AMD - Zen chitchat

[Corky34]

As I understand it, AMD's results are valid (they don't appear to be affected by this HPET implementation 'bug'). Also, other sites should be unaffected as this is down to the way Anandtech test.

Yes AMD results are valid AFAICT.

I'm not even sure it's a bug TBH as despite them saying some overclocking/benchmark programs force HPET on it seems to be more a snafu of Anandtech's own making.

Maybe I'm being to harsh on Anandtech and forcing Windows to rely on a single timer is a common thing for overclocking/benchmark programs to do but IMO I'd consider that to be a bad practice, it's not the usage of HPET that i take issue with it's the forcing of Windows to only use a single timer.

[watercooled]

I say 'bug' in inverted commas, but it does seem like Intel's implementation causes problems that don't happen on the AMD platform; the difference in Intel goes up to 76%.

[Corky34]

Apologies i missed the inverted commas.

TBH i wouldn't consider it a hardware related 'bug' as IIRC even in the days of using the RTC clock there were additional checks done on timings, besides if it is a 'bug' with Intel's implementation it's unlikely to be a purposeful bug as the difference results in a 76% lower result in some cases, i.e when using more than one source for timings their results increase, at least i think that was the conclusion I'm in a bit of a rush so haven't dbl checked.

[scaryjim]

... Maybe I'm being to harsh on Anandtech and forcing Windows to rely on a single timer is a common thing for overclocking/benchmark programs to do ...

3DMark not only requires use of HPET for "valid" results, but refuses to run in default configuration if HPET isn't available: you specifically have to tell it not to use HPET if you can't enable it (my new laptop has no option to enable HPET, for example).

[Corky34]

I think there maybe some confusion, enabling HPET in the BIOS only enables/exposes an additional timing mechanism for software to use, setting HPET in the BCD file for Windows tells it, and by extension any software running on it, to only use the HPET timer, 3Dmark requiring it to be enabled in the BIOS for valid results is understandable, setting Windows to use only HPET for its timing is what would be wrong.

[scaryjim]

... 3Dmark requiring it to be enabled in the BIOS for valid results is understandable, setting Windows to use only HPET for its timing is what would be wrong.

Yeah, I was just pointing out a very common piece of benchmarking software that pushes you onto HPET whenever possible. It's indicative of a general expectation of/reliance on HPET in benchmarking, and I can totally understand why a reviewer whose entire job is benchmarking CPUs would prefer to choose their own timers and would use the most accurate one available. It's a perfectly understandable mindset, and it's not something you'd expect to have an incidental side-effect of tanking game performance by over 40% in some cases.

The additional performance hit Intel are seeing here is basically exactly what you'd expect from an IO intensive task following the application of meltdown patches. It's a bit edge case, and I think it's right that AT are changing the benchmarking suite this year to use the default enabled, not forced HPET settings now they've discovered that forcing it causes anomalous results on Intel CPUs. OTOH, I also think the decision to force HPET in the first place was valid, reasonable, and sound. And it has highlighted an issue that Intel have either not known about or have chosen not to disclose. I'm disinclined to lay any blame on AT for this one.

[Corky34]

I'm not a reviewer and i don't run (m)any benchmarks so can't pretend to know the mindset of someone whose job it is to do that, it would drive me nuts TBH.

What i was trying to say is that Windows and AFAIK all software has typically never relied on just a single timer, normally to get an accurate measurement of time you use two or preferably more timing sources and constantly run checks against the two with highest resolutions, back in the day IIRC that was the RTC and the PCI clock generator.

I understand what you're saying about testers wanting to control all the variables and be as accurate as possible, and I'd personally put this down to one of those things, blaming anyone for it is perhaps a bit strong but it was interesting to read what caused it and reassuring that there are testers who go to such lengths, i think that speaks volumes about Andantech's CPU guy.

