Gigabyte publishes R9 3950X overclocking guide and benchmarks

[QXC]

Source: https://www.gigabyte.com/FileUpload/Global/multimedia/2/file/548/988.pdf

 

Via: https://www.tomshardware.com/news/amd-ryzen-9-3950x-overclocking-watercooling-4.3-ghz,40555.html (Site uses very heavy javascript, be warned)
 

TL;DR

• Gigabyte achieved a 4.3GHz all core OC 1 hour P95 stable @ 1.4v (4.4ghz R15 stable)
  • This suggests that those juicy 3800X dies are in fact being set aside for the 3950X... or Gigabyte was using a well binned chip.
• They cooled this using an EK "EK-KIT" P360 and maxed out around 100c.
  • The EK-KIT is basically a custom CPU only 360mm loop.
  • The Tom's article somehow interprets this as "the 16-core Ryzen 9 3950X is as easy to cool as an 8-core part." somehow. It's still a very hot chip.
• The 3950X scored 3932 stock CB R15, 4384 points at 4.3GHz w/ XMP, and 4475 at 4.4GHz
  • Fast scores, but a short benchmark that favors AMD. Still, these scores are in the ballpark of the 2990WX and 7980XE (and refreshes).

 

Personally I'm curious as to if this is NDA'd, Tom's historically isn't the kind of place to break NDA. Motherboard manufactures like to leak things but not usually like this. I'm also a bit skeptical about this information being posted without third party reviews. Although it was much different, I'm reminded of the AMD PBO video that alluded to 4.75GHz core speeds. This is obviously much more direct, but I wonder if it was intentional. To grab attention away from Intel's two stage paper launch of the HDET price cut chips perhaps?

November here we come.

[Doobeedoo]

Pretty cool though. 

[pas008]

Nice but its 1 benchmark

Hoping that 4.3 or 4.4 is doable on all core also

That might kill the gaming gap

[Deli]

How many people are interested in upgrade to 3950X?

[vanished]

I'm curious why they had such trouble cooling it.  Based on scaling up 3900x power consumption from 12 cores to 16, a D15 should handle it nicely.  We're talking under 250 W for sure even when overclocked, and under 200 at stock.  Clockspeeds appear to be as expected and as promised as well, but that heat output has me a bit concerned and confused.

[justpoet]

Interesting.  I've long been considering the 3950 for a new gaming/workstation build, so I really look forward to reviews on it.  I'll probably be most of a year out from making it though, so by then I'll probably be waiting for the next generation of chips anyway…and graphics cards as well…but this is still highly intriguing.

[S w a t s o n]

2 hours ago, Ryan_Vickers said:

I'm curious why they had such trouble cooling it.  Based on scaling up 3900x power consumption from 12 cores to 16, a D15 should handle it nicely.  We're talking under 250 W for sure even when overclocked, and under 200 at stock.  Clockspeeds appear to be as expected and as promised as well, but that heat output has me a bit concerned and confused.

What was the pump and fan speeds? 100C for p95 runs at 4.3/4.4GHz sounds insane

[justpoet]

1 hour ago, S w a t s o n said:

What was the pump and fan speeds? 100C for p95 runs at 4.3/4.4GHz sounds insane

From product description, there's a fairly wide variance in just the fans.

"Radiator fan: EK-Vardar F4-120ER (500-2200rpm) (3pcs)"

Interestingly, in the pump specs:

"

- Maximum pressure head: 3.9m
- Maximum flow: 1500L/h
- Maximum system temperature: 60°C
"

 

But, looking at the actual pdf from gigabyte, they're basically smashing every voltage up, then also making the load line much more aggressive too, while doing this all core OC.  That later part (load line) has been shown to generally cause a lot of issues for heat and the like (see various GN videos, including with Kingpin and Tin) due to pumping voltage even further.  They were normally running about 97°C after an hour of p95, with the first 8 cores averaging a bit over 8W/core and the second 8 cores averaging about 7W/core.  I'm not sure if that is 1 chiplet getting less, or if it is the SMT core portion using less, but is interesting either way.  8x8=64 + 7x8=56 = 120W rough CPU core use under 16 core P95.  Add a bit for the IO/SOC portion and that's how much heat you need to deal with to sustain 4.3Ghz all core OC with aggressive voltages at about 97°C.  Interestingly, their VRM was up to 83°C during this as well.

 

Various benchmarks may run a bit faster in other configurations as well, since they only used XMP 3200 RAM.

 

Also interesting, their MB was defaulting to what would be 3.3Ghz base clock, rather than the 3.5Ghz base clock on the 3950 spec sheet.  Getting to 4.3 all core isn't bad, considering it is designed to just have a short term single core boost clock to 4.7.

[williamcll]

4475 is impressive, isn't that like, just half the performance of EPYC 7742?

