10th/11th gen intel performing better than 13th in regards to latency?

[iameatingjam]

Thoughts? Is there anything to this? I mean, I guess it doesn't matter for most use cases but still, coming from 11th to 13th gen, its not exactly exciting to think it was more of a side-grade when you consider io. And could this perhaps be a factor in that thread I saw here last week showing intel getting worse ssd numbers than amd?

Also the loss of avx-512 sucks when I do a lot of ps3 emulation. At first I heard 13th gen was still faster even without it, but now I hear its more up in the air and varies from game to game. Though the games that I've played so far seem to work fine.

 

[Lew Zealand]

I just watched the video and Brian clearly states that this is an edge case though one that he directly experienced while editing his videos. And an Intel engineer at Computex acknowledged how it could happen without explaining any tech details. So it's something that the average user will probably never see but some people could run into it if they have a high-end previous system to compare with like Brian had.

I wouldn't be concerned about it at all if I had a 13900K.

 

[evernessince]

It's an edge case for sure but we'll see when Intel moves to a chiplet based architecture if latency gets worse as they did for Zen 1.

The bigger question might be, when is Nvidia going to fix the DPC latency issue that's been in their drivers for two decades now? I have the exact issues as brain in the video.

 

[A Computer Guy]

It's an edge case for sure but we'll see when Intel moves to a chiplet based architecture if latency gets worse as they did for Zen 1.

The bigger question might be, when is Nvidia going to fix the DPC latency issue that's been in their drivers for two decades now? I have the exact issues as brain in the video.

I believe Zen2 introduced the chiplet architecture for AMD Ryzen.

 

[R-T-B]

when is Nvidia going to fix the DPC latency issue that's been in their drivers for two decades now?

They supposedly have that pushed for next release. They made a big blog kerflaw about it and everything.

 

[N/A]

Could it be that the latency is added because of Nvme running in Gen4 mode?
Also 10th gen CPU doesn't have a dedicated PCIE lanes for Nvme in the CPU. All data is going through the QPI (PCIe4x) link and that will lower the overall bandwidth and add latency as well when bottlenecked.
Downgrading to it makes no sense, for productivity just use the CPU attached Nvme and lower the PCIe speed to gen3 should also lower the latency.
Whatever driver IO or clock generator moved to the PCH is unbelievable how it would affect things. Is data waiting for some feedback from it?

 

[evernessince]

I believe Zen2 introduced the chiplet architecture for AMD Ryzen.

Zen 1 was chiplet, Zen 2 was just a further modularization. You had Zen 1 Threadripper parts that scaled all the way up to four chiplets.

They supposedly have that pushed for next release. They made a big blog kerflaw about it and everything.

Hopefully they actually do fix it. I've been waiting far far too long.

 

[tabascosauz]

After 2 videos I'm still not convinced Bryan's 100% ruled out his own builds and Windows installs?? I appreciate that he's put some effort into establishing commonality between the 10850K and 13900K systems but there's not much extra effort made to rule out just about anything else.

OC, mem and BIOS settings? What BIOS? How many times did he reinstall Windows to repeat the test on either setup? Exactly what sort of stuff he's loading Windows down with? Windows builds being used? Did he allow Windows Update to auto-fetch any platform-related drivers? Exactly how many reviewer anecdotes did he collect corroborating (for example) the specific Office mouse jumpiness bug, and what sort of specs? Regarding DPC latency tests, are we ruling out Nvidia as the culprit during every test? Regarding random observed "slow Windows" symptoms, they pretty much remain completely anecdotal, and no word as to how consistently reproducible they are.

Being a self-described "edge case", if ignoring all other potential factors that have not been eliminated, seems like most of the behaviour can still be described as Intel still needing to fine tune Thread Director behaviour for applications like Premiere, and maybe MS has work to do as well. No need to go off the deep end with vague speculation regarding "I/O driver being moved off the CPU"?

Core-to-core latency, negative effect of E-cores, and the Alder Lake/Raptor Lake ringbus are largely a known factor as well. AT and Chips&Cheese both covered this stuff a long time ago. The overall latency picture for the ringbus is not rosy but not horrible (progressively worse starting from 8th gen), and overall AMD runs a tighter ship within the CCX. E-cores being online does have some perceptible negative impact on stuff like cache performance, but all in all nowhere near the severity of the symptoms Bryan is showing.

He's promised more testing and maybe even Ryzen testing so I guess I'll keep waiting on this to unfold.

But man........linking screenshots of random "X feels snappier than Y" forum posts (including from TPU) is pretty funny. I thought he's trying to convince everyone that all of this *isn't* just seat-of-the-pants feeling? And the "statement" he got from Intel sounds like nothing more than a vague red herring that the Intel rep threw at him to stop bothering them - doesn't sound like a promise that Intel is aware or investigating or doing anything really.

And I hope Bryan doesn't need a reminder how much of a dumpster fire 11th gen was. Even if latency is the cause, 11th gen certainly isn't the saviour here.

 

[A Computer Guy]

Zen 1 was chiplet, Zen 2 was just a further modularization. You had Zen 1 Threadripper parts that scaled all the way up to four chiplets.



Hopefully they actually do fix it. I've been waiting far far too long.

Ah yes I forgot about threadripper.

 

[iameatingjam]

And I hope Bryan doesn't need a reminder how much of a dumpster fire 11th gen was. Even if latency is the cause, 11th gen certainly isn't the saviour here.

I had 11th gen, not him. He's going for 10th gen for video editing. But your points are fair.

 

[tabascosauz]

It's an edge case for sure but we'll see when Intel moves to a chiplet based architecture if latency gets worse as they did for Zen 1.

The bigger question might be, when is Nvidia going to fix the DPC latency issue that's been in their drivers for two decades now? I have the exact issues as brain in the video.

Probably Navi31's fanout link is a better comparison for EMIB. IFOP is super slow. Maybe the cores becoming a compute tile can shed some weight, allowing intra-tile latency to be good again like in Skylake.

I think I remember reading in AT's piece on 13900K that E-core communication takes this weird slow route that probably explains why the latency is closing in on IFOP-level bad when E-cores are involved. Of course, that by itself still doesn't explain what's going on in the video.

https://twitter.com/i/web/status/1500609374414417928

I had 11th gen, not him. He's going for 10th gen for video editing. But your points are fair.

I see, my mistake. It's not my intention to discredit Bryan. He's easily proved that the alleged problems are real; it's just that there's still a long way to go to establish any real link to what he thinks is the culprit.

 

[aaa社保了啊]

I think it has to do with memory performance and latency [architecture] introduced by CPU caching. Of course, this is also related to SSD performance. Daily data opening speed is also tied to SSDs.