Discussion AMD Ryzen MegaThread! FAQ and Resources

[jdwii]

Juan do you have any idea why Ryzen is performing this way in games that can use 8 cores?

 

[salgado18]

Spoiler

Phoronix has a review of Ryzen core scaling: http://www.phoronix.com/scan.php?page=article&item=amd-ryzen-cores&num=1

The most interesting part I think is this: scaling is off, it's a bad benchmark, but every situation in which there is only one CCX active, performance is a lot better than with two CCX. Makes me sad for the 4-cores with 2+2

But also very hopeful for the APUs

That bench is weird. 3+0 faster than 4+0? 1+1 faster than 2+2? I wonder 0+0 must be a real winner here

Don't have too much hope on the APUs. Yes, they have a full CCX, but they use a different CCX with half the L3 cache than its big brothers. Half the cache will hurt performance in games.

That bench is stupid hahahaha

But it is one of four game benches they did. The other Dota bench was useless, and the Middle Earth one showed no scaling at all (possibly GPU bound). So, from that article, that's the best they have.

In any case, it is interesting to note that single-CCX scores are higher than any other double-CCX result. I really want to see those results on a full game benchmark on Windows, like from Tom's or Anandtech. Surprised there aren't many of those around.

Don't you think the interconnection latency is a much bigger issue than the size of the L3 cache? A while back, we were discussing that the difference in performance from L3 CPU and no-L3 APU Pilediver was negligible, when trying to estimate Zen's improvements.

 

[juanrga]

Juan do you have any idea why Ryzen is performing this way in games that can use 8 cores?

RyZen is optimized for throughput. That is the reason why RyZen shines on GPU-like workloads, which consist of independent, or almost independent, threads: Blender, Handbrake, CineBench,... Ryzen doesn't shine on latency sensitive workloads with inter-dependent threads: games.

 

[gamerk316]

Juan do you have any idea why Ryzen is performing this way in games that can use 8 cores?

RyZen is optimized for throughput. That is the reason why RyZen shines on GPU-like workloads, which consist of independent, or almost independent, threads: Blender, Handbrake, CineBench,... Ryzen doesn't shine on latency sensitive workloads with inter-dependent threads: games.

It also doesn't help the Linux scheduler is designed for minimum latency over maximum throughput, which makes things look worse in that bench.

 

[thegentlewoman]

There is no problem with averages. If a "patch" improves performance by 10% on a given game, but the improvement is 0% in other nine games then the average is 1%. The same happened with the Bulldozer patches. One game got 10% improvement a pair of games got 4% improvement and almost everything else performance remained unchanged. That is why AMD mentioned that the Bulldozer patches increased performance by 1--2%:

AMD was pretty honest when it described the performance gains FX owners can expect to see from this update. In its own blog post on the topic AMD tells users to expect a 1 - 2% gain on average across most applications.

However, now AMD itself claims that the W10 scheduler is working fine for RyZen and that there is not need for any update:

We have investigated reports alleging incorrect thread scheduling on the AMD Ryzen™ processor. Based on our findings, AMD believes that the Windows® 10 thread scheduler is operating properly for “Zen,” and we do not presently believe there is an issue with the scheduler adversely utilizing the logical and physical configurations of the architecture

No patch is coming because there is not anything to fix. What is more, as part of the investigation, AMD has determined that an outdated version of Sysinternals Coreinfo utility was responsible for originating the incorrect topology data that has been widely reported in the media.

If there's nothing wrong with how Win10 handles Ryzen why did Microsoft say they were releasing an update for it in their batch Windows Updates due for release in April ?

The update is addressing other issues, power management issues:

The average thing is crystal clear -.

What if I run win 8? \ 8.1 ? what is going on? Do I have to go on Win 10 absolutely? Might I encounter issues like crashes with internet, media play, gaming, rendering, things like that?

 

[Rogue Leader]

Apparently a reminder is in order here. There are rules, and if you don't follow them at a minimum your post will be deleted, from there on you may be sanctioned or banned permanently.

1. Stay ON TOPIC. This is about AMD Ryzen, not Intel's new CPUs, nor AMD's new Graphic Cards.

2. No GRAPES. G.R.A.P.E.S. (thats an acronym) means guns, religion, abortion, politics, economics, sexuality, and anything related to them.

3. Disagree peacefully, attack the message not the messenger.

If you have noticed your posts being edited or deleted, you are the problem here, just a hint.

 

[8350rocks]

It is just there in the article...

I still can't find it, lol. I can't be that blind, can I? xD

There should be better drivers and bios by the time ryzen 5 launches.

Same arguments we had after the BD launch. There's only so much software can do to optimize a hardware architecture.

