Quote from: LL on September 24, 2021, 02:27:33
After the CPU core, the most significant improvemnt was the appearance of SSD.
Now we need that working storage and memory be the same speed.
That's not going to happen. Also, what do you mean by speed? Because latency matters a lot and differentiates the different types of memory.
The reality is that a CPU core crunches data a lot faster than storage can supply it and it's probably going to stay that way. You need intermediate layers to bridge this gap. It's just a matter of how many and which technology. We don't need storage to be as fast as operating memory. That would be important if you wanted to remove the operating memory. Take a layer out of the system. But flash-based storage has another weakness - limited durability. Meaning, you don't want to be storing intermediate results in it and induce unnecessary wear. Exactly because memory stores intermediate results, there is higher demand for high throughput and low latency. And because storage doesn't, it doesn't have to be as fast. There are flash based memory devices that are a lot faster than any SSD. They are closer to SDRAM performance. But in order to boost durability, which is very important in an operating memory, they're volatile. Because most of the wear is incurred when making a write persistent. You can use them as a very expensive, but very fast storage where you're taking advantage of the much faster bus (perhaps in a database server). Or as cheaper, non-persistent memory, so that you can install more memory for the same money to aid data crunching.
Sure, fast storage gives us options. And in this, CXL is the big thing. As it has a lot lower latency than PCIe. Big HBM caches on processors are another big thing that should come. But storage as fast as operating memory? Persistent operating memory is an attractive idea but it currently carries a significant cost.