Zen 2 details announced

[maxzor]

Originally Posted by mackerel

There is also some chance my other mobo might work better with fast ram.

I guess if you want to get highest ram speed and overclock IF X570 is needed :<

[hansl]

https://www.extremetech.com/computin...pu-comparisons

Look like X570 boards draw ~40W more power than X470 for same gen2 Ryzen chips.

Particularly interesting is that they found that newer Prime95 version used significantly less power. Any ideas what specific changes to code would have caused that?

Originally Posted by maxzor

I guess if you want to get highest ram speed and overclock IF X570 is needed :<

What are the ram speed limitations of X470?

[maxzor]

Originally Posted by hansl

https://www.extremetech.com/computin...pu-comparisons

Look like X570 boards draw ~40W more power than X470 for same gen2 Ryzen chips.

Particularly interesting is that they found that newer Prime95 version used significantly less power. Any ideas what specific changes to code would have caused that?


What are the ram speed limitations of X470?

About the power issue, other than the bit highly power draw of the chipset itself and PCIE4 needs, no real clue.


I may have spoken ahead of myself.
There is not really a limitation by chipset, but some motherboard RAM traces look to hinder stability at high RAM speeds. X570 in that regard is likely to have this better worked out, because zen2 launched with higher RAM speed support...

[M344587487]

Originally Posted by maxzor

I guess if you want to get highest ram speed and overclock IF X570 is needed :<

Yes. Good quality X470, B450 and B550 boards (when they come out) can possibly do well (probably ~3600 if you choose well, B550 may also be better quality than expected) but you need to be more discerning with these chipsets. It's only with the X570 that many of these manufacturer's are giving the high end a proper go of it, some of the low end is shocking as is some of the midrange even.

Originally Posted by hansl

https://www.extremetech.com/computing/294596-amds-x570-chipset-draws-so-much-power-its-warping-cpu-comparisons

Look like X570 boards draw ~40W more power than X470 for same gen2 Ryzen chips.

Particularly interesting is that they found that newer Prime95 version used significantly less power. Any ideas what specific changes to code would have caused that?

From memory the new chipset draws something like 5-8W more than the old one under load. What that link fails to show is how the benchmarks they were running were finishing quicker on the boards that drew more power, aka the chips were sustaining higher frequencies. See this comparison of the same chip on different chipsets: https://youtu.be/oRaZ2Txv13M?t=405

So not a massive concern but the extra chipset power is something to avoid if possible. Then again if you're pushing an overclock a good efficient VRM (like you'd expect on an X570 board) goes some way to making up for the extra chipset draw.

Originally Posted by hansl

What are the ram speed limitations of X470?

Varies by board, I've seen 3466 specified but am only looking at the affordable boards. They were made for zen+ where 3600 was pretty much the top end that could be achieved at a push, that's not to say higher isn't possible but it'll largely be down to SKU quality.

Something that may be of interest is this AM4 tier list based on VRM quality: https://www.reddit.com/r/Amd/comment...u_decide_on_a/

After a quick gander I'm leaning towards the MSI B450M PRO-VDH Plus as a reasonable choice for an entry level Ryzen 3600 build at up to DDR4 3200, and an MSI B450i Gaming Plus AC as a reasonable choice for a Ryzen 3900X at DDR4 3466 (with luck, DDR4 3600). For a top end possibly overclocked 3900X/3950X at as high a RAM/IF speed as possible it'd have to be an X570 board, probably the Gigabyte X570 Aorus Xtreme.

[maxzor]

Originally Posted by mackerel

4096k FFT iter/s for 1, 2, 6 workers tested with 29.8b5:
8086k 187.59 180.16 170.94
3600 309.71 214.21 193.38

Correct me if I am stupidly wrong, we are talking LL :
this means for 1 worker
8086k 5.3 ms/iter
3600 3.2 ms/iter

What is then wrong with the mersenne.org CPU benchmark page?
It shows 15 ms/iter for top CPUs like 8086k, 9900K.

I am also nurturing from the various up-to-date stickied threads.
Radeon VII is 1 ms/iter at 4M FFT but is it LL?
Any reading on the difference between LL and PRP?
mersenne.org math section seems massively outdated.

