Discussion AMD Ryzen MegaThread! FAQ and Resources

[8350rocks]

Benches used stock settings for memory. Overclocking RAM favors both Intel and AMD, as has been demonstrated again and again in this thread. OC RAM is not a magic bullet that will reduce a ~15% microarchitectural gap to 0%. I am not going to copy and paste all the graphs again. I will just re-add two of the more favorable cases: one game and one compute

Also I got info from a RyZen user that just installed the new AGESA that supposedly fixes RAM latency on RyZen. Yes, the new AGESA reduces latency by about 6 nanoseconds, but he lost OC capacity and is now running back RAM at 2400MHz. Any minor improvement on same clocks latency has been lost due to reduction in frequency. In his own words:

If anything, the Agesa 1004a update caused me problems. It killed my RAM overclock. Have to run at 2400 now. So AMD is moving in the wrong direction with optimizations and fixes.

Can you use valid sources? I'm not sure where those foreign benchmarks are coming from. And no idea what their test setups were. RAM directly affects Infinity Fabric, and has a much greater affect on system performance than RAM does on Intel! And each users experience with Ryzen will vary with each system motherboard, updates, etc!

Uh? I am using Hardware.fr because they have provided the most extensive review of Zen. They are a well-known top site and they have tested virtually anything about RyZen: IPC, latency, throughput, BIOS updates, CCX-CCX interconnect, SMT, core parking, scheduler, power management policies,... and it is simpler to me to go their site and take anything that I need, instead goggling a dozen of reviews to check what site has the info I need, like this bench showing how Skylake improves performance on ARMA3 with faster RAM

RAM affect directly Infinity Fabric and this is only relevant if you are using a memory-bound workload. And memory bound workloads also run faster on Intel chips when the RAM is overclocked.

Arma3 is known to respond to faster RAM above and beyond almost any other game. Only Fallout 4 is as sensitive to RAM speed as Arma 3 is.

Look at the Division benchmark, for example...where a 6700K sees next to no gain for faster RAM speeds. The examples you provide are not a complete picture. You are merely citing what is convenient to try to support your own statements.

For example, Lame seems to show a deficit in encoding...however...Eurogamer used handbrake.

http://www.eurogamer.net/articles/digitalfoundry-2017-amd-ryzen-7-1800x-review

Handbrake shows a H.264 difference in favor of Ryzen by ~10%, while H.265 shows a difference in favor of the 6900K of about ~5%. I think handbrake is vastly more representative of what a typical person would use to encode video...but to each their own.

There are also many more examples that countermand your statements. For example, computerbase.de shows Ryzen is typically about 8% behind at most in nearly any application workload. In several, Ryzen is even better. I would link the site, but they would likely remove the link because you have to translate the page to english from german. However, I very much trust computerbase.de over any other review site.

 

[goldstone77]

Spoiler

Benches used stock settings for memory. Overclocking RAM favors both Intel and AMD, as has been demonstrated again and again in this thread. OC RAM is not a magic bullet that will reduce a ~15% microarchitectural gap to 0%. I am not going to copy and paste all the graphs again. I will just re-add two of the more favorable cases: one game and one compute

Also I got info from a RyZen user that just installed the new AGESA that supposedly fixes RAM latency on RyZen. Yes, the new AGESA reduces latency by about 6 nanoseconds, but he lost OC capacity and is now running back RAM at 2400MHz. Any minor improvement on same clocks latency has been lost due to reduction in frequency. In his own words:

If anything, the Agesa 1004a update caused me problems. It killed my RAM overclock. Have to run at 2400 now. So AMD is moving in the wrong direction with optimizations and fixes.

Can you use valid sources? I'm not sure where those foreign benchmarks are coming from. And no idea what their test setups were. RAM directly affects Infinity Fabric, and has a much greater affect on system performance than RAM does on Intel! And each users experience with Ryzen will vary with each system motherboard, updates, etc!
"Interestingly, the gains from 2667MHz to 3200MHz were approximately double those from 2133MHz to 2667MHz, despite an identical 533MHz increase. Does this mean increasing returns as the frequency gets higher?" Intel doesn't get a 10% performance increase going from 2667MHz to 3200MHz! https://community.amd.com/servlet/JiveServlet/showImage/2-2790040-118998/Bench.PNG

Just so you know a poor memory controller and cheap manufacturing which is why the "Infinity Fabric" needs high frequency memory to compete more in line with Intel setups with slower speeds is NOT a good thing and its not a excuse, and its even worse that its hard to get Ryzen to even perform at those speeds as of today.

