While the internet explodes into a verbal war between everyone it seems, there is a common thread appearing pretty much everywhere and that is the sheer confusion to the SSD situation. Basically in every aspect it is just confusion. It is pretty understandable though because besides when SSDs first really became popular in PCs and started seeing those first big boosts beyond HDDs, there hasn't been much care or thought about storage devices. Who cares right, they hold your games and that is it.
But over the past several years, something started happening. While our GPUs kept getting better, our CPUs kept getting better, everything kept getting better, something odd occured. SSDs were also getting much faster, but what was happening was games were no longer loading any faster. They did at first, why not now?
And this leads into a simple statement Cerny made during his presentation.
Evolution vs. Revolution
While there was this true revolution in POTENTIAL data throughput from the SSDs, there was never an equal revolution in ARCHITECTURE to support this revolution. So you started seeing this giant wall appear stopping any advancements from occuring with PC throughput potential.
YouTube: AMD PCIe 4.0 NVMe vs Intel PCIe 3.0 NVMe vs AMD SATA vs Intel SATA ...
Just go to that comment section and watch the sheer confusion to what the hell is going on. Most pretty much state it is the developers fault. " Programming with HDDs in mind " type thing. I have heard this many times in many places also, but ... No. There is no magical code that developers have been ignoring for some odd reason that would somehow make the process of loading data go from 30 seconds to 3 seconds on PC SSDs. They are just sitting back waiting 10 years for consoles to have SSDs.
So why will the PS5 be able to load games within a second? What sort of special code did Sony develop that will be ran? ( Lol sorry had to ). In reality it has nothing to do with software but everything to do with hardware.
In the PC world, despite having SSDs for 10+ years now, no hardware manufacturer has ever taken it upon themselves to make anything that assists the CPU in the task of loading data. It is still 100% reliant on the CPU to handle everything.
And that was the point Cerny was making. With this revolution in data throughput potential there has to be an equal revolution in architecture to keep up with the data coming in.
Unfortunately when data is pulled off the disc it is not just thrown to RAM immediately and it's done. Boy wouldn't that be great. But no. Many many things have to happen to that data before it can be tossed into RAM or to the CPU caches or to the GPU caches or to wherever the end location is. And since every bit of data needs to be worked on, something has to work on it. Which in PCs right now that is the CPU.
Unfortunately CPUs are like the house wife. Everyone wants them to do everything for them. Run my A.I run my audio run my game scripts run my API commands decode this remap that keep up with player inputs blah blah blah. It's a very busy device and unlike a GPU that has thousands of little workers doing everything all together like a well oiled machine, it has usually 8 workers doing all these different things.
So one thing it absolutely has no headroom for is decompressing and everything else that happens with a file once it is pulled from the SSD. It gets overloaded extremely fast and simply cannot keep up anymore, because it has this whole factory of jobs already queued up and a very small amount of workers to do it all.
Which is where the I/O Complex comes in. Like Cerny stated to decompress 5.5GB/s if you left it up entirely to the CPU that process alone would take 9 current Zen2 cores to handle the workload ( there are only 8 Zen2 cores in the entire PS5/Series X CPU ). The process of directing the data where to go next after decompression would again take yet another 1-2 Zen2 cores if that process was left up to just the CPU ( you go here you go there you go there ). So we now have the need for 10-11 Zen2 cores just to handle the beginning part of the data stream for a 5.5GB/s throughput load of data. Then you have another Zen2 core or two for SSD I/O, Memory Mapping and Coherency which are other steps the data can go through before it is fully ready to be placed in any caches or memory addresses once it's been fully broken down.
And that is the Revolution in architecture that the PC space has yet to receive. There is absolutely no hardware that was ever built to handle the data flow.
The revolution of data throughput was never met with a proper revolution of architecture.
This will become abundantly clear once therse new consoles come out and they are loading everything within 1-3 seconds while PCs are still taking 20-60 seconds to load the same games. It needs that Revolution. When will it occur? I have no idea. I have a feeling though that AMD is actually working on some custom I/O silicon in the upcoming Zen3s and that is why it won't support current motherboards but could be wrong there. If the CPU manufacturers do not do it who will. Motherboard manufacturers? They could put the decompressors and DMAs somewhere close to the NVMe ports certainly. Or perhaps there is room on the SSDs themselves to put the hardware needed? Who knows. But really for everyone to benefit it would need to be upcoming entire families of processors for everyone to receive the benefit. Enter Intel / AMD.
Anyway. Rant over. Just wanted to throw this out there because it actually is a very public example of what Cerny was talking about. Evolution vs. Revolution.