Discussion AMD Ryzen MegaThread! FAQ and Resources

[juanrga]

CCX inter-communication doesn't seem to be the problem.

Indeed, that is not the problem as it was demonstrated by PcPer, and Hardware.fr analysis. Those reviews measured CCX-CCX latencies and the work of the Windows scheduler, or just selectively disabled cores on a 8-core Ryzen and measured performance of 2+2 and 4+0 configurations. One game got a huge ~20% improvement in performance when running on the 4+0 chip, but on average the improvement was less than 5%, confirming that CCC-CCX latency wasn't the problem.

CCX inter-communication doesn't seem to be the problem.

I've been reading and one possible theory is that the PCI-E controller for the dGPU is kind of terrible and needs high performance ram

The problem doesn't affect only games, which implies the source cannot be the PCIe controller. The problem is, as some of us said before launch, on the memory controller

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

The whole of RyZen including its memory controller is optimized for throughput, and this affects any latency-sensitive application, including games.

Overclocking RAM reduces the access latency and the problem is reduced. It is also the reason why AMD has released the new AGESA/BIOS with ~6ns reduction on latency. Unfortunately 6ns aren't enough because the gap is ~5x higher

 

[cdrkf]

CCX inter-communication doesn't seem to be the problem.

Indeed, that is not the problem as it was demonstrated by PcPer, and Hardware.fr analysis. Those reviews measured CCX-CCX latencies and the work of the Windows scheduler, or just selectively disabled cores on a 8-core Ryzen and measured performance of 2+2 and 4+0 configurations. One game got a huge ~20% improvement in performance when running on the 4+0 chip, but on average the improvement was less than 5%, confirming that CCC-CCX latency wasn't the problem.

CCX inter-communication doesn't seem to be the problem.

I've been reading and one possible theory is that the PCI-E controller for the dGPU is kind of terrible and needs high performance ram

The problem doesn't affect only games, which implies the source cannot be the PCIe controller. The problem is, as some of us said before launch, on the memory controller

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

The whole of RyZen including its memory controller is optimized for throughput, and this affects any latency-sensitive application, including games.

Overclocking RAM reduces the access latency and the problem is reduced. It is also the reason why AMD has released the new AGESA/BIOS with ~6ns reduction on latency. Unfortunately 6ns aren't enough because the gap is ~5x higher

According to your table- a 98ns latency is not 5x a 69ns latency...?

 

[randomizer]

According to your table- a 98ns latency is not 5x a 69ns latency...?

The gap is 5x higher. 98-69 = 29ns gap. 29/6 ≈ 5.

 

[clutchc]

On the first page of this thread, there is a disclaimer stating "As a WARNING for new PC builders, DO NOT get high speed RAM in excess of 2933mhz for Ryzen at this time. Or it is pretty much guaranteed that you will suffer instability."

Does this still hold true, or is it safe to buy faster RAM yet?

 

[cdrkf]

According to your table- a 98ns latency is not 5x a 69ns latency...?

The gap is 5x higher. 98-69 = 29ns gap. 29/6 ≈ 5.

What a bizarre way of of measuring it.... This is the problem with metrics like this, depending on what you compare and what you use as your point of reference you can make something quite minor (98 ns vs 69 ns is hardly catastrophic) and make it sound much much worse. The actual difference between the architectures is circa 30%. Removing 6 ns off that gap brings it closer in a meaningful way.

This reminds me of some of the various GPU benchmarks (from either side) where they take a chart and move the '0' line up to exaggerate the differences. The difference between 100 and 105 fps looks insignificant from a baseline of 0, but if you move that up to 90 it gives the appearance of '50% more fps' 😛

 

[8350rocks]

On the first page of this thread, there is a disclaimer stating "As a WARNING for new PC builders, DO NOT get high speed RAM in excess of 2933mhz for Ryzen at this time. Or it is pretty much guaranteed that you will suffer instability."

Does this still hold true, or is it safe to buy faster RAM yet?

No...most MBs will now safely support higher end 3200 kits. G.Skill seems to have the most compatible options. Any Trident Z, Flare X, and even the better cas timing models of Ripjaws seem to work. Also, Team Dark has a cas 14 3200 kit that seems to work as well.

