• Welcome to TechPowerUp Forums, Guest! Please check out our forum guidelines for info related to our community.

Ryzen 3000 memory controller has "half the performance" on single CCD CPUs

This is the exact discussion I've been searching desperately for for the past few days... I'm a VFX artist by day, indie filmmaker by night who needs to build a new PC for a personal project-- I've already bought all parts except CPU and need to complete the build ASAP, but I'm really worried about this whole "half the memory for writes" thing on the new Ryzen's, because I suspect CG animation filmmaking workflows will be a 1% sort of thing vs. who these processors seem to be built for (gamers, etc.).

I animate in 3D (Maya) 2D (After Effects), and edit in Premiere (though I may switch to Resolve soon). A huge part of my workflow is being able to play back previews of animation in real time (the less dropped frames, the better). I'm not terribly tech savvy and only build a workstation every 5-7 years, so it's hard for me to discern what involves writing to RAM vs. reading from it, but it seems to me that something we call "RAM previews" in AE would be writing to RAM, and I do that a lot, and really need it to be fast. Render/export speed is less important to me, as I do that far less often and can let my computer render while I'm doing other things. You guys seem to really understand a lot about this issue (I've had trouble finding people that do), so I'd appreciate it very much if you could tell me whether you think my workflow would improve or suffer if I went with, say, a 3600x vs. a 2600x, etc.? Thanks!
You should fill out your system specs in your profile, its helpful knowing where your coming from.
The thread title is completely wrong, the memory controller and memory are unaffected theoretically since these days other components can read and write to memory as well as the CPU and it is technically the link between the CPU cores and the io die/ memory controller not quite the same since a lot of other io uses that route.
If i were you though I would hold on a bit, someone will investigate this ,likely intel and any real issues will be found. I
Don't think it an issue but getting either the 3900x which isn't affected or threadripper or an Intel quad memory platform would provide an increase in memory bandwidth.
All this is my thoughts though and I'm holding onto my cash for now :).
 
Last edited:
This is the exact discussion I've been searching desperately for for the past few days... I'm a VFX artist by day, indie filmmaker by night who needs to build a new PC for a personal project-- I've already bought all parts except CPU and need to complete the build ASAP, but I'm really worried about this whole "half the memory for writes" thing on the new Ryzen's, because I suspect CG animation filmmaking workflows will be a 1% sort of thing vs. who these processors seem to be built for (gamers, etc.).

I animate in 3D (Maya) 2D (After Effects), and edit in Premiere (though I may switch to Resolve soon). A huge part of my workflow is being able to play back previews of animation in real time (the less dropped frames, the better). I'm not terribly tech savvy and only build a workstation every 5-7 years, so it's hard for me to discern what involves writing to RAM vs. reading from it, but it seems to me that something we call "RAM previews" in AE would be writing to RAM, and I do that a lot, and really need it to be fast. Render/export speed is less important to me, as I do that far less often and can let my computer render while I'm doing other things. You guys seem to really understand a lot about this issue (I've had trouble finding people that do), so I'd appreciate it very much if you could tell me whether you think my workflow would improve or suffer if I went with, say, a 3600x vs. a 2600x, etc.? Thanks!
The one benchmark that was affected directly was SuperPI 32M. I would imagine that anyone who uses RAM drives can benchmark this easily also. For a single user, 22GB/s still seems to be plenty of bandwidth for any work you might be doing. If you use a single-CCD ryzen 3000 to run (say) a VM for more than 2 people, then they will be doing RAM writes at DDR2 speeds all of a sudden which may be noticeable.
 
GUYS,
Don't bother looking at data that doesn't matter. Focus on real-world benchmarks. It reminds me of the graphics card troll saying "but this card has more VRAM bandwidth" when they completely missed the point that there was lossless data compression and the card was scoring higher in benchmarks anyway compared to the card they were trying to say was better...

