[SOLVED] A better "should I upgrade for PCIe 4.0" question

[bumblebee953]

So the general understanding is that since PCIe 2.0, it took until a 2080 Ti to exhaust the x16 lane. This meant these mobos have been future-proof as far as GPUs go, since pre-GeForce 480 days. That's a LOT of generations.
So really the better question is - will it take another 6, 7, 8 generations (about 10 years) for PCIe4.0 to matter? This is a more important consideration for people that are upgrading their mobos now and asking if they should wait or make the switch to team Red.
As graphics cards become more powerful more quickly, is it reasonable to assume they will take up PCIe bandwidth faster? How long is the PCIe 4.0 tech expected to remain "overkill"? If RTX 50s series saturate the PCIe 3.0 x16 slot, then absolutely it is far more wise and economical to upgrade to 4.0 support right now, no?

 

[Darkbreeze]

Currently, there are no graphics cards that saturate the PCIe 3.0 bus. Not even the 2080 ti does. So, without going too far down the rabbit hole, I think it's pretty likely that generational performance is likely to take just as long to get to the point where the PCIe 4.0 bus is saturated or is getting close to it, as it has for the 3.0 bus to do so since moving on from PCIe 2.0.

Maybe the RTX 30 series cards will show us a need for the PCIe 4.0 bus, that's possible, but as it stands right now, existing cards don't show any real benefit compared to 3.0, so I think you are right on the money when you say PCIe 4.0 is still an unnecessary improvement, at the moment, but might not be before long. I would think it's going to be four or five years at least before we even start thinking about whether or not PCIe 4.0 is becoming a choke point or not.

 

[DSzymborski]

By the time PCIE 4.0 is required, you'll likely already buying motherboards with PCIE versions past this. The PCIE version shouldn't be driving this decision. We've only just gotten to the point where a 2080 Ti sees a couple extra frames from being on 3.0x16 vs. 3.0x8.

 

[bumblebee953]

I would think it's going to be four or five years at least before we even start thinking about whether or not PCIe 4.0 is becoming a choke point or not.

I think that is the main consideration here right? For example I've been sitting on my PCIe 2.0 board since 2010. I need to upgrade it soon because my build is currently heavily bottlenecked by my CPU, but for PCIe it's lasted a long time, which makes my old mobo a very good investment all around.

It'd be nice if a PCIE 3.0 board bought today will last that long.

I also haven't even talked about SSDs as apparently that has a bigger impact (or maybe not due to the serial access thing) with PCIe version? What would that be like 5 years down the road?

 

[Zerk2012]

A NVME Drive is already stupid fast. I would not really be concerned about 4.0 unless you were doing a ton of file transfers.

 

[DSzymborski]

I think that is the main consideration here right? For example I've been sitting on my PCIe 2.0 board since 2010. I need to upgrade it soon because my build is currently heavily bottlenecked by my CPU, but for PCIe it's lasted a long time, which makes my old mobo a very good investment all around.

It'd be nice if a PCIE 3.0 board bought today will last that long.

I also haven't even talked about SSDs as apparently that has a bigger impact (or maybe not due to the serial access thing) with PCIe version? What would that be like 5 years down the road?

I don't think I'd buy a PCIE 3.0 board today, but I also wouldn't upgrade from a PCIE 3.0 board I already had unless I had substantial unrelated reasons to upgrade.

 

[bumblebee953]

I don't think I'd buy a PCIE 3.0 board today, but I also wouldn't upgrade from a PCIE 3.0 board I already had unless I had substantial unrelated reasons to upgrade.

Fair enough. But for someone like me that needs to update mobo for an overdue CPU upgrade, would you suggest I either switch to team red or wait till blue gets proper 4.0 support (possibly rocket lake)?

 

[bumblebee953]

A NVME Drive is already stupid fast.

That's only relative. 10 years ago people were saying no computers will EVER need more than 8gbs of RAM. The question is how long will NVMe on 3.0 remain "stupidly fast"?

 

[drea.drechsler]

....
As graphics cards become more powerful more quickly, is it reasonable to assume they will take up PCIe bandwidth faster?
...

There are a ton of negativists who can make their case quite eloquently in the absence of data. Just go watch this EXCELLENT video that asks and answers (or tries to) this exact question as best it can be right now. There simply aren't enough PCIe gen 4 GPU devices to test with but while it may not be a significant performance uplift right now there is some slender potential. Maybe...maybe not.

I think we're soon enough to find out a lot more with the spate of RTX 3000 reviews coming forward. If it's a pressing enough issue just hold on motherboard decisions until after the reviews. Hopefully a few of the more thorough reviewers will provide at least an overview of gen3 vs gen4 performance with the appropriate cards.

That, of course, assumes the test protocols Nvidia signs them up to allow for it. But maybe at least one intrepid reviewer forks over the money for their own gen4 capable 3000 card and protocols be da??ed.

 

[Zerk2012]

That's only relative. 10 years ago people were saying no computers will EVER need more than 8gbs of RAM. The question is how long will NVMe on 3.0 remain "stupidly fast"?

Well about 10 years ago consumer SSD's were just coming out at a very high price and the NVME drives are 3 to 4 times faster.

EDIT for the memory 10 years ago I had 16gb then and 32gb now.

 

[Darkbreeze]

There are a ton of negativists who can make their case quite eloquently in the absence of data. Just go watch this EXCELLENT video that asks and answers (or tries to) this exact question as best it can be right now. There simply aren't enough PCIe gen 4 GPU devices to test with but while it may not be a significant performance uplift right now there is some slender potential. Maybe...maybe not.