Intel processors also improve with higher frequency memory by a decent margin so any true not bias reviewer should test both processors at the same ram frequency.

Edit by cheap i mean smart but with a compromise

Infinity Fabric operates much like the FSB did, and that is why RAM affect performance the way it does.

Yes and that was all in well Amd's decision on how they made the architecture, not a good excuse.

Not sure what you mean by excuse? That implies something is wrong? The architecture works as they intended it too. They can make custom systems easily just by adding more cores, because of Infinity Fabric, and that is a win for them. And if in the future they have support for higher frequency RAM beyond 3200MHz that will be a plus.

 

[8350rocks]

Spoiler

Benches used stock settings for memory. Overclocking RAM favors both Intel and AMD, as has been demonstrated again and again in this thread. OC RAM is not a magic bullet that will reduce a ~15% microarchitectural gap to 0%. I am not going to copy and paste all the graphs again. I will just re-add two of the more favorable cases: one game and one compute

Also I got info from a RyZen user that just installed the new AGESA that supposedly fixes RAM latency on RyZen. Yes, the new AGESA reduces latency by about 6 nanoseconds, but he lost OC capacity and is now running back RAM at 2400MHz. Any minor improvement on same clocks latency has been lost due to reduction in frequency. In his own words:

If anything, the Agesa 1004a update caused me problems. It killed my RAM overclock. Have to run at 2400 now. So AMD is moving in the wrong direction with optimizations and fixes.

Can you use valid sources? I'm not sure where those foreign benchmarks are coming from. And no idea what their test setups were. RAM directly affects Infinity Fabric, and has a much greater affect on system performance than RAM does on Intel! And each users experience with Ryzen will vary with each system motherboard, updates, etc!
"Interestingly, the gains from 2667MHz to 3200MHz were approximately double those from 2133MHz to 2667MHz, despite an identical 533MHz increase. Does this mean increasing returns as the frequency gets higher?" Intel doesn't get a 10% performance increase going from 2667MHz to 3200MHz! https://community.amd.com/servlet/JiveServlet/showImage/2-2790040-118998/Bench.PNG

Just so you know a poor memory controller and cheap manufacturing which is why the "Infinity Fabric" needs high frequency memory to compete more in line with Intel setups with slower speeds is NOT a good thing and its not a excuse, and its even worse that its hard to get Ryzen to even perform at those speeds as of today.

Intel processors also improve with higher frequency memory by a decent margin so any true not bias reviewer should test both processors at the same ram frequency.

Edit by cheap i mean smart but with a compromise

Infinity Fabric operates much like the FSB did, and that is why RAM affect performance the way it does.

Yes and that was all in well Amd's decision on how they made the architecture, not a good excuse.

Not sure what you mean by excuse? That implies something is wrong? The architecture works as they intended it too. They can make custom systems easily just by adding more cores, because of Infinity Fabric, and that is a win for them. And if in the future they have support for higher frequency RAM beyond 3200MHz that will be a plus.

support up to 3600 is already in the pipeline...beyond that, I cannot say...

 

[jdwii]

8350rocks digital foundry will quite honestly disagree with you they even showed I3 with high speed ram vs I5 sandy-bridge results. Higher frequency memory can make quite the difference in gaming with Skylake and above probably even my processor but i can't test it.

http://www.eurogamer.net/articles/digitalfoundry-2017-intel-core-i5-7600k-review
http://www.eurogamer.net/articles/digitalfoundry-2015-intel-core-i3-6100-review


To add more here is ryzen in their results doing the same thing
http://www.eurogamer.net/articles/digitalfoundry-2017-amd-ryzen-7-1700-1700x-vs-1800x-review

 

[goldstone77]