If an R9-3900X scores higher in a benchmark for a program YOU USE then go ahead and get it. Don't even worry about the architectural reasons beyond knowing what DDR4 capacity, CL value etc to get
 
Thanks for your help, everyone who replied! This was kind of what I was wondering, actually-- whether "memory write speed" actually means anything in the real world, or whether there are a ton of other factors at work that might make that one number irrelevant. So far only Puget Systems has done an After Effects or Premiere benchmark that I can find (and Blender, which is pretty similar to Maya), but I'm a tad concerned that most benchmarks for content creators seem to center around render and export times, when, in the real world, these numbers matter much less for most of us than how fast we can complete a RAM preview, or how many frames get dropped during playback in a Maya or Premiere timeline (which I think is a caching thing... I admittedly don't know much about how that works on the hardware end).

Here's a link to the Puget benchmarks, by the way... wish they'd talk about what goes into their "overall score" as it's harder to relate to real-world performance than, say, FPS number for gamers: https://www.pugetsystems.com/labs/a...MD-Ryzen-3rd-Gen-CPUs-for-Video-Editing-1522/

...from those benchmarks it would appear that the 3600 out-performs even the 2700x pretty massively, so maybe the memory write speed is something that doesn't matter at all in these programs.

Also, special thanks to the person who clarified that this means 22Gb/s can be written to RAM-- since RAW 4k video is, I think, 2Gb/minute of data, and most of what I work on is broken up into shots or sequences that are often nowhere close to a minute per section that needs to be previewed at a time, I can't imagine how, even with a lot of effects applied in AE or Premiere, etc. I could ever approach needing to load more than 22Gb/s into the RAM... so, I'm thinking the only possible real-world issues would be the potential multi-user scenario described.
 
Now now, no need to be so shy, I'm sure there won't be any hard feelings if you share with everyone...
Lol........no, I'm pretty sure there would be.
 
I'm more curious about the max bandwidth of IF at 1800MHz or 1900MHz (for capable silicon) and what this means for all data (especially PCIE) that needs to go through the IF.
DRAM speeds are orders of magnitude faster than PCI-e.

Even full slot 16x PCI-e 4.0 maximum theoretical speed is 32GB/s ( 16 GT(transfers)/s per bit , x 16 = 256 GT / 8 = 32), and no GPU today or in the near future is capable of sucking up CPU data so fast.
A 4x NVMe drive is only 8 GB/s theoretical max, and even that "reduced" memory bandwidth on R7 3700x (measured at over 28 GB/s) is several times faster.

Also be aware that this is pure CORE to MEMORY transfer speed, that doesn't count DMA access

DMA = Direct Memory Access, a function of the "North Bridge" (Now called I/O Die) allowing PCI devices to read and write in memory WITHOUT the CPU.
Basically, the CPU only tells the DMA Controllers " Copy NNN pages from here to there", and then the DMA does it by itself, creating an interrupt message to the CPU when done.
During that time, the CPU is free to do other things...

Most GPU texture loads, or NVMe transfers take place via DMA, so the write bandwith of IF is meaningless in those situations. No bytes are actually moved into and out of the core/cache/CCD/CCX

This has been part of CPU architectures since EVER, even the original IBM PC had a DMA controller chip (called Intel 8237)

Of course, the DMA controller became faster and faster, and today it is integrated in the PCI-e fabric, so any PCI-e device can access memory or another PCI-e device (That's how CrossFire / SLI works actually, espcially CrossFire without a bridge), but it works for any PCI(e) device, including USB controllers, Network, SATA controller, ANYTHING actually.

So yea, stop worrying about Zen 2 memory write speed, it's a total non-issue.
 
On this review, aida memory write on 3700x seems okay, I wonder why.