 

[clutchc]

On the first page of this thread, there is a disclaimer stating "As a WARNING for new PC builders, DO NOT get high speed RAM in excess of 2933mhz for Ryzen at this time. Or it is pretty much guaranteed that you will suffer instability."

Does this still hold true, or is it safe to buy faster RAM yet?

No...most MBs will now safely support higher end 3200 kits. G.Skill seems to have the most compatible options. Any Trident Z, Flare X, and even the better cas timing models of Ripjaws seem to work. Also, Team Dark has a cas 14 3200 kit that seems to work as well.

Great to hear, thanks. Then it's no longer necessary to stick to the board's memory QVL? Seems that is what a lot of Ryzen reviewers were recommending awhile back.

 

[jdwii]

On the first page of this thread, there is a disclaimer stating "As a WARNING for new PC builders, DO NOT get high speed RAM in excess of 2933mhz for Ryzen at this time. Or it is pretty much guaranteed that you will suffer instability."

Does this still hold true, or is it safe to buy faster RAM yet?

That needs to be edited out as one can get stable performance Amd even has a new update coming soon for 3200+

 

[clutchc]

On the first page of this thread, there is a disclaimer stating "As a WARNING for new PC builders, DO NOT get high speed RAM in excess of 2933mhz for Ryzen at this time. Or it is pretty much guaranteed that you will suffer instability."

Does this still hold true, or is it safe to buy faster RAM yet?

That needs to be edited out as one can get stable performance Amd even has a new update coming soon for 3200+

Does that include the early issue about populating all 4 DIMM slots with faster RAM? Previously, that was a problem. In fact, the Gigabyte board's QVL I was looking at did not validate any RAM for 4 slots; just 1 and 2 slots.

 

[jdwii]

On the first page of this thread, there is a disclaimer stating "As a WARNING for new PC builders, DO NOT get high speed RAM in excess of 2933mhz for Ryzen at this time. Or it is pretty much guaranteed that you will suffer instability."

Does this still hold true, or is it safe to buy faster RAM yet?

That needs to be edited out as one can get stable performance Amd even has a new update coming soon for 3200+

Does that include the early issue about populating all 4 DIMM slots with faster RAM? Previously, that was a problem. In fact, the Gigabyte board's QVL I was looking at did not validate any RAM for 4 slots; just 1 and 2 slots.

No that is still an issue but i expect future updates to fix that soon we will have a newer update that should make ram speeds higher using 4 dimm slots. Personally i'd buy two 8GB sticks of high speed(3200Mhz) single rank memory over 4 slots plus you can upgrade to more if one needs to.

Ryzen might have done well with a quad channel memory controller.

 

[juanrga]

According to your table- a 98ns latency is not 5x a 69ns latency...?

The gap is 5x higher. 98-69 = 29ns gap. 29/6 ≈ 5.

What a bizarre way of of measuring it.... This is the problem with metrics like this, depending on what you compare and what you use as your point of reference you can make something quite minor (98 ns vs 69 ns is hardly catastrophic) and make it sound much much worse. The actual difference between the architectures is circa 30%. Removing 6 ns off that gap brings it closer in a meaningful way.

It all depends on what is one measuring...

If we are measuring the gap, then it was ~28ns and the optimized BIOS/AGESA reduced this by ~6ns, which means only ~1/5 of the gap was reduced, because the gap was about ~5x higher than 6ns.

If we are measuring performance, things get a bit more complex. 98ns vs 69--70ns is a noticeable 40--42% higher latency, and one of reasons why one microarchitecture is ~10% behind the other microarchitecture, clock-for-clock, on applications.

Latency has been reduced by ~6%

We can estimate that performance on applications would improve by ~0.25 fraction of the ~6% reduction, which means performance would increase by ~1.5%, which is so low that falls within the margin of error. Not a surprise that the new BIOS/AGESA didn't change performance

There is no noticeable performance gains because the gap was about ~5x higher. Indeed 5x * 1.5% = 7.5% ≈ 10%.

