing CUDA code which will be deployed in the field, but it's a surprisingly capable workstation. Where else do you get 64 GBs of unified memory attached to a somewhat capable GPU?

The Mac Studio, that's where. I believe Nvidia can and will pull of similar feats.

View on HN

> Where else do you get 64 GBs of unified memory attached to a somewhat capable GPU?

AMD has many APUs with decent GPUs, you won't run CUDA on them but the graphics capabilities are there.

ROCm can run on some APUs, and then you have accelerated PyTorch enabling running Whisper and other fun models like that.

Not decent at all in the sense that ROCm never supported any of them.

Oh and sadly limited to 128-bit memory buses. Orin is at twice the mem bw in this regard, hopefully AMD will address that.

If they do, most likely Windows and ChromeOS will be the only ones being supported by OEMs.

They could be their own OEM. They're doing it for the Jetson, they're doing it for their shield devices.

Put a downsized version of the grace hopper chip in there, call it AI-DevBox and it will sell.

Reply on news.ycombinator.com