Official AMD Ryzen Benchmarks, Reviews, Prices, and Discussion

[krumme]

GMI and xGMI are basically the same link, but with different type of configuration (width), speed and error tolerance.

GMI is the internal form (inter-die) and has bandwidth of 42.6GB/s per direction at 2666MHz MEMCLK.
xGMI is the external form (inter-node, i.e. EPYC only), has the same bandwidth at the same MEMCLK but different speed, width and error tolerance.

Thanx but a few ??:
Isnt threadripper using xgmi then as well as epyc?
How can xgmi have same bandwith as gmi? And what is meant with "speed"? And how is "width" defined physical?

 

[The Stilt]

SP3r2 which is used by TR doesn't support 2P and therefore it has no xGMI. Meanwhile SP3 Epyc parts do support 2P and that means xGMI is present as well.
It appears that both SP3 and SP3r2 use the same socket, which means that a huge amount of the pins will be unused on SP3r2 (four memory channels, 64x PCIe links, xGMI, etc).

xGMI is half the width of GMI, but operates at twice the speed.

 

[formulav8]

xGMI is half the width of GMI, but operates at twice the speed.

To your knowledge or testing, is possibly better latency more important than bandwidth?

 

[krumme]

SP3r2 which is used by TR doesn't support 2P and therefore it has no xGMI. Meanwhile SP3 Epyc parts do support 2P and that means xGMI is present as well.
It appears that both SP3 and SP3r2 use the same socket, which means that a huge amount of the pins will be unused on SP3r2 (four memory channels, 64x PCIe links, xGMI, etc).

xGMI is half the width of GMI, but operates at twice the speed.

Thanx. Making those SP3 boards with twice the speed must be a huge task. Sound to me like it will pose great demands. High end stuff...

 

[T1beriu]

PCPer tests the new patch that brings Ryzen optimization to Rise of the Tomb Raider.

Full analysis.

 

[hotstocks]

@TheStilt, I purchased two more G.skill 3600 C16 8gb b-die single sided modules, and while it trains (but takes a long time), it is not stable for mem tests compared to when I only had 2 modules 100% stable. What do I need to set procD to in ohms? Or do I have to raised soc voltage above 1.1v or ram volt above 1.4v when using 4 instead of 2 modules at 3466mhz C14. I am on 9935. Thanks

 

[imported_jjj]

Raven Ridge shown at Computex.
If anyone knows the package size,do share so we can measure the die.

 

[T1beriu]

I don't think the package size has been released. If people can find a good photo (LE: from other publications) maybe we can extrapolate the die size using the size of resistors on the package as reference.

 

[imported_jjj]

The screen grab was from the live stream, better ones can be obtained from the recording

EDIT

 

[csbin]

http://www.tomshardware.com/news/amd-threadripper-vega-pcie-lanes,34581.html


AMD demonstrated ThreadRipper's performance in a few scenarios, including the now-famous Blender rendering test of the Ryzen CPU and logo. The render required a mere 13.04 seconds, which is much faster than the 36 seconds AMD demonstrated with the Ryzen 7 1800X prior to its launch.

 

[csbin]

 

[tamz_msc]

http://www.tomshardware.com/news/amd-threadripper-vega-pcie-lanes,34581.html


AMD demonstrated ThreadRipper's performance in a few scenarios, including the now-famous Blender rendering test of the Ryzen CPU and logo. The render required a mere 13.04 seconds, which is much faster than the 36 seconds AMD demonstrated with the Ryzen 7 1800X prior to its launch.

There were two versions of the demo - one that was released initially was at a higher sample count and that required 36 seconds on the 3.4GHz Ryzen 8 core without turbo. The one demoed at a closed door event was a lower sample count and that demo was released a day or two later for clarification. In the 'correct' version it was around 26s for the same configuration.

Anyway, we are getting very near 2x scaling, which means that clock speeds are going to be in the ~3.4GHz range.

 

[itsmydamnation]

GMI and xGMI are basically the same link, but with different type of configuration (width), speed and error tolerance.

GMI is the internal form (inter-die) and has bandwidth of 42.6GB/s per direction at 2666MHz MEMCLK.
xGMI is the external form (inter-node, i.e. EPYC only), has the same bandwidth at the same MEMCLK but different speed, width and error tolerance.

Do you know the topology of the 4 SOC solution? is it mesh or ring? I assume its ring based of the original Fud leaks around the APU.

 

[krumme]

Agree. Looks like 180w tdp and 3.5 to me. Anyway -Very- nice scaling and perf if true. That fabric...looks like a game changer.

 

[deasd]

That's quite quite big. Now what about the heatsink

 

[Veradun]

That's quite quite big. Now what about the heatsink

Apparently they just chose to use the same package as EPYC, meaning a lot bigger than what would be needed. Probably to reduce costs for a very low volume market segment.

 

[krumme]

Apparently they just chose to use the same package as EPYC, meaning a lot bigger than what would be needed. Probably to reduce costs for a very low volume market segment.

Yep. Btw is there a server segment of 12/16c worth adressing at the same time?

 

[lolfail9001]

There were two versions of the demo - one that was released initially was at a higher sample count and that required 36 seconds on the 3.4GHz Ryzen 8 core without turbo. The one demoed at a closed door event was a lower sample count and that demo was released a day or two later for clarification. In the 'correct' version it was around 26s for the same configuration.

Anyway, we are getting very near 2x scaling, which means that clock speeds are going to be in the ~3.4GHz range.

You do understand that they could run third version of demo this time instead? Blender does not have linear scaling with core counts, so yeah.

 

[DrMrLordX]

No one would have bought that A12 notebook

If it had 2x4 GB of RAM and a 128GB SSD, lots of people would have bought it, and liked it. That's what it was designed for after all. The R7 1700 is a notebook is basically silly. Awesome in a way, but still silly.

 

[tamz_msc]

You do understand that they could run third version of demo this time instead? Blender does not have linear scaling with core counts, so yeah.

Doesn't make sense to create even more confusion. Even the Tom's article from which I got the 13.04s result reports that the 1800X does it in 36s.

There was already a controversy with incorrect settings back in the day.

 

[Veradun]

Yep. Btw is there a server segment of 12/16c worth adressing at the same time?

Probably that is the workstation domain, that will be addressed by Threadripper-like EPYCs for major OEMs to ship "certified" workstations.

 

[coercitiv]

The R7 1700 is a notebook is basically silly. Awesome in a way, but still silly.

It's not silly per se, but if we're talking about a gaming laptop then I agree. However, some professionals may be willing to pay nice dollars for a 8c/16t power house inside a well built laptop.

 

[Malogeek]

Doesn't make sense to create even more confusion. Even the Tom's article from which I got the 13.04s result reports that the 1800X does it in 36s.

There was already a controversy with incorrect settings back in the day.

Sites posting that in articles should use logic to understand the numbers don't match, find out why and post the correct information.

 

[DrMrLordX]

It's not silly per se, but if we're talking about a gaming laptop then I agree. However, some professionals may be willing to pay nice dollars for a 8c/16t power house inside a well built laptop.

I just wish we had more notebook options with 65W TDP chips/APUs. Opens up a lot of options really. Most OEMs don't want to crack 15W anymore, even for an APU with no dGPU.

 

[krumme]

You do understand that they could run third version of demo this time instead? Blender does not have linear scaling with core counts, so yeah.

Yeaa they could and the results would be meaningless then. Whats the purpose then? I dont find it likely.