 

[8350rocks]

X399 platform chipset diagram from AMD: http://www.tweaktown.com/news/56845/amds-high-end-x390-x399-mobo-dual-ryzen-cpus-possible/index.html

Looks like the 16 core 32 thread monster is real. Quad channel memory as well on that one.

 

[salgado18]

X399 platform chipset diagram from AMD: http://www.tweaktown.com/news/56845/amds-high-end-x390-x399-mobo-dual-ryzen-cpus-possible/index.html

Looks like the 16 core 32 thread monster is real. Quad channel memory as well on that one.

Socket AM44? Also, sounds like a server part to me...

 

[8350rocks]

X399 platform chipset diagram from AMD: http://www.tweaktown.com/news/56845/amds-high-end-x390-x399-mobo-dual-ryzen-cpus-possible/index.html

Looks like the 16 core 32 thread monster is real. Quad channel memory as well on that one.

Socket AM44? Also, sounds like a server part to me...

Looks like MCM socket with 2x 8 core 16 thread chips put together.

 

[jdwii]

Just so others know the Wall is real 1.425V fails at 4.0Ghz during testing but at 3.8Ghz i'm running at 1.3V just fine.

I ran at 1.425V as a suicide voltage as i would not run above that.

 

[Gon Freecss]

Just so others know the Wall is real 1.425V fails at 4.0Ghz during testing but at 3.8Ghz i'm running at 1.3V just fine.

I ran at 1.425V as a suicide voltage as i would not run above that.

It's relatively safe to run the chip at 1.45V. Try it at that voltage.

 

[jdwii]

Just so others know the Wall is real 1.425V fails at 4.0Ghz during testing but at 3.8Ghz i'm running at 1.3V just fine.

I ran at 1.425V as a suicide voltage as i would not run above that.

It's relatively safe to run the chip at 1.45V. Try it at that voltage.

Based on my testing at such settings i'd be at 4.0Ghz stable but it would result in 65-75C temps during benchmarks and pushing max temps for stress tests. Where 3.8 results in 50C temps at 1.3V running benchmarks and 55-60C during stress tests.

3.9Ghz is probably were i will end up. As i'm a fan of cooler temps.

 

[Gon Freecss]

Just so others know the Wall is real 1.425V fails at 4.0Ghz during testing but at 3.8Ghz i'm running at 1.3V just fine.

I ran at 1.425V as a suicide voltage as i would not run above that.

It's relatively safe to run the chip at 1.45V. Try it at that voltage.

Based on my testing at such settings i'd be at 4.0Ghz stable but it would result in 65-75C temps during benchmarks and pushing max temps for stress tests. Where 3.8 results in 50C temps at 1.3V running benchmarks and 55-60C during stress tests.

3.9Ghz is probably were i will end up. As i'm a fan of cooler temps.

What's your chip?

 

[juanrga]

X399 platform chipset diagram from AMD: http://www.tweaktown.com/news/56845/amds-high-end-x390-x399-mobo-dual-ryzen-cpus-possible/index.html

Looks like the 16 core 32 thread monster is real. Quad channel memory as well on that one.

Socket AM44? Also, sounds like a server part to me...

It is the original server [strike]SP3[/strike] SP4 platform has been re-purposed for desktop.

 

[truegenius]

https://www.youtube.com/watch?v=Ew4A5fYyKv4&t=60s

Hey i actually like this video comparing how the Phenom II x6 does today

i was excited to see the phenom x6 in action, then i observed these which didn't let the phenominal cpu give it all it had.
1) hyperx savage kit they use have tRFC time = 260ns, and it is most critical timing i saw in phenom, its as critical as 160ns vs 300ns will make a 1866mhz cl9 ram to perform like 1333mhz ( at best 1600mhz ) cl9 ram in different games ( specially memory hog ), here we are talking about 10% fps difference between 1866mhz @160ns vs 300ns time ( my board have 160 and then 300 after that thus can't test 260ns ).
2) 1600mhz cl11 ! seriously ? probably no body use this loose timing and this much slow speed ram
3) stock cpu-nb speed even at overclock tests of 1090t ! its known that phenom is bottlenecked by imc i.e, cpu-nb speed, so overclocking only cpu speed won't do anything special indeed it will give impression of "dimnishing" returns, the guide they used was not guding them correctly to unlock full potential of phenom.