"CPU being tested
AMD Ryzen 7 3700X (3.6GHz – 4.4GHz Boost) Zen 2

TEST SETUP

Cooling be quiet! Dark Rock 4
Motherboard MSI Prestige X570 Creation
Memory 32GB Corsair Vengeance LPX 3000MHz
Graphics Radeon RX 480 8GB
Storage Corsair LX 512GB SSD
PSU Corsair RM 1000 80 Plus Gold Certified PSU "


index.php



Stepping cpu defect? It cant possibly be, can't it?
 
Last edited:
On this review, aida memory write on 3700x seems okay, I wonder why.

"CPU being tested
AMD Ryzen 7 3700X (3.6GHz – 4.4GHz Boost) Zen 2

TEST SETUP

Cooling be quiet! Dark Rock 4
Motherboard MSI Prestige X570 Creation
Memory 32GB Corsair Vengeance LPX 3000MHz
Graphics Radeon RX 480 8GB
Storage Corsair LX 512GB SSD
PSU Corsair RM 1000 80 Plus Gold Certified PSU "


index.php



Stepping cpu defect? It cant possibly be, can't it?

Looks like decent performance, you have the latest bios?
 
Thanks for your help, everyone who replied! This was kind of what I was wondering, actually-- whether "memory write speed" actually means anything in the real world, or whether there are a ton of other factors at work that might make that one number irrelevant.
The memory copy bandwidth is up over Zen+, and comes close to tying with Intel's copy bandwidth. Might be one those other factors at work.
 
On this review, aida memory write on 3700x seems okay, I wonder why.

"CPU being tested
AMD Ryzen 7 3700X (3.6GHz – 4.4GHz Boost) Zen 2

TEST SETUP

Cooling be quiet! Dark Rock 4
Motherboard MSI Prestige X570 Creation
Memory 32GB Corsair Vengeance LPX 3000MHz
Graphics Radeon RX 480 8GB
Storage Corsair LX 512GB SSD
PSU Corsair RM 1000 80 Plus Gold Certified PSU "


index.php



Stepping cpu defect? It cant possibly be, can't it?
Whoa O! Makes me wonder if some 3700X are actually multi-CCD's with disabled CCX's?
 
Whoa O! Makes me wonder if some 3700X are actually multi-CCD's with disabled CCX's?
That is very possible and I had suspected that ever since the first leaks/rumors from Adored (many months ago) about using multiple dies.

The thing is, unless people start de-liding them en-masse, so those "4+4" 3700X-es are found, there is no way of knowing.
That memory bandwidth benchmark might be an indicator, but it's not proof.

This video is from DECEMBER last year, almost 2 months before even AMD showed that 1-chiplet + 1 IO die at CES

Yes, he got some of SKU "numbers" wrong, and obviously prices.... but that's all marketing.

What did he got right?
Let's see...
126575


In his chart:
Ryzen 3 3300X - 6c ... 3.5 base, 4.3 boost ... became the 3600X : Base Clock 3.8GHz, Max Boost Clock 4.4GHz (better!) - https://www.amd.com/en/products/cpu/amd-ryzen-5-3600x
Ryzen 5 3600 - 8c ... 3.6 base, 4.4 boost ... became the R7 3700X: Base 3.6, Boost 4.4 (identical !!! just the name changed) https://www.amd.com/en/products/cpu/amd-ryzen-7-3700x
Ryzen 7 3700 - 12c ... 3.8 base, 4.6 boost ... became R9 3900X: Base 3.8, Boost 4.6 (identical !!!) https://www.amd.com/en/products/cpu/amd-ryzen-9-3900x
Ryzen 9 3800X - 16c ... 3.9 base, 4.7 boost !? ... "too good to be true" became known as R9 3950X: Base 3.5 (not quite), Boost 4.7 (identical) https://www.amd.com/en/products/cpu/amd-ryzen-9-3950x

Now that the launch has happened, I am extremely surprised how ACCURATE the specs are, especially considering the MEASURED power consumption in the reviews, not the ones declared by AMD.
Variations in names of the products or prices DO NOT MATTER, as those can be changed from one second to another before launch.