Spoiler

Benches used stock settings for memory. Overclocking RAM favors both Intel and AMD, as has been demonstrated again and again in this thread. OC RAM is not a magic bullet that will reduce a ~15% microarchitectural gap to 0%. I am not going to copy and paste all the graphs again. I will just re-add two of the more favorable cases: one game and one compute

Also I got info from a RyZen user that just installed the new AGESA that supposedly fixes RAM latency on RyZen. Yes, the new AGESA reduces latency by about 6 nanoseconds, but he lost OC capacity and is now running back RAM at 2400MHz. Any minor improvement on same clocks latency has been lost due to reduction in frequency. In his own words:

If anything, the Agesa 1004a update caused me problems. It killed my RAM overclock. Have to run at 2400 now. So AMD is moving in the wrong direction with optimizations and fixes.

Can you use valid sources? I'm not sure where those foreign benchmarks are coming from. And no idea what their test setups were. RAM directly affects Infinity Fabric, and has a much greater affect on system performance than RAM does on Intel! And each users experience with Ryzen will vary with each system motherboard, updates, etc!
"Interestingly, the gains from 2667MHz to 3200MHz were approximately double those from 2133MHz to 2667MHz, despite an identical 533MHz increase. Does this mean increasing returns as the frequency gets higher?" Intel doesn't get a 10% performance increase going from 2667MHz to 3200MHz! https://community.amd.com/servlet/JiveServlet/showImage/2-2790040-118998/Bench.PNG

Just so you know a poor memory controller and cheap manufacturing which is why the "Infinity Fabric" needs high frequency memory to compete more in line with Intel setups with slower speeds is NOT a good thing and its not a excuse, and its even worse that its hard to get Ryzen to even perform at those speeds as of today.

Intel processors also improve with higher frequency memory by a decent margin so any true not bias reviewer should test both processors at the same ram frequency.

Edit by cheap i mean smart but with a compromise

Infinity Fabric operates much like the FSB did, and that is why RAM affect performance the way it does.

Yes and that was all in well Amd's decision on how they made the architecture, not a good excuse.

Not sure what you mean by excuse? That implies something is wrong? The architecture works as they intended it too. They can make custom systems easily just by adding more cores, because of Infinity Fabric, and that is a win for them. And if in the future they have support for higher frequency RAM beyond 3200MHz that will be a plus.

By claiming all you need to do is buy faster memory to get more in line with Intel processors that offer the same performance with lower frequency ram is in excuse, not to mention a added cost and at this time even more time to get it stable.
Amd did a module like design to save money on manufacturing but it causes issues such as needing faster memory in applications that can make use of more then 4 cores. If i had a guess that is why we are seeing I5 level of gaming performance out of the processors when we should be seeing 6900K level of performance in gaming.

In Watch dogs 2 a stock 6900K beats a 5Ghz 7700K but 1800X competes with an I5. But in all productivity workloads that can use 8 cores the 1800X actually beat the 6900K.

I even asked this at the Amd tomshardware discussion and i got a reply basically saying optimization and so forth but no real technical reason to why this is happening.

Several games already use 8 cores and the 6900K beats a 7700K in some of those games but the ryzen chip competes with an I5.

By which Juan claims that Ryzen isn't a low latency processor i believe he is saying when a program uses more then 4 cores latency increases and performance doesn't scale inline with an 8 core Intel processor in latency based applications which of course is gaming.

First I didn't say anything about Intel, or trying to bring AMD in line with Intel. This is a brand new architecture, and to actually take full advantage of it there are some requirements. In that past the FSB was very similar in which RAM had a great deal of affect on system performance. This is not an Intel clone trying to imitate how intel gets work done. Not all games are optimized for Ryzen, but all games have been optimized for Intel for years. Gaming engines with just a small optimization like AOS have showed significant improvement in FPS with Ryzen architecture vs. Intel's huge advantage in IPC. This optimizations shows that lower clock Ryzen with a lower performing IPC doesn't take away from it's performance. And it's at the point of diminishing returns when it comes to gaming for the average gamer. Watch this video https://www.youtube.com/watch?v=RZS2XHcQdqA

 

[jdwii]

"Not all games are optimized for Ryzen, but all games have been optimized for Intel for years."

I personally think Very few games are optimized for anything, nothing but Amd fanboy lies or Intel lies to claim so. This is why i laugh when i read this. All optimizations are basically telling the game how to treat SMT on Ryzen.

