Ryzen: Strictly technical

Page 59 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
Status
Not open for further replies.

imported_jjj

Senior member
Feb 14, 2009
660
430
136
HPET is enabled, absolutely. I will have to retest, though, as I did not realize it at the time - so thanks for catching that :p

I think the memory could have been running at 3200, despite what the screenshot says. Others on the forums have stated seeing similar behavior with the 0083 BIOS. I ran a couple other simple benchmarks and they were entirely consistent (2300 in CPU-z ST, for example), but I pushed overclocking too far with Ryzen Master and had to clear the CMOS.

The fact that memory bandwidth scales with CPU clocks in the 2933MHz results is way weird. At stock you get very low BW and for the 2 overclocks, the BW is too high.
3200MHz wouldn't explain this so it has to be something else.

EDIT: Any changes in 4k writes perf with the new AGESA? Would be nice to see some fixes in the SSD area too.
 
Last edited:
May 11, 2008
19,303
1,130
126
There is something i am wondering about, when using several logic units on different clock domains, it is necessary to synchronize when data goes from one logic unit to another logic unit while both having different clock domains. I wonder if there are memory speeds where ryzen just no longer benefits from the increased memory speed.

For background information about synchronizing between different clock domains.
http://www.eetimes.com/document.asp?doc_id=1279906
 

looncraz

Senior member
Sep 12, 2011
722
1,651
136
The fact that memory bandwidth scales with CPU clocks in the 2933MHz results is way weird. At stock you get very low BW and for the 2 overclocks, the BW is too high. 3200MHz wouldn't explain this so it has to be something else.

EDIT: Any changes in 4k writes perf with the new AGESA? Would be nice to see some fixes in the SSD area too.

Actually, I've seen this with every scaling test I've done - Ryzen memory performance goes up with CPU clocks. This isn't really unexpected when you consider all of the logic and cache running at core clocks. The DF interface, for example, seems to run at the L3 speed, whereas the DF itself runs at memory clocks.

Running at DDR4-3200 would explain it pretty well, I think. But a funky PLL might do so as well.

More testing to do. I've mostly focused on NVMe benchmarks. 4k-64Thrd results seem low, but everything else is higher than on Intel. Real world testing (copy tests, for example) show higher performance than Intel and latency is lower.

Most of my SSDs are well traveled, so they don't all have their full new write speeds in any event... which can be a major factor to consider. My 850 Evo, however, still does - and performs rather similar on Ryzen as it did with my Sandy Bridge, with an edge for Ryzen overall - especially in 4K-64Thrd write results, which are doubled.
 

mattiasnyc

Senior member
Mar 30, 2017
356
337
136
I've got a question regarding the PCIe implementation...

I've been looking at the x370/b350 motherboards and I understand that the 'higher end' boards have two x16 slots that run x8/x8 when both are populated. Both of those pipe straight into the CPU from what I can gather.

What I don't understand though is why all of the boards with legacy PCI (as well as some others) have a single x16 to the CPU, and then for the second x16 it'll max out at x4 but drop to x1 if the PCIe x1 slot(s) are occupied. It just seems like the two should be ultimately unrelated to each other since the legacy PCI slots by definition hang off of the chipset, not the CPU. Is this just a coincidence or?



PS: I have a legacy PCI card that I want to keep one way or the other (because if I don't I'm likely looking at a $800-$1,500 additional cost for pro-audio), and all the PCI cards have this odd 'coincidence'. Unfortunately I'd also like the option of running two GPUs where one runs video editing processing and the other powers monitors, i.e. x8/x8 off of the CPU.... Not finding mobos for that combination yet... And this is what I have read and thought I understood:

ryzen-processors-support-ecc-memory.png
 
Last edited:

looncraz

Senior member
Sep 12, 2011
722
1,651
136
I've got a question regarding the PCIe implementation...

I've been looking at the x370/b350 motherboards and I understand that the 'higher end' boards have two x16 slots that run x8/x8 when both are populated. Both of those pipe straight into the CPU from what I can gather.

What I don't understand though is why all of the boards with legacy PCI (as well as some others) have a single x16 to the CPU, and then for the second x16 it'll max out at x4 but drop to x1 if the PCIe x1 slot(s) are occupied. It just seems like the two should be ultimately unrelated to each other since the legacy PCI slots by definition hang off of the chipset, not the CPU. Is this just a coincidence or?