*What i find most interesting is the unanswered questions, if the HPET timer does throw a wobbly on Intel systems why hasn't it been picked up in other OS' like Linux, is it caused by Windows like what we saw (iirc) when Ryzen system came out of sleep and people thought they could OC them more but it ended up being a timing issue, if Windows has been automatically discounting the HPET timing as unreliable in some situations how come its gone undiscovered since the days of Windows 8.

*I'm not expecting answers BTW it's just i love a good detective story that involves what went wrong or how something works in computers.

[CAT-THE-FIFTH]

6C/12T Ryzen 1 and 2 compared with 6C/12T CFL at the same clockspeed:

https://www.techspot.com/article/161...-core-8th-gen/

Looking at the application benchmarks there is not much in it,so gaming must be being affected by lack of optimisations plus things like latency.

[DanceswithUnix]

What i was trying to say is that Windows and AFAIK all software has typically never relied on just a single timer, normally to get an accurate measurement of time you use two or preferably more timing sources and constantly run checks against the two with highest resolutions, back in the day IIRC that was the RTC and the PCI clock generator.

On the PC you traditionally only had the RTC chip to work with, and that is really low precision. For short timings, these days you have the CPU timing registers.

TBH I would have thought the best way to get timings would be to run a second PC and spit a packet across the network to declare the start and end of test. There would be a network delay but that would be just a few milliseconds and the same for start and end so should cancel out. That would allow something like a Raspberry Pi running a network time client (so high accuracy) to record timings without having to worry about core migrations, load level, missed interrupts and core parking that make this hard on modern PCs.

I did have a quick play at benchmarking when I was unemployed so had some spare time a year ago, it didn't drive me nuts but was more work than I fancied for a hobby. The start of the site is still up if you want a laugh, www.linuxbench.com But hey, I was playing with Linux so I could script my benchmarks

[Corky34]

On the PC you traditionally only had the RTC chip to work with, and that is really low precision. For short timings, these days you have the CPU timing registers.

I could have sworn i remember something about not being allowed to use IRQ 0 for configuring the old sound blaster 16 in DOS because it was used for timing system events, maybe i need to take one of those tests for senility.

EDIT: A look at Ryzen 2's voltage curves.

[watercooled]

CD at Semiaccurate spoke about Intel's 10nm still having problems a while back. Just seen this announced: https://www.anandtech.com/show/12693...m-cpus-to-2019

[CAT-THE-FIFTH]

CD at Semiaccurate spoke about Intel's 10nm still having problems a while back. Just seen this announced: https://www.anandtech.com/show/12693...m-cpus-to-2019

I wonder if AMD can get their performance CPUs to 7NM before Intel can get to 10NN?? Let's hope GF actually manages to deliver on their promises!

[CAT-THE-FIFTH]

That is some good increases over Ryzen MK1 with games like PUBG,Fortnite and WoW:

https://www.computerbase.de/2018-04/...raft-1920-1080

[Corky34]

Interesting deep dive into PB2 on [H]ardOCP...

Wraith Prism aside, Precision Boost 2 is the real deal. I have spent the last week trying to shoot holes in it, and those holes are just not there. AMD has simply over-delivered when it comes to Precision Boost 2. I also do not recall ever giving a CPU "feature" an award in the last 20 years, but that is exactly what we are going to do today. Precision Boost 2? Color me impressed. AMD just won the Great Turbo War of 2018.

[CAT-THE-FIFTH]

The B450 chipset will be out in July and the new Z490 in June:

https://www.tomshardware.com/news/bl...pus,37001.html
https://www.anandtech.com/show/12711..._medium=social

Edit!!

AT test the Acer Swift 3 with a Ryzen 7 2700U:

https://www.anandtech.com/show/12709...n-meets-laptop

They think part of the reason the Intel laptops have better battery life is down to the use of DDR4 in the AMD laptops and LPDDR3 in the Intel ones.

[Biscuit]

Be interesting to find out what Z490 actually is in comparison to X470