Discussion AMD Ryzen MegaThread! FAQ and Resources

[8350rocks]

Spoiler

Spoiler

AMD asked them to publish 1080, 1440, and 4K reviews together. They did not ask them to *not* do 1080p, they just asked them to show higher resolutions. This was clarified by Lisa Su in the reddit AMA and was characterized as Steve making a gross uncharacterized jab.

Other reviews confirm GN claims. For instance pcgamer claims: "AMD's first suggestion was to test at 1440p or 4K, which is complete bunk. Testing higher resolutions will absolutely put Ryzen's gaming performance closer to on par with Intel, but only in the sense that running higher resolutions shifts the bottleneck to the GPU. Even AMD's FX-series performs relatively close to Intel in many games, provided you're running at 4K. But if you want to know how Ryzen compares to Core i5/i7 when the GPU isn't the bottleneck, you need to test at lower resolutions."

Note that AMD didn't suggest them to test both 1080p and 4K (or 1440p), but to test only at 4K (or 1440p) to hide the problem by generating a GPU bottleneck.

Also, as noted by GN, AMD public demos of gaming on Ryzen were run exclusively at 4K generating a GPU bottleneck and hidding the problem to everyone.

Additionally..Steve from GN published a review stating that GPUs are better for video encoding. Except that they are not used for professional video encoding...at all.

GN is rigth. Funny enouh it is the own AMD which has been saying us for years that GPUs are better for that kind of workloads. I still recall when AMD announced to everyone that tasks as Handbrake would run faster on a GPU than in a CPU. The reason is that GPUs are more optimized for throughput whereas CPUs are more optimized for latency. This is also the basis for HSA (Heterogenenous System Architecture) which identifies CPUs with LCUs (Latency Compute Units) and GPUs with TCUs (Throughput Compute Units).

Steve also contradicted himself in his review...he ignored the minimum FPS where Ryzen was running neck and neck with all the Intel parts, but the maximum FPS was lower.

I saw him discussing minimum FPS also and they were lower than in an i5 in several titles.

Additionally, they tested in DX12 exclusively, used odd blender settings, and other weird methodology. Joker tested in DX11, and his results are reflected by other DX11 tests.

Not true. GN also tested DX11 titles. Joker tests were unprofessional (large marging of error) and irrelevant because he chose odd settings that generated a GPU bottleneck with the GPU at 99% of load.

This reddit nitpick of GN's review pretty much sums it up: https://www.reddit.com/r/Amd/comments/5xgonu/analysis_of_gamersnexus_r7_1800x_review_an_i5_in/

GN conclusions agree with all other reviews that also found that Ryzen is not good at gaming. There are a fair good consensus among reviewers.

PcGamer: "The AMD Ryzen 7: plenty of power, but underwhelming gaming".

Arsctechnica: "an excellent workstation CPU, but it doesn't game as hard as we hoped".

Extremetech: "Zen is an amazing workstation chip with a 1080p gaming Achilles heel"

And of course TomsHardware: "It’s hard to recommend the Ryzen 7 1800X over Intel's lower-cost quad-core chips for gaming, especially given the Core i7-7700K's impressive performance."

I think the best comment on that reddit thread was made by the user that writes: "aka: Joker got the biased results we wanted out of hundreds of other reviewers and that is why we like him and are throwing GN and others under the bus".

All of this leads me to a few conclusions regarding the launch:

The BIOS was very buggy, and this was reflected in numerous Asus reviews, Gigabyte seemed to fare much better overall in that regard, especially in regards to memory timings and higher memory clock support.

Problems have been reproduced on non-Asus mobos.

The CCX cross threading is a major issue, supposedly windows is patching it next month, we will see. Windows 7 posts gaming numbers for Ryzen that are anywhere from 10-16% better according to the Stilt. Performance in other areas sees a significant improvement as well.

The CCX approach with a split LLC is weird. It is AMD fault, not everyone else. In any case, I don't expect large gains from patching Windows. Recall the Bulldozer scheduler, it improved performance by about 2--5%.

Agree 100% nicely said. I actually expected Fans to react this way but i never ever expected Amd to do this stuff i really didn't i thought that type of stuff was in the past(Bulldozer days).

Eitherway if you can watch Jayz latest video he did show that Ryzen is not being reported right in heaven benchmark and heaven benchmark was only using 1 core. If games are doing something like that it could be a reason why. I mean Ryzen does really well in everything but gaming that is why i find the results to be odd.

" No one talks about Cinebench scores anymore since Ryzen is now beating its counterpart there" I do Ryzen is great for everything but gaming where its mediocre.