i was thinking why they included 1055t with 1090t as they are same chip different core speed, then i saw overclocked settings and overclock results of them, 1055t @3.85ghz was consistently ahead of 1090t@4ghz in games by 3-4% despite 1090t's 4% higher core clock speed.
so if we normalize the core clock speed then we get roughly 7-8% difference between them, and since we normalized the core clock then the only thing left causing this difference is cpu-nb and ram speed difference which was 20% and 12.5% respectively, and since cpu-nb means higher ram efficiency ( and l3 speed ) thus we can say that that 12.5% ram ( accompanied by 20% cpu-nb ) gave 7-8% more performance than stock "uncore" part settings. And it also backs my points above.

This also show that ages old arch is benifiting 7-8% from 12.5% memory speed increase then games are being bottlenecked by ram speeds, despite having much less ipc than ryzen, thus a cpu with ryzen leve ipc and so much cores will surely benifit from ram speed ( in other words getting crippled by poor ram latency ).


BTW @jdwii do you have ( or can get hold off of ) ram with different tRFC time to check how much ryzen benifit from this timing ( or can change this timing by around 2x factor ( eg 160ns vs 300 ns ) ? i am wondering is it only phenom that is sensitive to this setting or is it other arch too.
btw best i will recommend is to compare smae speed and cl single rank 2x ns ram vs dual rank 1xns ram ( where "x" is tRFC time ) as these 2 are the differences between gskill sniper and hyperx ram that i have.
also test winrar's inbuilt bench, if it gets better result with ram settings then so games will.
i have gta5's test result with different cpunb and ram speed, but its on phenom so may not be a good idea to share them here unless highly needed, i think 😛 .

kingston's spec sheet is full of information and show this setting in first page in their spec sheet, others don't even have this much informative spec sheet. kudos to kingston for that ( but they have 260ns time thats more than gskill sniper stick's 160ns that i have )

 

[goldstone77]

https://www.youtube.com/watch?v=Ew4A5fYyKv4&t=60s

Hey i actually like this video comparing how the Phenom II x6 does today

i was excited to see the phenom x6 in action, then i observed these which didn't let the phenominal cpu give it all it had.
1) hyperx savage kit they use have tRFC time = 260ns, and it is most critical timing i saw in phenom, its as critical as 160ns vs 300ns will make a 1866mhz cl9 ram to perform like 1333mhz ( at best 1600mhz ) cl9 ram in different games ( specially memory hog ), here we are talking about 10% fps difference between 1866mhz @160ns vs 300ns time ( my board have 160 and then 300 after that thus can't test 260ns ).
2) 1600mhz cl11 ! seriously ? probably no body use this loose timing and this much slow speed ram
3) stock cpu-nb speed even at overclock tests of 1090t ! its known that phenom is bottlenecked by imc i.e, cpu-nb speed, so overclocking only cpu speed won't do anything special indeed it will give impression of "dimnishing" returns, the guide they used was not guding them correctly to unlock full potential of phenom.

i was thinking why they included 1055t with 1090t as they are same chip different core speed, then i saw overclocked settings and overclock results of them, 1055t @3.85ghz was consistently ahead of 1090t@4ghz in games by 3-4% despite 1090t's 4% higher core clock speed.
so if we normalize the core clock speed then we get roughly 7-8% difference between them, and since we normalized the core clock then the only thing left causing this difference is cpu-nb and ram speed difference which was 20% and 12.5% respectively, and since cpu-nb means higher ram efficiency ( and l3 speed ) thus we can say that that 12.5% ram ( accompanied by 20% cpu-nb ) gave 7-8% more performance than stock "uncore" part settings. And it also backs my points above.

This also show that ages old arch is benifiting 7-8% from 12.5% memory speed increase then games are being bottlenecked by ram speeds, despite having much less ipc than ryzen, thus a cpu with ryzen leve ipc and so much cores will surely benifit from ram speed ( in other words getting crippled by poor ram latency ).


