AMD RyZen CPU Architecture for 2017

[Cyan]

https://www.msi.com/Motherboard/support/B350M-GAMING-PRO.html#down-bios

This is your motherboard, right? The latest BIOS released should have AGESA 1.0.0.6

Edit: Or not, MSI has not updated the BIOS since early May so that should be in the next update

The bios you mention is the one I have, yes. It is a 16MB file I installed from a pendrive.

I went with Asus Prime X370-Pro and G.Skill RipJaws V F4-3200C14D 2x8GB (Samsung B-die) memories to make sure I can hit 3200MHz memory. Cooling will be done by BeQuiet Silent Loop 240mm AIO and rest of the components will come from my current i7 4790k rig.
BTW I already found beta 0801 BIOS with AGESA 1.0.0.6 for my board. Fun starts tomorrow as I received my board today, but have no time to build it yet!

Enjoy! I think The Witcher 3 which I know you like a lot, could be a good test for your PC. In fact The Witcher 1, which I have on PC, and The Witcher 2 -have the X360 and the PC version, as I usually purchased games for the PC too, just in case I built a new PC, like I did as of recently- are both demanding benchmarks, which is most surprising in the case of the original, because it has been a while.

 

[Malo]

I went with Asus Prime X370-Pro and G.Skill RipJaws V F4-3200C14D 2x8GB (Samsung B-die) memories to make sure I can hit 3200MHz memory. Cooling will be done by BeQuiet Silent Loop 240mm AIO and rest of the components will come from my current i7 4790k rig.
BTW I already found beta 0801 BIOS with AGESA 1.0.0.6 for my board. Fun starts tomorrow as I received my board today, but have no time to build it yet!

You'll have no issues with the Samsung b-die memory of course. I just tried 0801 this morning and there's definitely better results for other types of RAM. Support for this board from Asus has been..... sparse at best. After initially being able to run 2933Mhz with my 3200 Hynix TridentZ at release, later BIOS updates reduced it to 2133 or 2400 only and loose timings. I'm now able to get back up to 3066Mhz and will play with it some more to see if I can get timings tighter. So there's progress at least.

 

[DavidGraham]

More interesting info on the Ryzen patch for Rise Of The Tomb Raider.

Firstly, it seems there was a bug in the game that hampered performance on NV GPUs and Ryzen CPUs.

An optimization was identified in texture management that improves the combination of AMD CPU and NVIDIA GPU. Overhead was reduced by packing texture descriptor uploads into larger chunks.

So that addresses the noise made about worse performance with Ryzen and NV GPUs. It's something related to the game itself. In fact, the developer stated that the biggest gains with this patch will come on Ryzen + GeForce configurations.
https://www.pcper.com/reviews/Graphics-Cards/Rise-Tomb-Raider-Gets-Ryzen-Performance-Update

Secondly, the developer improved the overall multi-core performance of the game:

Rise of the Tomb Raider splits rendering tasks to run on different threads. By tuning the size of those tasks – breaking some up, allowing multicore CPUs to contribute in more cases, and combining some others, to reduce overheads in the scheduler – the game can more efficiently exploit extra threads on the host CPU.

In fact this optimization not only helped AMD CPUs, but Intel CPUs as well.

we went ahead and tested the Village level; a level that puts more stress on the CPU and will bottleneck a lot of players. And we are happy to report that this patch significantly improves performance in that level.

http://www.dsogaming.com/news/the-l...mproves-dx12-performance-in-cpu-bound-scenes/

 

[Silent_Buddha]

So basically they had to do something special to make it perform adequately on NV GPU + Ryzen CPU. Something they didn't have to do for NV+Intel, AMD+Intel, or AMD+Ryzen.

Regards,
SB

 

[Ethatron]

So basically they had to do something special to make it perform adequately on NV GPU + Ryzen CPU. Something they didn't have to do for NV+Intel, AMD+Intel, or AMD+Ryzen.