PS: I have a legacy PCI card that I want to keep one way or the other (because if I don't I'm likely looking at a $800-$1,500 additional cost for pro-audio), and all the PCI cards have this odd 'coincidence'. Unfortunately I'd also like the option of running two GPUs where one runs video editing processing and the other powers monitors, i.e. x8/x8 off of the CPU.... Not finding mobos for that combination yet... And this is what I have read and thought I understood:

ryzen-processors-support-ecc-memory.png

All PCI-e 2.0 slots come from the PCH. 3.0 slots come from the CPU.

PCI will be coming from the PCH.

You should note that you can probably manage to run graphics off PCI-e 2.0 x1, but you will need to take care about what card you choose. If you can wait a while, though, a board that can suit your needs more directly will eventually be released.
 

mattiasnyc

Senior member
Mar 30, 2017
356
337
136
Yeah, that's what I figured more or less. I just find it curious that there aren't any boards with legacy PCI and PCI-e 3.0 x8/x8.... Oh well, I may have to go with an adapter board instead.
 

looncraz

Senior member
Sep 12, 2011
722
1,651
136
Yeah, that's what I figured more or less. I just find it curious that there aren't any boards with legacy PCI and PCI-e 3.0 x8/x8.... Oh well, I may have to go with an adapter board instead.

That would certainly be a better long-term proposition. I'd suggest buying one and testing out how it works before fully investing into a legacy free platform.
 
  • Like
Reactions: Drazick

imported_jjj

Senior member
Feb 14, 2009
660
430
136
Actually, I've seen this with every scaling test I've done - Ryzen memory performance goes up with CPU clocks. This isn't really unexpected when you consider all of the logic and cache running at core clocks. The DF interface, for example, seems to run at the L3 speed, whereas the DF itself runs at memory clocks.

Running at DDR4-3200 would explain it pretty well, I think. But a funky PLL might do so as well.

More testing to do. I've mostly focused on NVMe benchmarks. 4k-64Thrd results seem low, but everything else is higher than on Intel. Real world testing (copy tests, for example) show higher performance than Intel and latency is lower.

Most of my SSDs are well traveled, so they don't all have their full new write speeds in any event... which can be a major factor to consider. My 850 Evo, however, still does - and performs rather similar on Ryzen as it did with my Sandy Bridge, with an edge for Ryzen overall - especially in 4K-64Thrd write results, which are doubled.

Your results with the DRAM at 2666MHz do not scale and not quite sure it would make any sense for memory BW to scale with core clocks. Some impact on latency maybe but not this huge impact on BW. Not aware of any evidence to back your claim either, quite the opposite.
The scaling is better than linear in your 2933 results. 41.7GB/s at 3.5GHz, 46.5GB/s at 3.8GHz and 48.9 GB/s at 3.9GHz for read.
On SSDs i was just asking if the new AGESA changed anything.

Edit: there are some suspicious results and some variance with the DRAM at 2666MHz too but the majority don't show scaling.
 
Last edited:

CrazyElf

Member
May 28, 2013
88
21
81
I'll try to do it in a couple of hours, I'll be natively running Linux Mint 18.1 while trying to figure out what I screwed up in my CCX benchmark utility - it will be a nice reprieve :p

UPDATE:

Results

Ryzen 7 1700X @ 3.9GHz, 16GB (2x8GB) DDR4-2933 16-16-16-38 1T.
Linux Mint 18.1, kernel 4.4.0-53-generic

Just a head's up everyone, it may be worth upgrading to Kernel 4.10 or newer for testing.

https://www.phoronix.com/scan.php?page=news_item&px=AMD-Ryzen-Newer-Kernel


Anyways, you can use the Ubuntu Kernel Updater (works with Mint too as Mint is based on Ubuntu)
http://www.teejeetech.in/2016/04/ubuntu-kernel-upgrade-utility.html

Alternatively:
http://kernel.ubuntu.com/~kernel-ppa/mainline/




What I'm really wondering is what this is looking like with the memory unlocked.

So far:
  1. The new AGESA largely gets rid of the SMT penalty (in other words, the SMT on Ryzen is even better than Intel's now)
  2. Faster RAM largely gets rid of the gaming penalties that we've seen (a few apps like 7-zip benefit as well)
  3. A few games here and there have even received patches (Ashes of Singularity, Total War: Warhammer, and Bathesda has indicated their desire to work with AMD)
So does that mean we are basically where we want to be?

  1. Haswell (or better than Haswell in a few cases) IPC
  2. Gaming gap is almost entirely gone now
  3. Power efficient (more so than Broadwell E)
  4. Very impressive SMT scaling

Only disappointing part is that we are stuck at 4 GHz OC it seems. Broadwell E can do 4.2 to 4.4 GHz at 24-7 voltages. Maybe with 14LPU and better IPC, we'd have something that could take on Skylake E in early 2018.
 
