NVME lanes (Chipset vs Processor)

[aQi]

Hello TPU,

I need to know some logical explanation please. Rocket Lake is coming with dedicated x4 lanes from processor to NVME. What is the major difference if the same lanes are attached to PCH instead of processor ?

Will there be any significant change ? Even if there is a change, what is it beneficial for ?

 

[X71200]

x4 is x4, what really matters is how much of the drive is filled and how it's been conditioned (Windows TRIM and stuff).

As long as you get 3.0 out of those lanes, both sides - let's say a PCI-E M.2 adapter VS plugging the stick on the board, should offer you similar speeds.

This of course applies if you have a setup similar to what I have in mind (some mid range NVMe drive).

 

[TheoneandonlyMrK]

Hello TPU,

I need to know some logical explanation please. Rocket Lake is coming with dedicated x4 lanes from processor to NVME. What is the major difference if the same lanes are attached to PCH instead of processor ?

Will there be any significant change ? Even if there is a change, what is it beneficial for ?

An nvme connected straight to the processor will always have that bandwidth available to it.
One attached to the PCH will incur a very slight latency increase and also runs a slight risk of being bandwidth starved in some scenario.
Rocket lake Ups the PCH to CPU link to x8 so it should pretty much mitigate this issue.
And the scenario to clog the PCH to CPU bus doesn't happen often.
Since few people have enough GPU or nvme storage attached to instigate the issue.
I had issues with raid 0 on three nvme due to this so it's not impossible.

 

[oxrufiioxo]

On Z590 the majority of the boards only have 1 gen 4 m.2 slot with some high end boards having 2 so depending on what drive you're using could make a big difference speed wise.

Assuming you're using a gen 3 drive it shouldn't make much of a difference though. I tend to always keep my OS in a slot directly connected to my cpu and any game/data drives going through the pch.

 

[Aquinus]

It's worth noting that the lanes off the PCH are shared with everything else that's off the PCH and that this all uses a shared DMI bus between the CPU and PCH. Generally speaking, using the lanes directly connected to the CPU will offer the best speed and latency. In reality, it probably won't make a perceivable difference outside of benchmarks.

 

[ThrashZone]

Hi,
20 lanes is all she wrote so second m.2 would matter and going off pch/ sata will just disable a couple sata ports.
Speed wise not much difference but you wouldn't want two m.2 in pci-e mode seeing you'd eat up gpu lanes ideally 16 lanes for it.

 

[aQi]

So technically the lanes from CPU specifically for NVME exclude the sharing issues of which PCH has. In real world i think the difference is not noticeable at all if someone here talks about latency. None the less if someone can share results perhaps.
I just noticed the lanes from CPU lacks RAID 0 support anyway.

 

[X71200]

NVMe RAID is kinda pointless, you're not helping your 4k speeds and the drives already do sequentials high. The whole point of M.2 is this, flexibility in a small form.

 

[aQi]

NVMe RAID is kinda pointless, you're not helping your 4k speeds and the drives already do sequentials high. The whole point of M.2 is this, flexibility in a small form.

On point

 

[ThrashZone]

Hi,
Yeah you really have to look at the manual to see what happens when using 2-m.2's
Could have gpu x16 1st m.2 at x4 and 2nd m.2 falls to x2 both being pci-e mode...

I love apex and dimm_2 slot for m.2 it's pch period

 

[kapone32]

NVMe RAID is kinda pointless, you're not helping your 4k speeds and the drives already do sequentials high. The whole point of M.2 is this, flexibility in a small form.

That depends entirely on what you do with your PC. For general use though you are right.

 

[aQi]

But wait there is more, what about those x16 raid cards out there ? and does the read write speeds in return benefit the loading times of softwares especially games ?

 

[kapone32]

But wait there is more, what about those x16 raid cards out there ? and does the read write speeds in return benefit the loading times of softwares especially games ?

I have an Asus M2 card with 4 660P 1TB. It depends on the Game. Where I find my drive shining is in RTS Games like Total War.

 

[Aquinus]

Where I find my drive shining is in RTS Games like Total War.

Are you saying that you can notice a perceivable improvement between using one NVMe and 4 in RAID? Is that just coming from your gut or do you have numbers to support that? I found very little difference between SATA RAID-0 of two SSDs and a single NVMe drive, so I question the improvement going from one NVMe to 4 in RAID. I'm sure synthetic benchmarks show the gain, but I seriously doubt that load times are perceivably different. The only use case I can imagine this to be used for is a database server or something along those lines. Even that might be a stretch.

 

[kapone32]

Are you saying that you can notice a perceivable improvement between using one NVMe and 4 in RAID? Is that just coming from your gut or do you have numbers to support that? I found very little difference between SATA RAID-0 of two SSDs and a single NVMe drive, so I question the improvement going from one NVMe to 4 in RAID. I'm sure synthetic benchmarks show the gain, but I seriously doubt that load times are perceivably different. The only use case I can imagine this to be used for is a database server or something along those lines. Even that might be a stretch.