Seriously people, Zen has been in incubation, what, three years? I'd imagine AMD had the BIOS/Drivers in a pretty good state given how long they were developing the hard\ware. It's not like they waited for the hardware design to launch before deciding to work on the software for crying out loud.

Caveat: Zen is not Bulldozer at launch. That is a very important distinction to make.

The performance differences are not stark enough to make or say Zen (as a uArch) is a failure like we all did when BD appeared.

Also, all CPU makers have had issues with their CPUs introducing new uArchs. I still remember Sandy having a lot of issues with freezing and the infamous B-revision for the chipset as well.

I do understand why you're mentioning it, but it's also important to put into context that Zen is not BD all over again.

Based on the quote from an uncredited board manufacturer in this article they may well have done that: http://www.legitreviews.com/one-motherboard-maker-explains-why-amd-am4-boards-are-missing_192470

Makes sense and I hope they start putting MoBos on the shelves soon. People is not that patient. Specially with a Microcenter near *_______*

Confirmed by AMD. The new chips are 3+3 and 2+2

http://www.anandtech.com/show/11202/amd-announces-ryzen-5-april-11th

Anandtech makes a pair of performance estimations. For instance they place the top 6-core Ryzen just behind the i7-7700k on multithreaded benches. This just agrees with what we have been awaiting for the 6C RyZen since before launch

Spoiler

Interesting. Making the 4C versions 2+2 (only) seems dumb. I wonder why they'd do that... A pure 4+0 arrangement would be a better purchase if there's a choice. Could they be saving them for something? I mean, the error should be comparable for die harvesting, right? As in, the chance of getting a full CCX inteact with just 1 or 2 other cores defective is just as big as getting 1 and 1, right? xD

Oh welp, I wonder if AMD is being straight forward here.

Cheers!

Some theory crafting on 2+2:

1.) Many, many, many more viable defective dies in that config.

2.) With the cores further apart, you get better heat spread, meaning that the thermal limits the 8 cores hit that prevents higher OC, may not be such a tremendous issue with a 4C part.

3.) Only way to do 4 cores with 16 MB cache is 2+2

With 16 MB cache on the 1500X, that is pretty compelling argument...add in that according to the stilt, there is a "gaming fix" coming soon at an architecture level through micrcode, and it may solve all the issues.

What I am thinking here, is that IF is tied to RAM speeds currently, such as it is...they may have left a back door to manually set the IF data rate. *IF* that is the case, the penalty for 2+2 versus 4+0 may become almost nothing. They could just set IF to run at max bandwidth independent of RAM, assuming this is the case.

The only other thing I could think of would be a hardware revision to manually set that mid-cycle, and he just misunderstood when they said update. If that is the case, then later chips would definitely be better than launch chips. Though, I am still unclear precisely what is going on...that at least gives a glimpse of the direction.

 

[juanrga]

What if I run win 8? \ 8.1 ? what is going on? Do I have to go on Win 10 absolutely? Might I encounter issues like crashes with internet, media play, gaming, rendering, things like that?

If you're building or upgrading a system around Kaby Lake or Ryzen, it's pretty much Windows 10 or bust.

http://www.pcgamer.com/microsoft-turns-off-windows-781-updates-for-ryzen-and-kaby-lake/

Apparently Windows 7 and Windows 8.1 will block any OS updates if you are running RyZen (or Kabylake).

 

[-Fran-]

Some theory crafting on 2+2:

1.) Many, many, many more viable defective dies in that config.

2.) With the cores further apart, you get better heat spread, meaning that the thermal limits the 8 cores hit that prevents higher OC, may not be such a tremendous issue with a 4C part.

3.) Only way to do 4 cores with 16 MB cache is 2+2

With 16 MB cache on the 1500X, that is pretty compelling argument...add in that according to the stilt, there is a "gaming fix" coming soon at an architecture level through micrcode, and it may solve all the issues.

What I am thinking here, is that IF is tied to RAM speeds currently, such as it is...they may have left a back door to manually set the IF data rate. *IF* that is the case, the penalty for 2+2 versus 4+0 may become almost nothing. They could just set IF to run at max bandwidth independent of RAM, assuming this is the case.

The only other thing I could think of would be a hardware revision to manually set that mid-cycle, and he just misunderstood when they said update. If that is the case, then later chips would definitely be better than launch chips. Though, I am still unclear precisely what is going on...that at least gives a glimpse of the direction.

Nah, 2+2 will always be worse than 4+0 unless you're hitting more than 8MB in L3. Simple reason is, to transit data from one set of CPU resources to the other you have an implicit penalty. That is what a heavy context switch entails. You have to move not only the cache contents (pointers and registers), but some other additional stuff for the CPU to know how to pick up the slack. If the Fabric is bound to the IMC and works at the same frequency (or something), then you might have same latencies as RAM access, which for L3 is just bonkers; plus you need to re-populate L1 and L2.