Sounds like normal business for a programmer.

 

[Clukos]

http://www.dsogaming.com/news/the-l...mproves-dx12-performance-in-cpu-bound-scenes/

One thing to note, they are comparing Intel Dx11 vs Intel Dx12 and not Dx12 v Dx12 with the previous patch and the current patch. The CPU improvements in Geothermal Valley from Dx11 to Dx12 existed long before the recent patch.

 

[CSI PC]

So basically they had to do something special to make it perform adequately on NV GPU + Ryzen CPU. Something they didn't have to do for NV+Intel, AMD+Intel, or AMD+Ryzen.

Regards,
SB

I think the context is specifically the DX12 implementation, and under DX11 probably find Nvidia has a bit more control in managing less ideal texture management *shrug*.
But I agree it is interesting this has more of a headache for Nvidia on Ryzen platform than Intel, I would not necessarily say this is just Nvidia's fault though, especially as there are DX12 games that run very well with Nvidia+Ryzen and follow AMD's trend:
Most recent testing comparing Nvidia and AMD in real world gaming scenarios and IMO with a reasonable game selection published May 26th: https://www.hardocp.com/article/2017/05/26/definitive_amd_ryzen_7_realworld_gaming_guide/1
Just note the 480 has been highly OC'd to represent 580 if looking at performance between 480 and reference 1060 ( the real interest IMO is the relative trend behaviour across all systems comparing AMD and Nvidia), also all CPUs overclocked to what is realistically achievable but notice Intel has slightly faster DDR4 as well so if looking at absolute CPU comparison figure they pretty good on Ryzen IMO.
Note they have been doing this testing for awhile so explains the games (pretty nice selection though as they avoid ones excessive biased for either Nvidia or AMD) and DDR4 settings.

It suggest there are no issues generally with Nvidia on Ryzen, apart from when we see situations such as the RoTR and some other games that are not really great examples of DX12 such as Total War:Warhammer/Deus Ex:Mankind/Hitman.
What does stand out though in their testing is just how much better Crossfire is on Intel with AMD GPUs compared to Ryzen, something is not right there which is strange considering one would expect CF 480+Ryzen would be top rather than having some big performance losses to Intel depending upon the game where they were also reasonably close with single 480.
So there are quirks for both AMD and Nvidia GPUs on Ryzen in certain game situations.
Cheers

 

[Jawed]

Firstly, it seems there was a bug in the game that hampered performance on NV GPUs and Ryzen CPUs.

So basically they had to do something special to make it perform adequately on NV GPU + Ryzen CPU. Something they didn't have to do for NV+Intel, AMD+Intel, or AMD+Ryzen.

Or, maybe, they had to undo the "optimisation" they made for NVidia's D3D12 driver to make it work properly on all CPUs.

EDIT: and I forgot, this is an NVidia-sponsored game. So the code that they wrote before Ryzen came out was probably code NVidia helped with. So the studio isn't necessarily the source of the original "optimisation", which it appears they have had to refine or undo.

 

[CSI PC]

Or, maybe, they had to undo the "optimisation" they made for NVidia's D3D12 driver to make it work properly on all CPUs.

EDIT: and I forgot, this is an NVidia-sponsored game. So the code that they wrote before Ryzen came out was probably code NVidia helped with. So the studio isn't necessarily the source of the original "optimisation", which it appears they have had to refine or undo.

Although that needs weighing up against the game sponsored was released as DX11 not DX12, which came as a later patch update and does not include some of the NVIDIA features that were only DX11.
Fair to say DX12 was done more for AMD with the emphasis on optimised with Async Compute (when the DX12 patch came out they heavily promoted this).