 

[juanrga]

It is interesting that you mention now computerbase, which not only agrees with what I say but even shows a bigger gap between Broadwell and Zen. Computerbase gives a 12% average advantage to the 6900k over the 1800x. Correcting for clocks we obtain Broadwell is ~25% faster than Zen clock-for-clock.

Computerbase shows 8% in both workload overall graphs.

14% on games and 9% on compute.

The total average is 11.5%, which I rounded to 12%, and correcting for clocks Broadwell is ~25% faster

1.115* 3.6 / 3.2 = 1.254

 

[juanrga]

Not all games are optimized for Ryzen, but all games have been optimized for Intel for years. Gaming engines with just a small optimization like AOS have showed significant improvement in FPS with Ryzen architecture vs. Intel's huge advantage in IPC.

All recent games have been designed with consoles in mind and consoles use AMD hardware. In fact I remember older discussions with certain people stating that since AMD was in both consoles, games would start to be optimized for AMD. I recall reading several articles about that. I recall an article on Eurogammer where they predicted that soon an FX-8350 would be better for games than i5 as a consequence of AMD being in both consoles.

AOS is a very special case. The game was broken on RyZen and performed at only 65% of the performance of Broadwell (both at 3.8GHz)

The patch has solved the problem and now RyZen is much much closer, but still behind Broadwell.

There is no magic 'fix' for RyZen microarchitecture, RyZen is what it is. At best AMD can fix the pair of broken games or pay some developers to optimize a pair of future titles for RyZen and minimize the performance gap. That is all.

The same happened with Bulldozer, after original launch AMD promised windows scheduler fixes and game optimizations. Several years after a pair of games saw huge double digit percent gains, but overall the improvement was a mere 2--4%.

 

[randomizer]

It works for everyone...not just a specific subset of people. So, in a manner, I suppose I am referring to over 7 billion ways. Myself, yourself, and this entire forum included.

I thought that might be the case. You are correct, of course. My statement was deliberately self-referential.

 

[ex_bubblehead]

I will only post this warning once as after this point there will be ejections.

Some of you will notice that posts have disappeared. There will be more as this thread is cleaned up.

This is a final warning that the infighting and personal callouts will cease.... Immediately!

 

[jaymc]

Hey people...

I'm interested to hear your thoughts on this video...
This guy run's Quake 2 in software on both Ryzen an Kaby Lake an then compares fps on each system..

He actually wrote the OpenGL Software Renderer he use's in the video, an also admits he optimized it (at the time that is) for AMD CPU's..
Although for some reason I do think he sounds quite genuine.. as usual I am probly wrong about this... do let me know what yis think.

The results are quite surprising !
Writing his own renderer eh..

Anyhow check this guy out:
https://www.youtube.com/watch?v=lN5mxFfkr7g

Jay

Edit:
Here's another "Analyzing the Core IPC of Ryzen 1800X in Blender"
https://www.youtube.com/watch?v=myPRNtifMaA

 

[Nope 1151]

Juan.　Do you think that AMD will use 14HPP in light of issues clocking Ryzen?

 

[8350rocks]

Juan.　Do you think that AMD will use 14HPP in light of issues clocking Ryzen?

I doubt they change process on Ryzen. Zen+ may be a possibility depending upon their goals for the architecture and performance...though I would not rate the odds better than low at this point.

 

[8350rocks]

XFR is 200MHz for the 1500X and that is for standard cooling. It could get higher clocks on an open case with low tempts, because the XFR turbo depends on temperature.

RyZen is not closing the gap significantly. Here you have the impact of the new AGESA update that addressed the memory latency issue, how you can check yourself the overall performance has not varied a bit

Here you have a recent graph with last data for RyZen including the new models. First games and below compute

If you compare the R5 1400 with the i7 2600k. You can easily deduce

RyZen IPC+SMT ~ Haswell
RyZen IPC ~ Sandy

which is the same conclusion obtained at launch day with the reviews of the 1800X.

Yeah... Well, No.