Perfectly parallel tasks that can access all of the L3 (or use all of it) will se a benefit as long as they don't jump around as well.

For the average case, no way 2+2 is better than 4+0. And that being said, if the APUs are 4+0, I won't be surprised that they might be even better at the same speeds, which would be kind of ironic. Even without L3 or a halved L3.

Cheers!

 

[blazorthon]

4+0 with 8MB L3 is definitely better. No inter-CCX communication outweighs the minor advantage of the extra L3 several times over. 8MB is plenty of L3 for any modern and reasonably soon future consumer workloads. If the APUs have no L3 at all, then there might be a problem since they don't have enough L2 capacity to pick up the slack. It'd probably be like Phenom II versus LLano/Athlon II with pretty significant performance differences in many workloads.

The reason Piledriver APUs and similar didn't have a big performane drop from the lack of L3 was because they had massive, partially shared L2 caches and the AMD L3 cache was very slow. Ryzen fixed the slowness of their caches and fixed the cache architecture to the much better performing small, fast caches + large, moderately fast cache configuration, but that means most of the performance improvement of Ryzen came from reorganizing the cache. Mess that up by cutting L3 entirely and you're not much better off than before.

Keep in mind that when Llano first came out, it had no L3 cache and only 512KB L2 caches for each core, not shared, like the Athlon II CPUs. Llano often lagged Phenom II (which was only really different on the CPU level by having a 6MB L3 cache) by 20% or so at the same frequency. Piledriver APUs come along and they have two huge 2MB L2 caches, each shared between two of the cores. That made a big difference.

 

[-Fran-]

I had a 965C at 3.9Ghz and still have my A8 3850 ~2.5Ghz... The 965 was indeed faster, but due to clocks only. At the same speed, they were the same CPU and the A8 felt faster with a GPU attached (4890 at the time). You can even go back to the benchies at Anand; the differences are mostly due to clocks and not the lack of L3. I remember having a *very* lenghty discussion with someone at the time with several benchmarks from different places and there were things for both arguments; so for me at the end they were the same.

That is to say, I do think having L3 is better, but not a necessity. If there's not much L3 usage, it's wasted space anyway.

Cheers!

 

[juanrga]

With 16 MB cache on the 1500X, that is pretty compelling argument...add in that according to the stilt, there is a "gaming fix" coming soon at an architecture level through micrcode, and it may solve all the issues.

I guess you got that from that AT "technical thread". It is a anything except technical and it is filled with multiple promises about a soon to be released fix:

• 
  ■ First it was BIOS 'bug' that was going to be solved with a BIOs update. It was demonstrated there is no real bug affecting performance, and thus no fix coming.
  ■ Then it was a 'bug' with SMT was going to be solved with a Windows patch. It was demonstrated SMT affects average performance only by 1% or 3% and thus no fix coming.
  ■ Then they changed again, this time it was the W10 scheduler was not working correctly due to a 'bug' and bouncing threads among CCX and they even said us that Microsoft was already working in a patch. Again it was demonstrated there is no bug, that the W10 scheduler works fine and again there is no fix coming.

Lots of analysis performed by review sites (Hardware.fr, PCPer, TechSpot, computerbase.de) and even AMD in their official blog

https://community.amd.com/community/gaming/blog/2017/03/13/amd-ryzen-community-update?sf62107357=1

has debunked all the fantasies and conspiracy theories in that "technical" thread. Last time I checked that thread, people as OrangeKrush was posting BS about that the current RyZen models in stores are in reality "engineering samples" and that the good models with everything fixed were going to be released soon. LOL!

 

[randomizer]

Some people go to great lengths to resolve their cognitive dissonance.

 

[jdwii]

Some people go to great lengths to resolve their cognitive dissonance.

I'm confused by that statement and if anything i don't see it in this thread. Ryzen is amazing for the money unless you are a gamer then its a second choice

 

[randomizer]

I'm confused by that statement and if anything i don't see it in this thread.

I was commenting on the post above mine, not making a statement about this thread (which I haven't read enough of to judge anyway).

 

[juanrga]

Some people go to great lengths to resolve their cognitive dissonance.

Indeed, after AMD wrote that official answer in their blog, the people in that "technical thread" began a funny conspiracy theory. That people pretended that the W10 scheduler issue was real, and that AMD was denying it because it was being obligated by Wintel (sic). LOL

This is the same people that attacked PCPer and even sent death threads to reviewers because RyZen reviews didn't say what they want to hear.