Again, this was 7 (SEVEN) months ago !

Also, I seriously believe that AMD has targeted those frequencies for boost (4.8-5.1), but TSMC failed to deliver... and we got what we got.
It's very possible that once the process is improved there will be a refresh that adds 300-400Mhz to the clocks.
 
Last edited:
This post will be updated to reflect what I can find. For now these are the aida benchmarks that I found on reviews.

First will be images of reviews without aida write memory write issue.

index.php




memtest.jpg



AIDA-64-MEM.jpg




tableau-ryzen-3600-crosshair-formula-x570-vonguru-3.jpg



index.php




memory-aida-mem-bandwidth.png




aida2.png





d02.png



Graphs_aida64memwrite.png



9051_190_amd-ryzen-3900x-3700x-zen2-review.png


 
Last edited:
This post will be updated to reflect what I can find. For now these are the aida benchmarks that I found on reviews.


AIDA-64-MEM.jpg




tableau-ryzen-3600-crosshair-formula-x570-vonguru-3.jpg


Wait a sec, some reviews are showing higher copy bandwidth than the i9-9900K?
 
finished the search, 2 reviews dont have the issue, all the others have the issue.


index.php




memtest.jpg


And the interesting thing about the last review is that the 3600 in his review dont have the problem too which indicates this might not be the cpu itself, software maybe?
 
So yea, stop worrying about Zen 2 memory write speed, it's a total non-issue.

Unless you write to RAM a lot, no? I'd imagine @xkm1948 would be interested in this. I don't know enough about RAM to say for sure.
 
finished the search, 2 reviews dont have the issue, all the others have the issue.


index.php




memtest.jpg


And the interesting thing about the last review is that the 3600 in his review dont have the problem too which indicates this might not be the cpu itself, software maybe?

Perhaps different versions of AIDA used? If so, which is the correct one?
 
Perhaps different versions of AIDA used? If so, which is the correct one?

Even if it was, how the 3900x is unaffected? ehhe, we could say ryzen 3xxx might have some problem, this is so complicated that it could be. It's funny, all reviews pointed that aida memory write issue on 3600 and 3700x and none has any answer what caused that and those people are professionals.

What I'm checking now is if benchmarks differ that much because of that issue.

Update, benchmarks dont differ that much, still within the margin of error, anyway I give up.
 
Last edited:
If the memory controller on Ryzen 3000 needs two CCDs for full utilisation of the bandwidth, it will be interesting to see how a possible GPU as the second CCD will perform. If it at least is guaranteed bandwidth for writes.
 
Or is it possible the reviewer was lazy and just reduced core and thread count on a 3900 to simulate a 3700 instead of testing a proper 3700?
 
Or is it possible the reviewer was lazy and just reduced core and thread count on a 3900 to simulate a 3700 instead of testing a proper 3700?

Unethical but yeah, it could have been.
 
in this video here 7:01 minutes

He says amd had to make a compromise here, so the low memory write issue is actually not an issue, so i guess this discussion is over.
 
A relative to the RAM & Zen2 topic video:
 
About that fast ram...

Does anyone here thinks it will matter A LOT MORE for the upcoming 3950X than it does for all the other CPUs with less cores?

3900X already shows slightly more difference for faster ram than 3600(X) (they both use 6-core CCD)...
Adding 4 more cores (8 threads) will certainly increase competition for memory access !

I have a good feeling that the fast 3600 CL16 RAM will shine bright for 3950X, and actually be faster than 3200 CL14 (which is not the case for current existing CPUs)
The increased IF bandwidth will also matter more when 8 cores need to exchange data with another 8 cores, than 6+6.
It's also possible that 3950X will boost better by having chiplets of higher quality, so more Mhz will need a tiny bit more RAM performance to work at max capacity.

I guess we'll see in 2 months time or so ;)
 
In situations that can use that core/thread count, sure... othewise, in the vast majority of situations, no.
 
Back
Top