While it is an Nvidia sponsored game, integral to the engine is actually Crystal Dynamic's/AMD's enhanced TressFX 3 as Pure Hair, PCGamesHardware showed this actually provided additional visual fidelity with snow and hair quality on AMD compared to Nvidia, scroll down to section on Pure Hair on AMD/Nvidia where picture of Lara in the snow and can move slide bar left/right: http://www.pcgameshardware.de/Rise-...451/Specials/Grafikkarten-Benchmarks-1184288/
Point being I do not think too much can be made about it being a sponsored game considering these points.

TressFX/Pure Hair is also in Deus Ex engine which also runs pretty poorly under DX12 on Nvidia relative to AMD.
Some detail of it used for Deus Ex Universe but still applicable to RoTR and Deus Ex Mankind: https://www.scribd.com/document/258...se-Projects-TressFX-3-0#fullscreen&from_embed
Cheers

 

[CSI PC]

One thing to note, they are comparing Intel Dx11 vs Intel Dx12 and not Dx12 v Dx12 with the previous patch and the current patch. The CPU improvements in Geothermal Valley from Dx11 to Dx12 existed long before the recent patch.

The following RoTR Ryzen retest helps I feel to add more context to the situation.
I use PCLabs sparingly (seems sometimes to be more Nvidia biased so prefer to be careful using the site for context) but they did a nice comparison of DX12 RoTR Ryzen update in 3 areas with each version:

We measured performance in three test sites. One of them is our standard tests used in place of the GPU, and two additional Citadel Plaza and Valley Farmstead , are more challenging for the CPU
...
For the measurement of performance in the game Rise of the Tomb Raider we used two test platforms: the first one is a platform from Intel i7-6700K CPU @ 4.7 GHz, our basic configuration in gaming performance tests. In the second platform, based on the AMD, we used the system Ryzen 7 1800X clocked at 3.85 GHz.

Moreover, tests were used for two graphics cards:

- Zotac GeForce 1080 GTX AMP Extreme
- Sapphire Nitro 580 RX +.

Measurements were made in DirectX 12 settings and Very High (1920 × 1080 SSAA 4 ×, 16 × AF). Drivers AMD's Crimson relive Edition 17.5.2, and Nvidia's drivers GeForce Game Ready driver 382.33.

Will need translating (I just use the Google Chrome translate): http://pclab.pl/art74383.html
The gains in the more busy CPU intensive maps show nice increases with the change to texture management for Nvidia GPUs.

Cheers

 

[Clukos]

No info on the ram speed used (that's important up to a point) but they do say the 6700k was running at 4.7GHz and the 1800x at 3.85GHz.

Edit: Got some images from different parts of the game with fps info, ram was at 3466CL14 and the 1700 at 3.9GHz, game settings were 720p dx12 w/ very high preset

Spoiler

 

[Cyan]

No info on the ram speed used (that's important up to a point) but they do say the 6700k was running at 4.7GHz and the 1800x at 3.85GHz.

Edit: Got some images from different parts of the game with fps info, ram was at 3466CL14 and the 1700 at 3.9GHz, game settings were 720p dx12 w/ very high preset

Spoiler

which program do you use to show the fps?

 

[Cyan]

I think the context is specifically the DX12 implementation, and under DX11 probably find Nvidia has a bit more control in managing less ideal texture management *shrug*.
But I agree it is interesting this has more of a headache for Nvidia on Ryzen platform than Intel, I would not necessarily say this is just Nvidia's fault though, especially as there are DX12 games that run very well with Nvidia+Ryzen and follow AMD's trend:
Most recent testing comparing Nvidia and AMD in real world gaming scenarios and IMO with a reasonable game selection published May 26th: https://www.hardocp.com/article/2017/05/26/definitive_amd_ryzen_7_realworld_gaming_guide/1
Just note the 480 has been highly OC'd to represent 580 if looking at performance between 480 and reference 1060 ( the real interest IMO is the relative trend behaviour across all systems comparing AMD and Nvidia), also all CPUs overclocked to what is realistically achievable but notice Intel has slightly faster DDR4 as well so if looking at absolute CPU comparison figure they pretty good on Ryzen IMO.
Note they have been doing this testing for awhile so explains the games (pretty nice selection though as they avoid ones excessive biased for either Nvidia or AMD) and DDR4 settings.