A single data point doesn't help compare "day1" to "current day". I still see in today's data that Ryzen is within 7% of Kaby Lake clock per clock when day1 was 12%-ish. Whatever the reasons, I don't really care, but like you say "the data is there". The most beaten to death horse, CB, seems to agree with me there.

Which, at the same time, makes me even more curious to explain *why* we see bigger differences elsewhere. Your default answer is "because latency". Well, again, no. That is not a convincing answer. Have more data that could indicate why? Something that you haven't thought of yet?

I have read that different RAM and cooling should be the biggest differentiation variables, but outside of Toms and Anands, I haven't read anything else around those topics.

At launch RyZen was 10--20% behind Broadwell clock-for-clock. Improved BIOS, Windows updates, and the extra cache in R5 models can have closed the gap by about 5%. But if you check the above recent data 4C/8T RyZen performs more like a 4C/8T Sandy Bridge than like a 4C/8T Kabylake.

It is interesting that you mention now computerbase, which not only agrees with what I say but even shows a bigger gap between Broadwell and Zen. Computerbase gives a 12% average advantage to the 6900k over the 1800x. Correcting for clocks we obtain Broadwell is ~25% faster than Zen clock-for-clock.

Computerbase shows 8% in both workload overall graphs.

14% on games and 9% on compute.

The total average is 11.5%, which I rounded to 12%, and correcting for clocks Broadwell is ~25% faster

1.115* 3.6 / 3.2 = 1.254

AMD has stated that they are only ~10% behind Kaby Lake clock for clock in an interview...I would say that contradicts your claims.

 

[juanrga]

Juan.　Do you think that AMD will use 14HPP in light of issues clocking Ryzen?

I guess you mean Glofo 14HP. My answer is "no". This is the node used for IBM power9 and I suspect IBM has some exclusivity clause, otherwise 14HP would appear on Glofo website as part of the product portfolio.

Even if 14HP was open to anyone, I don't see AMD using it. This is an expensive process node made on a SOI substratum, which is optimized for 4GHz and higher clocks. AMD couldn't use it for mobile chips, neither for servers, GPUs, or APUs, which account for most of Zen volume production. This would force AMD to port the design to different process nodes and AMD doesn't have the money to do that.

 

[juanrga]

AMD has stated that they are only ~10% behind Kaby Lake clock for clock in an interview...I would say that contradicts your claims.

My claims are based in reviews. What AMD says in an interview is rather irrelevant.

Reviews have measured RyZen is about 10% behind Broadwell on compute and about 20% behind on games both clock-for-clock. That is the reason why the R5 1400/1500X perform like a Sandy/Ivy R7 with similar clocks.

10% behind Broadwell implies about 13% behind Kabylake/Skylake for compute and last reviews also confirmed this:

AMD still lags in IPC to Intel, so a 4.0 GHz AMD chip can somewhat compete in single threaded tests when the Intel CPU is around 3.5-3.6 GHz, and the single thread web tests/Cinebench results show that.

4/3.5 = 1.142 ==> 14%

4/3.6 = 1.111 ==> 11%

The average is 12.6% for compute, which agrees very well with my 13% claim for compute.

 

[Nope 1151]

Could we stop quoting longer quote posts? It's probably making others (along with me) suffer on mobile.

Thanks.

 

[Gon Freecss]

@juanrga;

Pretty sure Zen's IPC is comparable to that of Haswell. So theoretically, the Ryzen 7 1800X should perform identically to the i7-5960X in games at the same clocks. However, there seems to be a few issues with Zen and gaming. In professional work, the Ryen chip performs around the same as the i7-5960X at the same clocks, and even surpasses it in others.

 

[jdwii]

Could we stop quoting longer quote posts? It's probably making others (along with me) suffer on mobile.

Thanks.

That's why i use the spoiler

 

[jdwii]

@juanrga;

Pretty sure Zen's IPC is comparable to that of Haswell. So theoretically, the Ryzen 7 1800X should perform identically to the i7-5960X in games at the same clocks. However, there seems to be a few issues with Zen and gaming. In professional work, the Ryen chip performs around the same as the i7-5960X at the same clocks, and even surpasses it in others.

I wish i had money so i can test Ryzen for myself but looking at benchmarks my favorite CPU is the 1500X comparing that to a 4 core 8 threaded Intel processor

Given that the 1500X has a 3700mhz turbo vs 4500mhz the 7700K has i try and see the difference after i deduct 21.6%.