  • Like
Reactions: lightmanek

looncraz

Senior member
Sep 12, 2011
722
1,651
136
Your results with the DRAM at 2666MHz do not scale and not quite sure it would make any sense for memory BW to scale with core clocks. Some impact on latency maybe but not on BW. Not aware of any evidence to back your claim either, quite the opposite.
The scaling is better than linear in your 2933 results. 42GB/s at 3.5GHz, 47GB/s at 3.8GHz and 49 GB/s at 3.9GHz.
On SSDs i was just asking if the new AGESA changed anything.

Bandwidth is a function of not just memory speed and the IMC, but of the CPU being able to process the data. I had higher bandwidth with core overclocks on Sandy Bridge, but that was rather minimal by comparison... but Sandy Bridge didn't rely on the same type of data mesh operating in multiple clock domains (IMC -> DF -> CCX fabric -> CORE).

The Ryzen 5 1400 results with DDR4-2400 CL14 on another computer entirely show scaling as well, but not as extreme.

Also, be careful to ensure you are looking at the same core counts in my results.

DDR4-2667 16-16-16-38 CR1

8x 3.0Ghz: 31501
8x 3.5Ghz: 36055
8x 3.9Ghz: 40639

It's important, too, to note that benchmarks that are sensitive to memory bandwidth also improve slightly more than just the CPU clocks would suggest - so it is a real effect.
 
  • Like
Reactions: Drazick

imported_jjj

Senior member
Feb 14, 2009
660
430
136
Bandwidth is a function of not just memory speed and the IMC, but of the CPU being able to process the data. I had higher bandwidth with core overclocks on Sandy Bridge, but that was rather minimal by comparison... but Sandy Bridge didn't rely on the same type of data mesh operating in multiple clock domains (IMC -> DF -> CCX fabric -> CORE).

The Ryzen 5 1400 results with DDR4-2400 CL14 on another computer entirely show scaling as well, but not as extreme.

Also, be careful to ensure you are looking at the same core counts in my results.

DDR4-2667 16-16-16-38 CR1

8x 3.0Ghz: 31501
8x 3.5Ghz: 36055
8x 3.9Ghz: 40639

It's important, too, to note that benchmarks that are sensitive to memory bandwidth also improve slightly more than just the CPU clocks would suggest - so it is a real effect.

Sorry but you are choosing the results that fit your theory, a theory that doesn't make sense and is not supported by any 3rd party data.
You make 2 very risky assumptions
- the DDR was at 3200 not at 2933 as displayed by AIDA
- the memory BW has extreme scaling with core clocks
Why do that before making sure it's not the obvious suspect, software.
I do get the core saturation point and could be interesting to look at that.

Maybe try Geekbench 3 , the memory tests?
 
Last edited:

looncraz

Senior member
Sep 12, 2011
722
1,651
136
Sorry but you are choosing the results that fit your theory, a theory that doesn't make sense and is not supported by any 3rd party data.
You make 2 very risky assumptions
- the DDR was at 3200 not at 2933 as displayed by AIDA
- the memory BW has extreme scaling with core clocks
Why do that before making sure it's not the obvious suspect, software.
I do get the core saturation point and could be interesting to look at that.

Maybe try Geekbench 3 , the memory tests?

I'm not making any assumptions, you are comparing results without taking into consideration the control results - they all show the same pattern of increased bandwidth with core clocks. And it's not just the one test - it's everything, including Geekbench 3 - though I'm not yet back to the DDR4-2933 testing.

Very few people do as much testing as I am doing or as many controls as I do.

Update:

This statement was made before I realized the actual scale of the variation. Agreement between benchmarks ended with AGESA 1.0.0.4, now only AIDA shows the absurd bandwidth scaling with CPU clocks.
 
Last edited:

imported_jjj

Senior member
Feb 14, 2009
660
430
136
I'm not making any assumptions, you are comparing results without taking into consideration the control results - they all show the same pattern of increased bandwidth with core clocks. And it's not just the one test - it's everything, including Geekbench 3 - though I'm not yet back to the DDR4-2933 testing.

Very few people do as much testing as I am doing or as many controls as I do.