It suggest there are no issues generally with Nvidia on Ryzen, apart from when we see situations such as the RoTR and some other games that are not really great examples of DX12 such as Total War:Warhammer/Deus Ex:Mankind/Hitman.
What does stand out though in their testing is just how much better Crossfire is on Intel with AMD GPUs compared to Ryzen, something is not right there which is strange considering one would expect CF 480+Ryzen would be top rather than having some big performance losses to Intel depending upon the game where they were also reasonably close with single 480.
So there are quirks for both AMD and Nvidia GPUs on Ryzen in certain game situations.
Cheers

Fallout 4 is so tough to run smoothly...it bottlenecks the CPU quite a bit.

 

[Clukos]

which program do you use to show the fps?

Rivatuner statistics server (RTSS), you can show a lot more than just FPS

 

[CSI PC]

Fallout 4 is so tough to run smoothly...it bottlenecks the CPU quite a bit.

Yeah good point and probably why they included it even though it is seen more as an Nvidia game due to its global illumination/god rays, that said HOCP primary context is how it performs across platforms rather than actual 480 vs 1060 review comparison so has a useful purpose here but it is usually a game I think should not be in specific GPU reviews.
Cheers

 

[Clukos]

Hit number 1 in the 3D Mark database with a 1700 in the Sky Diver benchmark: http://www.3dmark.com/3dm/20292499

 

[Gubbi]

Does anybody know if the DRAM controllers are in any way coupled to each individual CCX ?

I'm wondering because up until now, all announced SKUs have one DDR4 channel per CCX; two for Ryzen, 4 for Threadripper, 8 for Naples.

Also, the L3s are victim caches. If they just act as victim caches for the local CCX, they'll end up being less effective in the many CCX case. For example, on Threadripper, if you have 32 threads crunching data using a central data structure (that doesn't fit in L2), each CCX will evict the same data over time, ending up with 4 copies of the same data in the L3s, lowering effective capacity.

If, instead, the L3s are tied to the memory channels, each L3 would only hold a copy of data from that particular memory channel.

Cheers

 

[3dilettante]

Does anybody know if the DRAM controllers are in any way coupled to each individual CCX ?

I have not seen any testing that was done with chips that had 1 CCX inactivated show much of a difference that could be attributed to DRAM being affected. The controllers connect to the data fabric, which connects to the CCXs and other elements.

If, instead, the L3s are tied to the memory channels, each L3 would only hold a copy of data from that particular memory channel.

Cache benchmarks show the L3s act like 8MB victim caches local to the CCX they are part of. If they were statically linked a DRAM channels, then all 16MB on chip would be detectable as being lower-latency than DRAM rather than latency rising as it approaches 8. The bandwidth of inter-CCX data transfers was also described as being somewhat modest at the bandwidth of one DRAM channel, which seems undesirable to make 1/2 of all L3 accesses have to cross.

 

[Anarchist4000]

If, instead, the L3s are tied to the memory channels, each L3 would only hold a copy of data from that particular memory channel.

AMD mentioned address interleaving before in relation to the fabric and per CCX/node makes sense from a design standpoint. Thats why I've been looking at it like a mapped cluster. The issue is that "many" doesn't describe 2 CCXs very well so the current cache policies are limited. I'm unsure how well they can bond links between nodes. An arrangement like Intel's Cluster on Die makes far more sense with a mesh interconnect like AMD mentioned in an EETimes article a while back.

The policies also make sense for different environments. VMs and HPC would have different access patterns.

 

[Clukos]

The bios you mention is the one I have, yes. It is a 16MB file I installed from a pendrive.

Lots of new BIOS for MSI boards: https://forum-en.msi.com/index.php?topic=283344.0

Check if there's one for your board with AGESA 1.0.0.6