The latest L1T video shows that drive as having 319MB/s qd1 speeds a 51 µs average access time, how is that possible? That’s 5-7x half of what a 970 evo (which also has a slc cache) can do, and getting pretty close to optane.
I thought that qd1 speed was limited by flash performance? How are they able to improve qd1 speed access time by this much with just controller optimizations?
Edit: those are PCMARK 10 speeds which are absolutely not the same as qd1 speeds. Average access time (which i assume correlates with QD1) is still half of a 970 evo, so the question of how such a low access time is achieved.
I haven’t watched the video (yet) so I may be way off base, but SSDs can have multiple flash chips work in parallel, essentially like RAID0 to increase the total performance of the drive. That could be labeled under “controller optimizations”
I did watch it but the details have left me already, he did poi t out there was 2 sets of… I want to say dram cache? On both sides of the stick. So more of and running in some sort of a raid style speed.doubling along with other improvements?
Maybe it does if they are splitting single 4k blocks over multiple chips, but i very strongly doubt that they are doing that, or that that would even help.
if anything, high capacity nand chips would have bigger block sizes ~ 64k used to be typical.
So, if you’re doing a 4k write with wear leveling, smaller writes turn into reading a page + writing a modified page somewhere and then updating a map (which is usually some kind of log based structure in a preallocated space).
If I was making a firmware and had dram, I’d just make all writes (metadata + data) go to a sequenced preallocated log.
Latency becomes ping-pong over pcie (2 microseconds give or take) + 10-40 microseconds (10microseconds in case of SLC).
So you get roughly 100k write iops per second (or like 350-400MB/s of queuedepth 1).
If you want to do raid stuff across different nand chips, and can afford a tiny capacitor that lasts a millisecond or two, you could go up to a few million iops.
This post was misleading, the stated speeds are PCMARK 10 speeds and not qd1 speeds. The post has been edited to reflect this. The pcmark 10 average access time of 51 µs is still half of that of a 970 evo, but there are apparently ssd’s that come close like the plextor M9P plus with 72 µs according to legit reviews.
But that 51µs is very close to optane’s 42 µs.
So a much smaller difference, but the question of why the average access time is so low still remains, Is that just due to PCMARK 10 not having random access patterns, and the controller being able to better predict & prefetch some of the data?
Don’t think that that’s a limit of optane itself, optane in the memory form factor can apparently achieve access latencies of 350ns (0.350 µs). So the pcie/controller/nvme protocol is probably the limit with optane ssds