AMD: Zen 3 Announced [2020-10-08]

I think this mention of memory bandwidth may explain why I was happy with my 6c/12t 3930k @ 4.5GHz for so long. I ran eight sticks of single-ranked 1600MHz CL7 1T 4GB DIMMs in the full quad channel configuration, which (even by today's standards) provides excellent memory bandwidth and latency numbers.

Intel Core i9-11900K Processor 16M Cache up to 5.30 GHz Product Specifications <- 50GB/sec memory bandwidth at "rated" speeds
Intel Core i7-3930K Processor 12M Cache up to 3.80 GHz Product Specifications <- 51.2GB/sec memory bandwdith at "rated" speeds (ten years prior to the above...)

it wasn't a framerate beast, however it always seemed very fluid to me. I now suspect at least part of it was likely due to excessive memory bandwidth...
 
The IMC of the 5800X3D I got is crazy good compared to my old 5800X, it can boot up to 4333 MT/s 1:1, 4267 MT/s 1:1 seems to be the best performance though

21330skp4.jpg
 
@Clukos Mighty impressive clocks on that IMC!! Timings you run on these DDR4 memories are also great, must be preforming really well in games with these :)
 
It wasn't a framerate beast, however it always seemed very fluid to me. I now suspect at least part of it was likely due to excessive memory bandwidth...

Which of course suggests that the bandwidth wasn’t excessive after all.
Bandwidth is underappreciated in benchmark(et)ing in general. The mere fact that in order to get consistent results, you need to run in a clean software environment where you don’t have a number of threads evicting the benchmarked process’ cache contents (which creates bus traffic and require refilling from memory) implies this. And benchmarks are typically selected to be largely bandwidth independent in the first place, because you are interested in (selling) advances in CPU/SoC technology, which being bandwidth constrained nullifies/hides.
 
The IMC of the 5800X3D I got is crazy good compared to my old 5800X, it can boot up to 4333 MT/s 1:1, 4267 MT/s 1:1 seems to be the best performance though

Any chance you could run same benchmark for the 2007 release of Crysis?

I'm curious to see if the 5800X3 can finally hit 60fps lock in all levels.
 
Looks promising!

crysisc6kqv.jpg

That's the worst I've seen it drop (apart from the saving stutters).

The game still looks so good even as it closes in on 15 years since release. And on top of that it still has some of the most impressive environmental destruction (especially trees) and random physics on objects in a game.

15 years on and still most games aren't nearly as good as Crysis at those things. Graphics have improved in the interim, but everything else Crysis still does better.

Once I get time to grab a 5800x3D and install it, I might have to play it again.

Regards,
SB
 
AMD is #1 on the Top 500 https://top500.org/lists/top500/list/2022/06/?page=1
1.1 Exaflops o_O https://top500.org/system/180047/
Its using 8.7million cores of 'AMD Optimized 3rd Generation EPYC 64C 2GHz' which is Zen3 Epycs, also AMD Instinct MI250X (I guess they get included in the core count & probably are the main contributor).

2.5* Rmax & over 3* Rpeak of the previous champion while using 8.8MW less electricity.

Bottom of the list is 1.65PF now. Crazy.
 
AMD is #1 on the Top 500 https://top500.org/lists/top500/list/2022/06/?page=1
1.1 Exaflops o_O https://top500.org/system/180047/
Its using 8.7million cores of 'AMD Optimized 3rd Generation EPYC 64C 2GHz' which is Zen3 Epycs, also AMD Instinct MI250X (I guess they get included in the core count & probably are the main contributor).

2.5* Rmax & over 3* Rpeak of the previous champion while using 8.8MW less electricity.

Bottom of the list is 1.65PF now. Crazy.
Also 3rd, LUMI uses same HPE Cray EX235a nodes with 64C Trentos and Instinct MI250Xs (it will get better results in november update, too, it's not all up yet)
 
I guess the 8.7 M cores figure adds the number of CPU cores and GPU cores which would be meaningless IMHO.
Yes. Using the CU-count of 220 as cores it roughly adds up to 8,730,112.

I couldn't be bothered to try to calculate the exact numbers of CPUs and GPUs exactly.
 
You might be tempted, then, to overclock your eight-core X3D CPU and eke out a little extra performance that way. Unusually, AMD actually locks out overclocking on CPUs with 3D V-Cache, and the reasoning behind this is that the cache die itself is very sensitive to extra voltage—like, "could-instantly-die-when-overvolted"-sensitive.

With that in mind, be aware then that overclocking utilities from the "big four" Taiwanese motherboard manufacturers (that being ASUS, ASRock, Gigabyte, and MSI) purportedly don't honor this behavior and will apparently allow you to crank the clocks and voltage to the sky if you so choose.
 
I recently downgraded from 5950X to 5800X3D for free and get some more on top of. Unlike what I read around it's actually a cool CPU. Haven't seen temps go beyond 75C at 115W load that you won't see other than prime95 or occt stress tests.

Since it's a locked CPU, tuning is very easy and straightforward. Most chips can do -30 CO. Set PPT/TDC/EDC to 114/75/115 and be done with it. You need to have latest agesa bios update installed for your board, otherwise you have to use pbo2 tuner app in Windows. In gaming it's around 50-60C. You can try BCLK overclocking but you can corrupt your OS and at most you'll get 100MHz more, which is negligible. RAM bandwidth is more beneficial than latency thanks to vcache but still you are fine just running default xmp settings.

Quick and dirty comparison below. It's usually 5 to 10% faster but when a game take advantage of Vcache, it can be around 20-30%. 5800X3D also has better IMC thanks to the latest B2 stepping, that I can oc RAM to 3733MHz which scores whooping +1% more than the ones in the table lol.

1680384622734.png
 
Back
Top