Question Ryzen 3700X low RAM write speed conundrum

[birdie]

People seem to have overlooked the fact that Ryzen 3700X "features" very low RAM write speeds:

1) https://www.overclock3d.net/reviews...n_7_3700x_ryzen_9_3900x_x470_vs_x570_review/9

2) https://www.tweaktown.com/reviews/9051/amd-ryzen-3900x-3700x-zen2-review/index3.html

which, according to TweakTown, is explained by:

"In memory bandwidth, we see something odd, the write speed of AMD's 3700X, and that's because of the CDD to IOD connection, where the writes are 16B/cycle on the 3700X, but it's double that on the 3900X. AMD said this let them conserve power, which accounts for part of the lower TDP AMD aimed for. AMD says applications rarely do pure writes, but it did hurt the 3700X's performance in one of our benchmarks on the next page. Memory latency is a bit high at stock, but you can overclock the memory quite easily".

Overclockers.com offers a similar explanation:

“This is an expected result. Client workloads do very little pure writing, so the CCD/IOD link is 32B/cycle while reading and 16B/cycle for writing. This allowed us to save power and area inside the package to spend on other, more beneficial areas for tangible performance benefits”.

This sounds extremely odd considering the fact that this CPU consumes up to 120W under load without any overclocking involved. Also memory copy seems not to be affected at all.

I wonder if someone with connections to AMD could dig deeper and find out the root cause of this anomaly because if it's real then the Ryzen 3600X CPU might feature "proper" write speeds despite it being a lower tier CPU.

 

[gdansk]

Also memory copy seems not to be affected at all.

This is the confusing thing to me. How can it copy (read and then write) faster than 16B/cycle? I'd like an explanation of that.

 

[birdie]

A similar discussion at TPU.

 

[Topweasel]

Well that settled any doubts I had in getting the 3900x. Multi chiplet for me. I doubt this will make a major difference. Doubt I will ever run into a scenario where I would be able to tell, but the way this chip runs, you can tell it waaaaay over engineered to keep premium performance on 2 chiplet designs. For the most part it shouldn't matter much which config you have but like this if there is a compromise to be made the dual chiplet CPU gets the advantage.

 

[lopri]

It sounds just like SLI'ed GPUs getting 2x theoretical bandwidth of a single GPU. Meaning the bandwidth is not shared between the chiplets and the theoretical advantage only shows up in microbenches.

Though I wonder why it does not affect read/copy bandwidths?

 

[BTRY B 529th FA BN]

What does the 3800X write at? Could that be a reason for the 105w tdp?

 

[.vodka]

What does the 3800X write at? Could that be a reason for the 105w tdp?

It's still a single CCD product, so it'll have the same write performance as the rest of the single CCD CPUs (3600/x, 3700x)

The 105w TDP bracket is just higher headroom for the XFR algorithm to do its boosting work.

AMD didn't sample the 3800x for review, so we don't know how much higher the average frequency is for the 3800x over the 65w 3700x.

 

[VirtualLarry]

I wonder, does this affect the R5 3600? Is read and write still faster than a 1st-Gen R5 1600, say for Distributed Computing applications, that sometimes are heavy on the memory-controller? Or does this mean that in certain corner-cases, that a 3600 might be slower than a 1600 at DC? (Although, given the clock-speed differences and IPC improvements, I have my severe doubts that such a problem would occur.)

Edit: As per a joke from @DrMrLordX , if Zen2 have 256bit FMACs, for AVX2, does this write-speed disparity affect that at all? Will AVX2 or AVX512 -using DC apps be severely affected? Or is the solution, for those, to buy the 3900X, and not worry? Might be the most direct solution, if this issue doesn't manifest with the 3900X.

 

[.vodka]

Seeing the performance I'd say AMD did their homework when doing things this way. There's no discernible performance loss because of the lower memory write speed on 1 CCD parts. It's no issue.

They wouldn't have done it this way if it would be noticeable on common workloads, although I'm sure there will probably be a corner case here or there that tanks relative to a 3900x or 3950x.

By the way the 3600 just obsoleted the entire 1xxx and 2xxx AM4 lineup at almost every workload (going even further with an all core OC to ~4.3-4.4GHz, the only part where all core OC makes sense), my good old 1700 will be finding a new owner soon.

 

[VirtualLarry]

By the way the 3600 just obsoleted the entire 1xxx and 2xxx AM4 lineup at almost every workload (going even further with an all core OC to ~4.3-4.4GHz, the only part where all core OC makes sense), my good old 1700 will be finding a new owner soon.

Yes, I agree, with everything except the all-core OC. Good luck with OCing a 3600 past 4.2Ghz. I think that you will need AT LEAST 240mm AIO WC to break past 4.2Ghz. With mine, PBO enabled, 95W TDP stock cooler from a R5 1600 (stock cooler for the 3600 is 65W), I still hit 92C, and don't break 3.9Ghz all-core.

 

[lopri]

I have been testing this out and it seems to me that the write speed is entirely bound by Infinity Fabric bandwidth, which kind of makes sense since that is what connects the CPU and memory.

 

[mtcn77]

It might be evident since it has been posted by OP; however "4600 DDR" is a whole new tier entirely.