[M344587487]

Originally Posted by maxzor

Correct me if I am stupidly wrong, we are talking LL :
this means for 1 worker
8086k 5.3 ms/iter
3600 3.2 ms/iter

What is then wrong with the mersenne.org CPU benchmark page?
It shows 15 ms/iter for top CPUs like 8086k, 9900K.

I am also nurturing from the various up-to-date stickied threads.
Radeon VII is 1 ms/iter at 4M FFT but is it LL?
Any reading on the difference between LL and PRP?
mersenne.org math section seems massively outdated.

Don't know what's wrong with the CPU benchmarks or how they should be interpreted. Regardless it looks like zen2 is the way to go from now on when it comes to CPUs.

gpuowl which Radeon VII uses does PRP. "PRP is the new LL", preferred over LL because of an efficient and reliable error check. Technically not a primality test but a composite test, doesn't matter in practice as it's been agreed that a first time PRP test that results in a positive result counts just like an LL test would for the purpose of recognition etc. Search this forum for PRP and Gerbicz error check with your favourite search engine, probably not the one on the site.



Radeon VII is actually more like 1 ms/iter at 5M.

[maxzor]

Originally Posted by M344587487

Technically not a primality test but a composite test, doesn't matter in practice as it's been agreed that a first time PRP test that results in a positive result counts just like an LL test would for the purpose of recognition etc. Search this forum for PRP and Gerbicz error check with your favourite search engine, probably not the one on the site.

Thank you^


Is the PRP workload also distributed by primenet as this suggests, or do you have to post your exponents in the forum, as you did with your username ("composite as heck" ? :< condolences...)


Will pull the trigger on a used 600€ Radeon VII then, it is a good time for buying it I believe, as a lot of kids feel bad about their card compared to the incoming navi 5800+, and are worried about the "discontinued news"... not keen on being a shark but heh :<[/OT]

[M344587487]

You can get PRP tests manually ( https://www.mersenne.org/manual_assignment/ ) then submit them ( https://www.mersenne.org/manual_result/ ), or let the primenet.py script that comes with gpuowl do it for you. I only posted on the forum as I was technically poaching the exponent from someone who was probably long gone (large exponents do not expire), as a proof of concept for the Radeon VII and for the mild entertainment value of doing a large exponent.

[Prime95]

Originally Posted by maxzor

What is then wrong with the mersenne.org CPU benchmark page?

The benchmarks page is also massively outdated. It shows the timing using a single core of the CPU.

This made sense back in the days when dual-cores were rare and prime95 was not limited by memory bandwidth. The page should be replaced with one that shows throughput numbers. Then the problem becomes the client does not report the memory speed -- so once again the chart is less-than-useful for comparing CPU speeds.

[hansl]

Originally Posted by M344587487

What that link fails to show is how the benchmarks they were running were finishing quicker on the boards that drew more power, aka the chips were sustaining higher frequencies. See this comparison of the same chip on different chipsets: https://youtu.be/oRaZ2Txv13M?t=405

While I agree it would have been much more useful if they could have included performance alongside the watt reading, or even better performance/watt, I still would not automatically assume the x570 board ran quicker.
The video you link was running different ram clocks so not an entirely fair comparison.
The article I had linked looks to have used the same 32GB of DDR4-3200 for all tests.

[M344587487]

Originally Posted by hansl

While I agree it would have been much more useful if they could have included performance alongside the watt reading, or even better performance/watt, I still would not automatically assume the x570 board ran quicker.

You're right you wouldn't assume it, by not showing the data they got a misleading clickbait fluff piece article out of poor reporting.

Originally Posted by hansl

The video you link was running different ram clocks so not an entirely fair comparison.

I missed that, but as it's different RAM clocks per CPU (not per board) the data is still adequate to show that the X570 board they tested was more aggressive in how it clocked:

This means for the benchmarks that you're about to see, the 3900X and 3700X were tested using DDR4 3200 CL14 memory while the [Ryzen] 3600 was tested using DDR4 3000 CL14 memory.

Originally Posted by hansl

The article I had linked looks to have used the same 32GB of DDR4-3200 for all tests.

But entirely ignored performance and in doing so misled you into thinking that the motherboard was using all of that power instead of it being mostly the CPU.