Bigger nvme, more speed?

[AshesAndWool]

Hey guys, I was having a conversation with my tech-friend who is known for having some questionable sources and plain making stuff up when it comes to certain things he says. We were discussing my new build where I told him I wanted a 1  tb Nvme boot drive over a 250 gb, both because I could use the space, and because Linus has a video where he explains that because bigger SSDs have more NAND modules they also have slightly better speed and longevity because you have more NAND modules to load onto. However he was claiming that this only applies to traditional SSDs for some reason he didn't explain, but I couldn't get why this would be the case because in my mind an m.2 NVME is just an SSD that connects to PCIE x4 instead of Sata. I would really love if someone could tell me if you get the slight increase in speed and longevity on NVME drives or if this only applies to SSDs, thanks.

Video I was referring to: 

 

[NewMaxx]

Yes, I can explain this.

 

A SSD controller's can interface with several channels at once, improving some aspects of performance via parallelization somewhat akin to striping/RAID-0. Consumer SSDs tend to have 2, 4, or 8 channels, with NVME drives tending to have 4 or 8. Within each channel the controller can switch (chip enable) between banks or dies for further improvement gains via interleaving, much as you can benefit from having two ranks of memory per channel in your PC. There are diminishing returns but it's always worthwhile to have at least two per channel, with faster NVMe drives capable of four per channel, or 8 x 4 = 32 dies.

 

Therefore drives do get faster as capacity goes up, but it's dependent on the controller and flash density. For example, Intel's QLC is 1Tb/die (128GiB/die), which means the Intel 660p doesn't hit peak performance until 1TB - 4 channels on its SM2263 controller, with two dies per channel - while the Rocket Q with that flash needs 2TB - 8 channels on its Phison E12(S) controller, two dies per channel. TLC is more commonly 256Gb/die (32GiB/die) although current 96L TLC is often 512Gb/die (64GiB/die). So a drive like the SATA-based Crucial MX500, with its four-channel SM2258 controller and 64L TLC, gets good (relative) performance even at 250GB (256GiB of flash): 4 (channels) x 2 (dies/channel) x 32 (GiB/die) = 256GiB.

 

Also, it's possible to use the old AHCI protocol (vs. NVMe) over PCIe, but this is uncommon. SATA and PCIe are interfaces while AHCI and NVMe are protocols, technically speaking. M.2 is a form factor with keying, so therefore also an interface similar to SATAe (SATA Express) whereby it can handle both SATA and PCIe interfaced SSDs and also either AHCI or NVMe protocol'd drives.

[AshesAndWool]

3 minutes ago, NewMaxx said:

Yes, I can explain this.

 

A SSD controller's can interface with several channels at once, improving some aspects of performance via parallelization somewhat akin to striping/RAID-0. Consumer SSDs tend to have 2, 4, or 8 channels, with NVME drives tending to have 4 or 8. Within each channel the controller can switch (chip enable) between banks or dies for further improvement gains via interleaving, much as you can benefit from having two ranks of memory per channel in your PC. There are diminishing returns but it's always worthwhile to have at least two per channel, with faster NVMe drives capable of four per channel, or 8 x 4 = 32 dies.

 

Therefore drives do get faster as capacity goes up, but it's dependent on the controller and flash density. For example, Intel's QLC is 1Tb/die (128GiB/die), which means the Intel 660p doesn't hit peak performance until 1TB - 4 channels on its SM2263 controller, with two dies per channel - while the Rocket Q with that flash needs 2TB - 8 channels on its Phison E12(S) controller, two dies per channel. TLC is more commonly 256Gb/die (32GiB/die) although current 96L TLC is often 512Gb/die (64GiB/die). So a drive like the SATA-based Crucial MX500, with its four-channel SM2258 controller and 64L TLC, gets good (relative) performance even at 250GB (256GiB of flash): 4 (channels) x 2 (dies/channel) x 32 (GiB/die) = 256GiB.

Wow, that is way more advanced than I thought this would be. Thank you so much for the explanation, it explains why the speed of the nvmes I am looking at increases by a lot from 250 to 500 gb, slightly from 0,5 to 1 tb, but doesn't change from 1 to 2 tb. 

[Enderman]

1) NVME drives ARE ssds.

 

2) A modern SSD already is going to live longer than you are, so the 'extra longevity' is only important for cases like datacenters where you're writing to them terabytes of data 24/7.

 

3) The speed increase is a few percent, nothing you'll be able to notice. If you don't need 1TB of space then don't waste money buying a bigger drive. An NVME drive is already faster than you need anyway, you would even be fine buying a regular sata SSD. Most programs and stuff won't see a difference.

[NewMaxx]

7 minutes ago, AshesAndWool said:

Wow, that is way more advanced than I thought this would be. Thank you so much for the explanation, it explains why the speed of the nvmes I am looking at increases by a lot from 250 to 500 gb, slightly from 0,5 to 1 tb, but doesn't change from 1 to 2 tb. 

The majority of consumer usage is small, random I/O, which does not benefit from parallelization. Usually 4K as that's a typical cluster size. A modern TLC SSD is probably writing pages (the smallest unit they can write) of 16KB, breaking your (larger) writes into page-sized sub-requests, which means the benefits from a larger SSD aren't necessarily significant. Although of course, larger drives also have larger SLC (write) caches and often more DRAM among other things which can improve performance. That being said, sequential performance and that at higher queue depths (more I/O in the pipeline) absolutely benefit from more dies over NVMe especially.