Ryzen main enemy really is the infinity fabric and lower clock speeds to kaby lake.

Be nice to see these two things fixed with Zen+ or Zen2 or whatever they call it.

I've also noticed and i'm sure others have as well that ryzen 1500X gets higher SMT returns then what i seen with Intel's HT.

 

[jaymc]

Ryzen 5 1600x overclocked to 5.9ghz on LN2:

http://www.digitaltrends.com/computing/amd-ryzen-5-1600x-overclock-59/

 

[randomizer]

Could we stop quoting longer quote posts? It's probably making others (along with me) suffer on mobile.

Thanks.

There's nothing wrong with quoting longer posts as long as you only quote the relevant bits. Quoting a quote of a quote of a quote only to respond with two sentences is probably not necessary.

 

[jdwii]

So, the more I keep looking at several benchmarks from several sites and hours of research it seems like the 1500X at times beats the 7700K in IPC in some cases by 3-6%. Most cases its 5-15% weaker in IPC and at times even cases such as 40%(one benchmark and another I found was 28%) but I guess this might be the optimization thing Amd was talking about, and it’s not a common result.

I’m confident that if Ryzen wasn’t being bottlenecked by its infinity fabric and if clock speeds could reach the same as kabylake Ryzen would perform much better than it does. In several gaming benchmarks.

1500X should be treated as equivalent to a stock 4770K when using 2933-3200mhz memory

Though I wish I had a chip myself so I could do testing but I don’t. I will be buying a Ryzen build one day however.

 

[TMTOWTSAC]

Ryzen 5 1600x overclocked to 5.9ghz on LN2:

http://www.digitaltrends.com/computing/amd-ryzen-5-1600x-overclock-59/

I saw this earlier on PCGamer, and the part that stuck out to me was the 2 GB of RAM. CPU-Z was showing it as 1398 MB, 12-12-12-24-57-1. Really makes you wonder how much performance improvement will be available from fine tuning the memory controller.

 

[juanrga]

@juanrga;

Pretty sure Zen's IPC is comparable to that of Haswell. So theoretically, the Ryzen 7 1800X should perform identically to the i7-5960X in games at the same clocks. However, there seems to be a few issues with Zen and gaming. In professional work, the Ryen chip performs around the same as the i7-5960X at the same clocks, and even surpasses it in others.

I have provided lots of benches and averages showing this is not true: I gave Zen vs Haswell at 3.5GHz and Zen vs Broadwell at 3GHz. And Zen IPC is clearly behind: about 10--20% behind Broadwell. This is the reason why 4C/8T RyZen performs like a 4C/8T Sandy/Ivy with similar clocks. I compared R5 1500X with i7 3770k and R5 1400 with i7 2600k. The microarchitecture of Zen is closer to Sandy than Haswell: 6-wide, 2AGU, 16FLOP/core vs 8-wide, 3AGU, 32FLOP/core.

As I said before launch, Zen is optimized for throughput, but fails at latency. We already knew before launch that Zen was going to shine on GPU-like workloads but suffer on games and any other latency sensitive workload. This is a discussion I had with CanardPC before launch

https://mobile.twitter.com/CPCHardware/status/836346777267761155/actions
https://mobile.twitter.com/juanrga/status/836503836260990976/actions
https://mobile.twitter.com/CPCHardware/status/836508218092302336/actions

AMD has designed Zen for certain server workloads (something similar did happen with Bulldozer) not for games. That is why I said since the launch day that gaming is not going to be fixed soon. There is no magic fix for a muarch designed for throughput. I have been saying this for months that those magic fixes (mobo-fix, SMT-fix, scheduler fix, 4+0 fix,...) that certain people has been promising in forums were fake. And reviewers and even AMD, in an official communication, have confirmed all what I said up to now.

The last hope was a new AGESA that was going to 'fix' the latency problem. Yes, latency has been reduced by ~6ns, but the performance has increased exactly by zero, because, I will say it again, there is no "magic fix".

If we are lucky Zen2 will be better at latency sensitive code, but Zen2 is not coming before 2019 or 2020.