News Intel Arrow Lake processors bottleneck PCIe 5.0 NVMe SSDs by 16%, limiting peak speeds to 12Gb/s instead of 14Gb/s

this performance gap may not be perceptible in everyday usage, but if you are paying a premium for the speed and cant get the full spec out of the component its yet another black eye for intel. At least its not burning out the CPU tho amiright?
 
  • Like
Reactions: bit_user
200 series getting to be more and more an Intel Vista moment but the bad news for Intel, there are real alternatives from of companies, while the alternative for Vista was still a Microsoft product.
 
Ssd transfers need to be reviewed on Linux. Micosoft don't push to the limit always have some interference of windows apps.
Whether or not that's an issue, they're comparing performance between two different CPU + motherboard chipset combos under the same software configuration. So, unless it's a driver problem (which you'd expect Intel to have told them, if true), the OS isn't the issue here.
 
Stop with the lake codename nonsense. Its useless to the consumer. I get that you want to sound informed and technical but literally after 25+ freaking “lakes” i just want to see the “Ultra 100” or “Ultra 200” series or “14th gen” etc. something that actually translates to real world marketing materials. INTEL DOES NOT PUT “ARROW LAKE” ON A LAPTOP BOX. STOP IT.
 
  • Like
Reactions: adbatista
Stop with the lake codename nonsense.
...
i just want to see the “Ultra 100” or “Ultra 200” series or “14th gen” etc.
The first sentence of the article states that they're talking about the Core 200S series CPUs. They further reference specific motherboard chipset numbers, which are an official product designation. Fair enough that they don't mention the CPU family of the Z790 boards, other than by its code name.

FYI, I'd also just mention that they do include a link to the source article, in the second sentence, for those wishing to see more details. There's also a discussion of the CPU models concerned, in the 3rd paragraph from the bottom.
 
Stop with the lake codename nonsense. Its useless to the consumer. I get that you want to sound informed and technical but literally after 25+ freaking “lakes” i just want to see the “Ultra 100” or “Ultra 200” series or “14th gen” etc. something that actually translates to real world marketing materials. INTEL DOES NOT PUT “ARROW LAKE” ON A LAPTOP BOX. STOP IT.
At the moment and most likely for the forseeable future, not even a direct mention makes much difference - if it is Intel and a modern CPU from them, it has a pretty high chance of having some issue attached to it. Now, if it came with news about something the user could do about the situation (like a BIOS update release), then definitely narrow it down and slap it in the title.
 
At the moment and most likely for the forseeable future, not even a direct mention makes much difference - if it is Intel and a modern CPU from them, it has a pretty high chance of having some issue attached to it. Now, if it came with news about something the user could do about the situation (like a BIOS update release), then definitely narrow it down and slap it in the title.

Why are they acting like their a brand new company? Hard to sell products if they have problems.
Nvidia is selling the second generation of GPUs that catch on fire (the power connector) and the first gen that got rid of physics, and they have one driver issue after the other lately, still the most sold gen ever.
AMD used to have CPUs that would outright explode (too much vsoc) , they fixed that and now they don't explode anymore they just quietly die or start failing being able to run your ram. (vsoc is the volts that go to the soc, that controls the ram)
Intel has many issues as well, but don't pretend that intel is the only one with products that have issues out there.
 
Does this problem go away with a normal ARL overclock where the dtd is sped up?
If so then much ado about nothing.
Let's focus on exactly what they said:

"Intel can confirm that the PCIe lanes 21 to 24 PCI Gen5 root port on Intel Core Ultra 200S series processors may exhibit increased latencies compared to the PCIe lanes 1 to 16 Gen5 root ports, owing to a longer die-to-die data path," a statement by Intel published by SSDReview reads. "However, any variations are contingent upon the specific workload and the capabilities of the PCIe endpoint device."

It sounds to me like a simple overclock won't entirely fix the problem. It'd be one thing, if the issue were a bandwidth bottleneck, but latency won't necessarily get fixed like that.

Also, if it were addressed by the recent overclocking mode, I'd have expected Intel to mention that. Maybe not, but I think it's likely they would've.
 
Pretty bad for a new platform but not something I'm going to notice.
Well, it's not exactly a new platform. They originally planned to use this platform for the desktop version of Meteor Lake, which got far enough that they had engineering samples of it. So, perhaps they could've seen these problems a year earlier, if they'd looked.
 