The results you have shared don't quite support the claim and no 3rd party data supports the scaling claim. The extreme scaling is rather absurd too.
Based on these results he core is borderline saturated at 3.5GHz-2666 and 3.8-3.9GHz assumed 3200MHz with something else providing more than 10% scaling for memory BW
3.5-2660=40.6
3.5-3200?=41.7
3.8-3200?=46.5
3.9-3200?=49
Less than 10% increase in core clocks between the 2 saturation points and 20% gains in read.
That's too much and on top of that you have 2933MHz reported instead of 3200 and maybe a flaw in the benchmark that saturates the core far too early with w/e workload is used.
 
Last edited:

looncraz

Senior member
Sep 12, 2011
722
1,651
136
The results you have shared don't quite support the claim and no 3rd party data supports the scaling claim. The extreme scaling is rather absurd too.
Based on these results he core is borderline saturated at 3.5GHz-2666 and 3.8-3.9GHz assumed 3200MHz with something else providing more than 10% scaling for memory BW
3.5-2660=40.6
3.5-3200?=41.7
3.8-3200?=46.5
3.9-3200?=49
Less than 10% increase in core clocks between the 2 saturation points and 20% gains in read (real and theoretical).
That's too much and on top of that you have 2933MHz reported instead of 3200 and maybe a flaw in the benchmark that saturates the core far too early with w/e workload is used.

This seems to be a bug with AIDA. I didn't recognize how large the climbing was, despite writing the numbers, LOL! Guess that can happen when running 1,000 different numbers through your head.

Geekbench 3 shows 2% scaling on the memory subtests with AGESA 1.0.0.4 using the 8C/16T configuration and the Ryzen 5 1400. However, the former AGESA showed 14%... between 3Ghz and 3.4Ghz with the 4C/4T configuration, which lines up pretty well with AIDA results.

That is far too much for CPU clocks to explain, as you said. No idea why, but my brain was thinking we were talking about 10% total, not 30%.

GB ST Memory, 4C / 4T, DDR4-2667 CL15, ASRock AB350 BIOS 1.41 Windows 7 (Windows 10)
3.0Ghz: 3500 (3635)
3.4GHz: 3997 (4225)
3.8GHz: 3953 (4315)

GB ST Memory, 8C/16T, DDR4-2667 CL14, C6H BIOS 0083 Windows 10
3.0Ghz: 3608
3.4Ghz: 3654
3.8GHz: 3688

Only 2% total (though I didn't restart with these new tests, contrary to the former tests). HPET always active.

I will play around again with DDR4-2933 tomorrow and test higher speeds on the Ryzen 5 1400 as well. Something weird is going on, that's for sure.
 

imported_jjj

Senior member
Feb 14, 2009
660
430
136
This seems to be a bug with AIDA. I didn't recognize how large the climbing was, despite writing the numbers, LOL! Guess that can happen when running 1,000 different numbers through your head.

Geekbench 3 shows 2% scaling on the memory subtests with AGESA 1.0.0.4 using the 8C/16T configuration and the Ryzen 5 1400. However, the former AGESA showed 14%... between 3Ghz and 3.4Ghz with the 4C/4T configuration, which lines up pretty well with AIDA results.

That is far too much for CPU clocks to explain, as you said. No idea why, but my brain was thinking we were talking about 10% total, not 30%.

GB ST Memory, 4C / 4T, DDR4-2667 CL15, ASRock AB350 BIOS 1.41 Windows 7 (Windows 10)
3.0Ghz: 3500 (3635)
3.4GHz: 3997 (4225)
3.8GHz: 3953 (4315)

GB ST Memory, 8C/16T, DDR4-2667 CL14, C6H BIOS 0083 Windows 10
3.0Ghz: 3608
3.4Ghz: 3654
3.8GHz: 3688

Only 2% total (though I didn't restart with these new tests, contrary to the former tests). HPET always active.

I will play around again with DDR4-2933 tomorrow and test higher speeds on the Ryzen 5 1400 as well. Something weird is going on, that's for sure.

If it were a few % it wouldn't necessarily be unexplainable but it's a lot more so yeah...
Will add a 3rd party result for context, could dig for more but i can't find anything to support BW scaling with core clocks in a large manner or any tests with very low clocks or for core saturation.
At the very least the graph shows no core saturation at stock for the 1700 with the DRAM at 3200.
index.php
 
Last edited:
  • Like
Reactions: looncraz

Insert_Nickname

Diamond Member
May 6, 2012
4,971
1,691
136
Oh well, I may have to go with an adapter board instead.