I have an X399 build. I have 2 2TB SATA drives in RAID 0, a single SATA drive, a 5tb HDD a single NVME and the RAID array. Using all of those drives FPS games load fastest on the SATA RAID. In terms of performance difference it is during Turns that you will see it most. Having said that you have to be able to test as close as possible to notice the difference as a single NVME is plenty fast but then that is 1 XPG S11 1TB vs 4 660P and there is a marked difference between those 2 in performance.

I have an X399 build. I have 2 2TB SATA drives in RAID 0, a single SATA drive, a 5tb HDD a single NVME and the RAID array. Using all of those drives FPS games load fastest on the SATA RAID. In terms of performance difference it is during Turns that you will see it most. Having said that you have to be able to test as close as possible to notice the difference as a single NVME is plenty fast but then that is 1 XPG S11 1TB vs 4 660P and there is a marked difference between those 2 in performance.

The first slide is my 660P NVME drive and the 2nd is my 4TB SSD RAID 0

 

[Aquinus]

I'm sure synthetic benchmarks show the gain, but I seriously doubt that load times are perceivably different.

This is what I was trying to get at.

 

[kapone32]

This is what I was trying to get at.

What I was showing is IOPs. Which do matter

 

[Aquinus]

What I was showing is IOPs. Which do matter

Yes, but I wouldn't say that it demonstrates this. I find the following claim to be anecdotal. Loading games isn't a pure I/O workload, so there is a point of diminishing returns when it comes to making a disk more and more fast. I'm questioning it because I didn't personally feel that the difference between SATA RAID and a single NVMe to really be all that different.

Where I find my drive shining is in RTS Games like Total War.

My point is that I don't think it's really that much faster than you think it is and synthetic benchmarks aren't going to reflect real world use.

 

[kapone32]

Yes, but I wouldn't say that it demonstrates this. I find the following claim to be anecdotal. Loading games isn't a pure I/O workload, so there is a point of diminishing returns when it comes to making a disk more and more fast. I'm questioning it because I didn't personally feel that the difference between SATA RAID and a single NVMe to really be all that different.

My point is that I don't think it's really that much faster than you think it is and synthetic benchmarks aren't going to reflect real world use.

Loading Games is one thing. Ending turns in an RTS Game is a completely different animal. As an example in the early stages of TWWH2 Mortal Empires there are over 100 factions that need to do exactly what you do in the campaign when you end your turn. For me it doesn't matter though as I have proven it to myself with the amount and types of storage I have used. As an example when I had a OCZ Revodrive as my main Shogun 2 gave me better FPS by about 18+ FPS vs SATA. I know that SATA is plenty fast and as I have stated before I know that for most Games it won't matter but I also know that Total War is one Game that cares about the type of storage you use. Try to load a 140 turn Campaign from a HDD and you will know what I mean.

 

[JcRabbit]

So technically the lanes from CPU specifically for NVME exclude the sharing issues of which PCH has. In real world i think the difference is not noticeable at all if someone here talks about latency. None the less if someone can share results perhaps.

I have a 960GB Intel Optane 905P NVMe drive and initially I wondered the same thing. I asked Intel and they said it would not make a difference, the max bandwidth of the drive itself being the limiting factor.

Anyway, I never-the-less tried connecting it to the PCIe CPU lanes (GPU going to 8x) and run some tests. While this was a couple of years ago and these tests were NOT exhaustive or even very conclusive, I seem to recall that Intel was right and at best there was no practical difference (as measured by CrystalDiskMark), and it might actually be a bit WORSE when copying data between drives (my other 2TB Samsung 970 EVO M.2 drive was still connected to the PCH, and with BOTH drives directly connected to the PCH, copy operations seemed to perform just a little bit faster, IIRC) .

Anyway, my conclusion was that it was definitely not worth bottlenecking the GPU to 8x instead of 16x, so I moved the Optane drive back into the PCIe slot connected to the PCH, where it stays to this day.

 

[rdkone]

I just moved my WD Black 750 1Tb nvme from chip based connection to cpu connection and per a popular HD speed tester results as follows (x570 asus tuff board, 5800x cpu)

Sequential Read : chipset interf. : 2453.001 MB/s, cpu interface read 2646 mb/s
Sequential Write : chipset interf. : 2864.501 MB/s, cpu interface write 3094 mb/s

This isn't all results but generally speaking there's obviously a gain.

 

[aQi]

I just moved my WD Black 750 1Tb nvme from chip based connection to cpu connection and per a popular HD speed tester results as follows (x570 asus tuff board, 5800x cpu)

Sequential Read : chipset interf. : 2453.001 MB/s, cpu interface read 2646 mb/s
Sequential Write : chipset interf. : 2864.501 MB/s, cpu interface write 3094 mb/s

This isn't all results but generally speaking there's obviously a gain.

That is 200+ MB/s on read and write but WB Black 750 1TB advertises it to be 3470MB/s on read and 3000MB/s on writes. Are we missing something here ?

 

[TumbleGeorge]

That is 200+ MB/s on read and write but WB Black 750 1TB advertises it to be 3470MB/s on read and 3000MB/s on writes. Are we missing something here ?

Real live vs advertising

 

[thesmokingman]

Real live vs advertising

The settings also matter and the speeds that WD advertises doesn't list what and how they got it.