Let's focus on exactly what they said:

It sounds to me like a simple overclock won't entirely fix the problem. It'd be one thing, if the issue were a bandwidth bottleneck, but latency won't necessarily get fixed like that.

Also, if it were addressed by the recent overclocking mode, I'd have expected Intel to mention that. Maybe not, but I think it's likely they would've.
Latency is an issue with 4k random, not sequential read. If you are reading 1GB and measuring the time it takes the latency needed to slow that down would make the same SSD have 4k random run at the speed of an HDD.

And Intel can't respond to everything. If they don't get the response just right the big techtubers will be all over it with their latest Intelbad videos for clicks.

But I have neither an ARL processor nor a PCIE5 SSD so I can't test and know any more than taking a guess.
 
Latency is an issue with 4k random, not sequential read. If you are reading 1GB and measuring the time it takes the latency needed to slow that down would make the same SSD have 4k random run at the speed of an HDD.
I'm a little unclear on your statement, here.

Looking at what they achieved with the PCIe carrier card, I'm reminded a bit of the results I saw, when comparing Raptor Lake's CPU-connected M.2 slot with its chipset-connected M.2 slot. 4k QD1 random IOPS were down substantially, but 1M results were basically unchanged. That's the kind of effect latency can have. Essentially, it's adding in gaps between when one transaction finishes and the next one starts. With larger transactions, that overhead can be negligible. It's with the smaller ones where it adds up.

So, my best guess is that there are multiple issues affecting their CPU-connected M.2 slot, including both latency and bandwidth. By switching to the PCIe slot, they avoid the bandwidth bottleneck (which is the main thing D2D clock should affect) and reduce the latency hit, but it's still not gone.

Sorry, I don't have a link to the Raptor vs. Raptor data I mentioned. If you want, I can try to find it.
 
I think it is more of a bandwidth bottleneck somewhere with the ARL system. Particularly the D2D to the little chiplet:
1746512269_guru3d.webp

Which seems overclockable, even though I don't have one to test: https://skatterbencher.com/2024/10/24/arrow-lake-d2d-overclocking/

As you know I have some Optane which, since is inherently lower latency than NAND, shows off latency issues more when going from the CPU to chipset lanes. I do see a difference, but it is small. And for some reason if I cut down the PCIe gen from 3 to 2 I see a larger drop in random read performance than the source for this article had. If I'm not mistaken I used to see almost a 30% drop in random read performance by running my PCIe gen3 Optane at PCIe gen2 speeds on my older chipset gen2 systems (that could also be Z97 chipset issues, maybe I'll reboot and recheck on this Z790 if it will let me run one of them at gen 2).

But to do a specific test on Raptor Lake I erased my Primocache setup (the 2 small 800p Optane drives were leftover from something I was doing a while back when I just had a 5775c and the release week 900p so I decided to use them to cache my 2 slowest drives, and also had a ramcache on my OS drive which would make the results look crazy if I left all of that enabled) and I ran Crystaldiskmark on them. The 905p has always been the fastest in random read, the 900p the slowest, and the 800p has been real close to the 905p. Maybe the 905p is getting a boost of 10MB/s? by being in the CPU m.2 slot, but that is about it.
Screenshot-178.jpg


And the Optane drives, since they are running 300+ MB/s in random 4k, show the latency is low enough to not be an appreciable bottleneck at 100MB/s just as a 300+fps CPU can show the GPU is not a bottleneck at 100 fps at the same settings.

It would do a better job of showing what the problem is if the source tried Optane in the differently routed slots as well. As that one can't fill the bandwidth, but does expose latency better.
 
OK I rebooted and set one of the 800ps that is on the chipset driver and my Hynix p41 to gen 2. The 800p is easy to see the difference, along with what looks like an increase in latency with reduction of speed over the same copper wires as before with nothing changed but PCIe gen. But I could have swore my 900p ran in the low 200s MB/s 4k q1 on Z97 chipset lanes. Also similar is apparent on the P41, but the bandwidth drop going from gen4 to gen 2 is a lot bigger.

Screenshot-179.jpg


So it could be D2D, but this doesn't prove it. Just is making it look a bit more likely. It could also be an extra D2D jump on the way to the SSD.
 
The 265K is still my choice if I'm building today. The price/performance balance is there now, and while I have a 9100 Pro most people are buying gen4 drives still due to cost. If the 2GB/sec loss is a problem, there's plenty of other options but I think the advantages outweigh the other options from Intel and AMD.