Something like this -should- work. But check first, there are no guarantees. Even if it works, it might not function 100%, some cards do not like bridges.

https://www.startech.com/Cards-Adapters/Slot-Extension/PCI-Express-to-PCI-Adapter-Card~PEX1PCI1

Does this mean that there is still native PCI available on AMD chipsets? I assumed they had to use PCIe-PCI bridges just like every current Intel board?

They do. I don't think anything AMD has had native PCI since the 900-series.
 

Timur Born

Senior member
Feb 14, 2016
277
139
116
It would be a very strange design if an onboard PCI bridge would slow down CPU bound x16 then. It's rather off-topic, but since the question came up, here is what I discuss in a Gearslutz (forum) thread:

The Startech card uses a Pericom Semiconductor PI7C9X113 PCI Express to PCI-X Bridge. This is a good thing, because I had good experience with the PI7C9X111 in another adapter. Both the Startech card and the other adapter allow me to drive RME PCI cards down to 32 samples buffer.

On the other hand I had a bad experience (as in did not work at all) with a PCI adapter card that looks similar to the Startech, but used an Asmedia based chipset. That being said, I think it was more a thing of these cards being cheap (10+ EUR compared to 40+ EUR for the Startech) than a problem of the chipset. Asus/Asrock boards usually use a current revision of the Asmedia ASM1083 chipset, which in the past I tested to be working properly, but older revisions caused trouble for some people. My Z87 Gigabyte board uses a IT8892E.

One drawback of the Startech card is that you absolutely need to use the power-cable, even though there should be enough power from the PCIe bus. My Pericom based adapter even drives two (2) PCI cards from the PCIe bus alone. It does provide an extra power-plug (SATA based), but I don't need to use it for the RME cards.

My PC case offers a single vertical slot opening. This allows me to combine the Startech card with an extender cable and then have a RME card properly installed in the case. You need to be careful with these extender cables, because the quality is not so great. I had one solder pin bent to its neighboring pin, which I noticed before using the cable and then fixed with a soldering iron.

If your old audio card doesn't work then this can be a sign that the card itself has some defect, too. The DSP MADI shown in the image sometimes causes my PC not to boot when it's used with the Startech + extender cable, a AES works without problems. In this case I happen to know that my MADI card had some defects for some time (aka needs repair), so the adapter problems are just a symptom of those card based defects.

imgext.php


My other mentioned Pericom based adapter is this Sintech one (bought via Ebay from China):

imgext.php


They also sell a 5.1/4" slot/external chassis based solution:

imgext.php
 

Timur Born

Senior member
Feb 14, 2016
277
139
116
According to Elmor the overheating emergency shutdown is initiated by Tctl, not by SIO CPU. Going by my measurements the shutdown temperature would be 115C Tctl then, not 110C SIO as I first thought.

This makes no difference concerning the shutdown problem, though, because the current Sense Skew defaults on an Asus SH6 still prevent the shutdown - and even hard throttling - from happening. A Code 8 happens shortly before emergency shutdown would/should happen and then keeps about 1.0 V Vcore applied to the CPU. For a system that its CPU cooling failing this seems somewhat dangerous to the CPU, doesn't it?!

I don't know how other mainboards implement this, though, so maybe it's an Asus related problem.
 

looncraz

Senior member
Sep 12, 2011
722
1,651
136
http://zen.looncraz.net/ now online.

I don't have time for the next few weeks, maybe as much as a month, to finish more than what has been done.

At that point in time I should also have more Intel results as I will be upgrading an Ivy Bridge Xeon system to Ryzen and will have the parts on consignment for a short while (long enough to run a series of benchmarks at 3Ghz with and without Hyper-threading).
 

tamz_msc

Diamond Member
Jan 5, 2017
3,724
3,554
136
http://zen.looncraz.net/ now online.

I don't have time for the next few weeks, maybe as much as a month, to finish more than what has been done.

At that point in time I should also have more Intel results as I will be upgrading an Ivy Bridge Xeon system to Ryzen and will have the parts on consignment for a short while (long enough to run a series of benchmarks at 3Ghz with and without Hyper-threading).
A big thanks for the incredible amount of effort you put into this work.
 

lobz

Platinum Member
Feb 10, 2017
2,057
2,856
136
http://zen.looncraz.net/ now online.

I don't have time for the next few weeks, maybe as much as a month, to finish more than what has been done.

At that point in time I should also have more Intel results as I will be upgrading an Ivy Bridge Xeon system to Ryzen and will have the parts on consignment for a short while (long enough to run a series of benchmarks at 3Ghz with and without Hyper-threading).
man.... you are a big balancing force to juanrga in the universe
 
Status
Not open for further replies.