DGX Spark

Depends on what you call "professional grade" The 400 billion model you just need enough unified ram or vram and some decent flops and bandwith. e.g. you can run the 671b deepseek model at 18 tok/sec on a mac studio m3 ultra with 512 gb of unified ram. Price is around 10k I think. No need for a gian

I own a M4 Max MacBook Pro 128GB RAM with a 4TB SSD. I recently bought a DGX Spark. Then returned it a few weeks later. I thought I was buying a smaller, slower Blackwell AI datacenter for my home lab when I bought a Spark. I was wrong. It’s a poorly-supported ARM knockoff of the sm120 ecosystem cal

I can't tell if people are serious when they defend the reason for the DGX Spark existing. I honestly started laughing thinking you were joking about tough workloads training small models til you started comparing and adding defenses and I figured you are being serious... I'm not trying to be disres

DGX is not for anything other than AI, and big models at that. It's a 5070ti speed wise. The 5070ti has 896GB/s memory bandwidth while the DGX Spark and the Minisforum S1 Max have 256GB/s memory bandwidth so it's literally 3.5x slower. You don't buy it for that. You buy it to run 200b parameter mode