My advice is to ignore that AT thread or to read it with a huge amount of salt.

 

[cdrkf]

Some people go to great lengths to resolve their cognitive dissonance.

I'm confused by that statement and if anything i don't see it in this thread. Ryzen is amazing for the money unless you are a gamer then its a second choice

Agreed- although I'd just add that *Ryzen 7* (based on it's price point / performance) is a second choice for gaming. I think the Ryzen 5 and Ryzen 3 series are likely going to be the gaming choice on the AMD side.

I know AMD aren't going to be able to compete for the very top spot, still anything below that they'll have a competitive option. As for comments comparing Ryzen to Dozer / Piledriver- yes in games the deficit was only around 20% on those, however the gaming performance is only one aspect that hurt the market share so badly. To my mind the main points were:

- Dozer was 20% behind at stock clocks, but it gets much worse with overclocking. Ryzen oc vs Kaby oc doesn't change the gap that much (they keep relative pace)

- Dozer didn't have the strengths that Ryzen does- at *best* the '8 core' (I still maintian they were really quads) in a well threaded application could only keep pace with the 8 thread i7. Sandy bridge E mopped the floor with dozer and AMD had no answer to the throughput of Intels enthusiast parts. Ryzen on the other hand *behaves like a real 8 core where it counts* and gives a similar Intel config a run for it's money

- Power consumption on the Dozer derrived parts was *much worse* than Intel. That killed performance on any small form factor builds / laptops which really hurt market share imo (though Excavator did at least iron that one out to some extent)

- AMD stopped updating it (at least on the desktop). I do think if we'd had Steamroller and Excavator based 8c desktop parts AMD could have kept much closer to Intel. However relying on Piledriver for so long (and just dropping prices on it) meant that the 8c parts started as an i5 competitor, then became i3 competitor and now you get a better gaming experience out of the $70 hyper-threaded Pentium. AMD have a road map in place with at least 2 new enhanced versions of Ryzen due over the next few years to keep pace with Intel so hopefully they will keep the new parts coming and not rely on 5 year old tech that is totally outclassed.

 

[8350rocks]

Some people go to great lengths to resolve their cognitive dissonance.

Indeed, after AMD wrote that official answer in their blog, the people in that "technical thread" began a funny conspiracy theory. That people pretended that the W10 scheduler issue was real, and that AMD was denying it because it was being obligated by Wintel (sic). LOL

This is the same people that attacked PCPer and even sent death threads to reviewers because RyZen reviews didn't say what they want to hear.

My advice is to ignore that AT thread or to read it with a huge amount of salt.

There are a lot of things I take with a grain of salt...but the Stilt seems to rarely propagate poor information.

I have also heard similar rumblings from other reliable sources saying the same. I just have no specifics behind what the fix actually entails...so speculation is all I can offer there. If I have more concrete information later, I will elaborate.

 

[eathdemon1]

https://www.youtube.com/watch?v=TId-OrXWuOE something is weird. ryzen only lost 5 to7% when they dropped to a 3+3

 

[juanrga]

https://www.youtube.com/watch?v=TId-OrXWuOE something is weird. ryzen only lost 5 to7% when they dropped to a 3+3

There is not anything weird. Reviews (Hardware.fr, computerbase.de) have shown that (4+0) is less than 5% faster than (2+2) on average. Some specific game like BF1 gets a huge double digit performance penalty from CCX-CCX interconnect, but overall the effect on gaming or applications is in the low single digit percent.

 

[jaymc]

This leak about a new 16 core zen states that the "gaming issues" have been "Ironed out" by new silicon revisions:

https://forums.anandtech.com/threads/ryzen-strictly-tec...

Poster goes on to say that these new silicon revisions will be available right down through all SKU's

 

[James Mason]

This leak about a new 16 core zen states that the "gaming issues" have been "Ironed out" by new silicon revisions:

https://forums.anandtech.com/threads/ryzen-strictly-tec...

Poster goes on to say that these new silicon revisions will be available right down through all SKU's

So does that mean only newer versions of the chips work correctly?

 

[jaymc]

Thats if this new silicon even does indeed FIX Ryzen completely. But yes only newer chips will have it... this is the rumour anyway.

 

[salgado18]

This leak about a new 16 core zen states that the "gaming issues" have been "Ironed out" by new silicon revisions:

https://forums.anandtech.com/threads/ryzen-strictly-tec...

Poster goes on to say that these new silicon revisions will be available right down through all SKU's

That link is broken, and a search in the forums didn't come up with it. Could you post again, please?

 

[Nope 1151]

I doubt that is true, that would be Zen+.
Amd cannot just magically improve L3 latency in a couple of days.