I recently remembered that Intel Optane was a thing for a minute. The shattered dream that was 3D Xpoint is now dead as a doornail. Intel discontinued this in 2022, right before the generative AI boom. But I forgot that Intel made a version of Optane for enterprise data centers that could fit into specialized DDR4 DIMMs, offering terabytes of capacity.
That got me thinking, did Intel discontinue something that allowed for ultra-high-capacity memory on a single system, just as everyone wanted as much memory as possible for AI inferencing? Isn’t that why people are buying maxxed-out Mac Studios, to run local models entirely in memory? Turns out… not really. In its latest incarnation, Intel could put a lot of capacity in Optane DIMMs, but the throughput was way too slow for anything we’d need for AI workflows, 40GB per second compared to 800GB on a modern Mac. So even with that capacity, the token output would be basically unusable.
I’m always fascinated by the technical road not taken. Optane goes on the glorious scrap heap with Transmeta CPUs and Foveon image sensors.