[Ravendarat]

On 10/5/2019 at 8:17 AM, Deli said:

How many people are interested in upgrade to 3950X?

I purposely held off on my 3900x purchase so I could see what the 3950x actually does when it hits the market, but im certainly in the discussion to buy one

[MeatFeastMan]

20 hours ago, williamcll said:

4475 is impressive, isn't that like, just half the performance of EPYC 7742?

It's ahead of the 9980XE apparently. Even when both are at stock.

 

https://hothardware.com/news/gigabyte-cranks-amd-ryzen-9-3950x-to-43ghz-across-all-16-cores

 

So I guess the 10980XE will just edge it out, but you'd darn well hope so considering the price difference.

 

I believe the 7742 did 8450 if I remember correctly from Linus' video.

[cj09beira]

On 10/5/2019 at 5:24 PM, Ryan_Vickers said:

I'm curious why they had such trouble cooling it.  Based on scaling up 3900x power consumption from 12 cores to 16, a D15 should handle it nicely.  We're talking under 250 W for sure even when overclocked, and under 200 at stock.  Clockspeeds appear to be as expected and as promised as well, but that heat output has me a bit concerned and confused.

its all about heat density, 7nm is hard to sool

[vanished]

5 hours ago, MeatFeastMan said:

It's ahead of the 9980XE apparently. Even when both are at stock.

Not surprising, the IPC and clockspeed difference is easily enough to make up a 2 core deficiency under a full core load.

edit: lol I just realized, that's exactly the kind of thing we would have been saying about Intel a few years back vs AMD.  How times have changed...

[WereCat]

On 10/5/2019 at 8:24 PM, S w a t s o n said:

What was the pump and fan speeds? 100C for p95 runs at 4.3/4.4GHz sounds insane

My 3900X runs up to 90°C on stock with NH-U12A cooler at around 1.33V when all cores are under load. 

 

This 3950X was pinned at 1.4V with 4 more cores so I am not really surprised that much by the temperature even if its a 360mm radiator cooling it. 

 

Another thing to consider is that AMD measures its temperatures differently than Intel so we can't directly compare the temperatures. 

 

And one more thing to point out. 

If I stress only one single core on my 3900X it also goes up to 80°C while drawing significantly less power. This to me suggests that either only the hottest spot is reported or simply the area of the heat is so small that it's physically impossible for the cooler to transfer the heat from it as it has a very small area of contact.

 

For comparison. 

My delided i7 4770k at 1.4V 4.6GHz was drawing more power than my 3900X at stock but it never exceeded 76°C with the same cooler. 

[Deli]

5 hours ago, WereCat said:

My 3900X runs up to 90°C on stock with NH-U12A cooler at around 1.33V when all cores are under load. 

 

This 3950X was pinned at 1.4V with 4 more cores so I am not really surprised that much by the temperature even if its a 360mm radiator cooling it. 

 

Another thing to consider is that AMD measures its temperatures differently than Intel so we can't directly compare the temperatures. 

 

And one more thing to point out. 

If I stress only one single core on my 3900X it also goes up to 80°C while drawing significantly less power. This to me suggests that either only the hottest spot is reported or simply the area of the heat is so small that it's physically impossible for the cooler to transfer the heat from it as it has a very small area of contact.

 

For comparison. 

My delided i7 4770k at 1.4V 4.6GHz was drawing more power than my 3900X at stock but it never exceeded 76°C with the same cooler. 

I'm a little confused. Ryzen master and HWinfo report different temperatures. Before the parts of the 280mm watercooling arrive. I'm using a Noctua U9S. The Ryzen master reports 82C after 30min AIDA64 stress test. While HWinfo reports 91C. Running stock, no PBO or auto OC.

[WereCat]

14 minutes ago, Deli said:

I'm a little confused. Ryzen master and HWinfo report different temperatures. Before the parts of the 280mm watercooling arrive. I'm using a Noctua U9S. The Ryzen master reports 82C after 30min AIDA64 stress test. While HWinfo reports 91C. Running stock, no PBO or auto OC.

HWinfo is supposed to be accurate but in this case I would rather believe the Ryzen Master as it can fully monitor the CPU sensors. 

 

That said, both of them may be accurate. 

For example, HWinfo can report the max temp which may be 90°C but that could be only temp spike that lasted few milliseconds. 

In Ryzen master you see the actual temperature so it would be easy to miss the spike. 

[Marucins]

This Gigabyte link has stopped working.
Is there another guide for AORUS boards showing how to overclock the 3950X / 3900X?

[williamcll]

1 hour ago, Marucins said:

This Gigabyte link has stopped working.
Is there another guide for AORUS boards showing how to overclock the 3950X / 3900X?

Try the following:
http://global.inside-test.gigabyte.com/FileUpload/Global/multimedia/2/file/555/995.pdf

https://intensewebs.com/images/dl/Gigabyte_3950X_OC_Guide.pdf