I think we're soon enough to find out a lot more with the spate of RTX 3000 reviews coming forward.

This is almost exactly what I said, with a few additional considerations. However, the AMD 5000 series graphics cards ARE PCIe 4.0 cards, and there was not any massive uptick in performance using those on PCIe 4.0 boards, so to say that it's just "negativists making their cases due to a lack of data" is rather a stretch IMO.

They are 4.0 cards and they don't show any substantial performance difference when used on 3.0 or 4.0 motherboards from any data I've seen so far. That does not, of course, mean that higher performance cards that are 4.0 compatible won't show a performance gain, but I need to SEE that in the results before I'm willing to say "Yes, it definitely shows up as a gain by comparison".

Unless I was misunderstanding what you were saying.

 

[Darkbreeze]

That's only relative. 10 years ago people were saying no computers will EVER need more than 8gbs of RAM. The question is how long will NVMe on 3.0 remain "stupidly fast"?

NVME on 3.0 ONLY comes anywhere near the limitations of the bus when performing very large, sequential transfers, between NVME devices that are of similar capability. For random operations, which is where 90% of your OS and applications "live", the throughput capability of the 3.0 bus probably still has several years to go before it would have started seeing problems from lack of pipe. A moot point now since everything new coming out will pretty likely be 4.0 compliant from here on.

And as far as a 3.0 board is concerned, I have a 3.0 board that was purchased in 2011. It is a 990fx Sabertooth board with an FX-8320. I'm repurposing it for use by my nephew as a beginners gaming machine since he doesn't need more than that for now, and am using my equally dated R9 280 in that system for him, but I've used my 2060 Super in it as well, and aside from the CPU limitations, it did just fine on any game that wasn't distinctly CPU bound. So that's lasted about ten years or so. I suspect any decent 4.0 board paired with an appropriately long-legged CPU will be able to do the same if you personal upgrade cycle isn't more along the five year variety like most of ours are.

I think we're going to hit a wall soon anyhow as far as seeing much in the way of dramatic improvements in CPU performance, which we kind of already have, and a lot of improvements are going to have to come by way of storage, memory, cache and refined pipelines unless something surprising comes along from one of the fabs. IDK, I'm no engineer, so not really my area, but aside from AMD catching up recently for the most part with Intel, we really haven't seen anything seriously impressive as far as any of the generational gains by either company that is anywhere along the lines of what we saw when the 2nd and 3rd Gen Intel CPUs came along.

 

[bumblebee953]

when performing very large, sequential transfers, between NVME devices that are of similar capability.

I keep seeing this being brought up. What exactly are large sequential transfers? I fairly often move large movies (mostly 1080 but some 4K quality ones too) between my different storages because I'm constantly running out of room here and there in addition to Steam games taking up space. Does a large transfer of say 20 movie files count as large sequential transfer? It's the one thing that I always wish was a little faster on my devices...(860 Evo, samsung T5, Crucial M4).

 

[hotaru.hino]

I keep seeing this being brought up. What exactly are large sequential transfers? I fairly often move large movies (mostly 1080 but some 4K quality ones too) between my different storages because I'm constantly running out of room here and there in addition to Steam games taking up space. Does a large transfer of say 20 movie files count as large sequential transfer? It's the one thing that I always wish was a little faster on my devices...(860 Evo, samsung T5, Crucial M4).

While the size is likely arbitrary, sequential transfer means just that: the data you want to pluck from the storage drive is all together and the drive doesn't have to jump around as much to grab what it needs. But for the sake of conversation, I think most people will agree that "large" is at least a gigabyte of data for this context.

SSDs are typically made out of NAND type flash. The layout makes it such that addressing is done by the page (anywhere from say 4K bytes to megabytes), then byte, then bit. Because of this, it's actually faster for NAND flash to read data that's together because the SSD doesn't have to jump around as much trying to pull data from one page or another. So while people think SSDs and flash storage is "truly random access" (i.e., access time is the same regardless of where the data is), it really isn't.

 

[Darkbreeze]

I keep seeing this being brought up. What exactly are large sequential transfers? I fairly often move large movies (mostly 1080 but some 4K quality ones too) between my different storages because I'm constantly running out of room here and there in addition to Steam games taking up space. Does a large transfer of say 20 movie files count as large sequential transfer? It's the one thing that I always wish was a little faster on my devices...(860 Evo, samsung T5, Crucial M4).

Anything bigger than 1GB, would be what I'd consider "large". And the larger, if it's a single file, the more you are likely to notice the speed difference. Sequential is generally either one or only a few very large files, while random operations are generally a whole bunch of small to medium sized files. That is equally true whether reading or writing. Random is, well, random. Sequential is typically one big data file. Movies would generally qualify in this regard because the bulk of the movie folder will usually be one, or a few, rather large files, and then maybe a couple of smaller ones that don't take much time to address.

But, you MUST have the same speed capability on BOTH drives to see these kinds of speeds on sequential file transfers, reads, writes, etc. If you are transferring a 5GB file from your NVME drive that's capable of reading a THEORETICAL 3500mbps and writing a THEORETICAL 3300mbps, to a 7200rpm hard drive, you are still going to be limited by the speed of the hard drive. If you are transferring to a 500mbps SATA SSD, you will be limited by that. Only if you have an NVME to NVME transfer would you see anything even approaching those advertised sequential speeds.

So, only when both drives are very fast, will anything you do be very fast. If you are transferring to or from a slow drive from a fast drive you will always be limited by the speed of the slower drive no matter what kind of drives we are talking about.