BTW @jdwii do you have ( or can get hold off of ) ram with different tRFC time to check how much ryzen benifit from this timing ( or can change this timing by around 2x factor ( eg 160ns vs 300 ns ) ? i am wondering is it only phenom that is sensitive to this setting or is it other arch too.
btw best i will recommend is to compare smae speed and cl single rank 2x ns ram vs dual rank 1xns ram ( where "x" is tRFC time ) as these 2 are the differences between gskill sniper and hyperx ram that i have.
also test winrar's inbuilt bench, if it gets better result with ram settings then so games will.
i have gta5's test result with different cpunb and ram speed, but its on phenom so may not be a good idea to share them here unless highly needed, i think 😛 .

kingston's spec sheet is full of information and show this setting in first page in their spec sheet, others don't even have this much informative spec sheet. kudos to kingston for that ( but they have 260ns time thats more than gskill sniper stick's 160ns that i have )

Here is a link to the AMD forums where I started a thread about infinity fabric. RAM latency was also tested for potential benefits. You might be able to answer some of your questions with the test they ran there. https://community.amd.com/message/2790538#comment-2790538

 

[jdwii]

Just so others know the Wall is real 1.425V fails at 4.0Ghz during testing but at 3.8Ghz i'm running at 1.3V just fine.

I ran at 1.425V as a suicide voltage as i would not run above that.

It's relatively safe to run the chip at 1.45V. Try it at that voltage.

Based on my testing at such settings i'd be at 4.0Ghz stable but it would result in 65-75C temps during benchmarks and pushing max temps for stress tests. Where 3.8 results in 50C temps at 1.3V running benchmarks and 55-60C during stress tests.

3.9Ghz is probably were i will end up. As i'm a fan of cooler temps.

What's your chip?

Ryzen 1700 with B350 TOMAHAWK I'm not 100% sure what voltage to use at max as the videos i saw from the MSI trip said 1.425V max while others say 1.45V and some even say as low as 1.35V. It's probably safe to say its to early and even Amd doesn't know what voltage to use at max. But i plan on keeping this CPU until at least Zen+ or Zen2 but i don't think i'll keep it past 5 years that is pushing it for me.

Plus i like having lower voltage so that means less heat and then less noise.

 

[nikoli707]

Just so others know the Wall is real 1.425V fails at 4.0Ghz during testing but at 3.8Ghz i'm running at 1.3V just fine.

I ran at 1.425V as a suicide voltage as i would not run above that.

It's relatively safe to run the chip at 1.45V. Try it at that voltage.

Based on my testing at such settings i'd be at 4.0Ghz stable but it would result in 65-75C temps during benchmarks and pushing max temps for stress tests. Where 3.8 results in 50C temps at 1.3V running benchmarks and 55-60C during stress tests.

3.9Ghz is probably were i will end up. As i'm a fan of cooler temps.

What's your chip?

Ryzen 1700 with B350 TOMAHAWK I'm not 100% sure what voltage to use at max as the videos i saw from the MSI trip said 1.425V max while others say 1.45V and some even say as low as 1.35V. It's probably safe to say its to early and even Amd doesn't know what voltage to use at max. But i plan on keeping this CPU until at least Zen+ or Zen2 but i don't think i'll keep it past 5 years that is pushing it for me.

Plus i like having lower voltage so that means less heat and then less noise.

cant you just dial in 3.9/4.0 on 2 or 4 cores then drop to 3.8 for all 8?

 

[cdrkf]

@nikoli, you can't currently overclock individual cores on ryzen, it's all or nothing.

 

[jdwii]

@nikoli, you can't currently overclock individual cores on ryzen, it's all or nothing.

Amd should however think about allowing it however another great thing to add to their software. Edit ha ha at 3.9Ghz i'm pretty sure this chip will be stable at 1.325V i'm still at 1.3V and its stable so far. Also from reading online i have a average chip at best and i'm getting this. Some are getting 3.9Ghz with just 1.25V.