"IS NOT RYZEN'S GAMING LINE OF PROCESSORS"

Then what will be the 4 and 6 core parts will look worse and i doubt they will be clocked higher then 300Mhz over the 1800X. 6 core part will be aimed at I5 prices where the I5 can be OC to 4.5Ghz easily, the 4 core part will look better being aimed at the I3 and i do think budget gamers will love that part.

From various comments you just made, are you telling me that if someone came to you and said all I do is video game, its all I'm interested in and I want to build the best video gaming system possible then you would hose them and building a system around the i7 6900K??? You wouldn't be in business for very long doing that to customers and then quote a couple outlier games where the i7 6900K beast the i7 7700K. Lets be honest, if someone asked you for gaming system you would recommend the i7 7700K over the i7 6900K for gaming as most of the time it has better gaming performance - and that isn't even taking into account the insane price Intel puts on its i7 6900K. Right now 8 core 16 thread processors are not the gaming line of processors, they are the workstation class of processors.

The gaming processors are the 4 core 8 thread and maybe even 6 core 12 thread processors. The gaming processors for Ryzen will be the R5 4 and 6 core processors, and they should boast better overclocking potential than their massive R7 brothers. As many others have stated as well Ryzen's gaming issue(s) aren't Bulldozer's issue. Bulldozer's low gaming performance was easy to understand from the get go, all one had to do was run any benchmark like Cinebench, look at the very low single core performance and bingo that is why it doesn't game in the enthusiast market. Ryzen is totally different, from all the benchmarks we have, with the powerhouse it is it should be gaming much, much better. Something is currently hobbling it, but it is far from the hardware issues Bulldozer had. Ryzen has very good IPC and very good single core execution, once the issue is found that is crippling its 1080p gaming potential it should hit the same gaming benchmarks as Skylake based on its very impressive workstation performance.

R5 Ryzen processors will bring AMD back into enthusiast gaming, but they won't dominate there, where they have a chance of dominating is with their R7 and Naples line in workstation and server application.

No i'd have them get a 7700K and according to IPC test's and frequency scaling on Ryzen OC i'd still recommend the 7700K as the best gaming processor 300mhz more on ryzen won't change that.

R5 will not bring Amd back in Enthusiast gaming it will make them a option again for people who do more then gaming. R3 will be Amd's major win CPU.

i5s are not even the recommended gaming processor anymore. Seriously...4 core i7s are what people are recommending for high end gaming.

As for enthusiast gaming, that depends, does it not? I intend to go 4K this time around, and I may run CF Vega GPUs to do 3 monitors...(still undecided about multi-monitor...to be honest...)

1440p Ryzen is fine, the differences are not substantial enough to warrant paying more for a 7700k. If you only play at 1080p...then maybe you look at a 7700k. I would only say sure if literally all you did was gaming, not capture software, no streaming, and no additional apps in the background. Simply because the 7700k is very nearly bottlenecking at 1080p in many games, and Ryzen has enough cores it would not impact performance in the same way.

Additionally...the 1% and 0.1% low FPS numbers are typically in favor of Ryzen, meaning a much smoother overall experience in spite of a 15-20% deficit in frames compared to a 5 GHz 7700k.

You said,"i5s are not even the recommended gaming processor anymore."

Intel is going to release new Intel 8th generation so i5 will have hyper-threading feature later this year. that is what I read it on Intel's plan on news. I assume i5 will be back with recommendation for gaming. But i5-7600k still strong and long time for 1080/1440p with decent graphic card. Also, I believe Ryzen 5 will challenge i5.

-DG

No, i5s will not receive hyper threading. They will just up core counts to 6 cores supposedly on 8th Gen in 2018 when it launches in Q2.

 

[8350rocks]

another issue is Intel will be releasing 8th gen core chips this year, which is said to be %15 faster than previous chips. while AMD has just released a new chip barely on par with intel 6th-7th gen, with the release of the new series AMD will be falling short again.

So AMD has again some rough times ahead. They will be strugling with devolopers, also will be releasing new drivers and BIOS updates and when they finally have a stable product ready for the market, will face a new opponent. This can cause AMD to loose its edge they have today.

Right now they may be selling a lot of product yet a lot more people like myself will be waiting to see the final results to make a purchase. I have been purchasing AMD chips since Athlon and I will again bu a ryzen chip in the next months but not before I see the platform is stabilized by means of operating system, BIOS and software updates.

If by this year you mean, in the next 15 months. then sure.

As it stands, Intel announced 8th Gen coming in Q2 2018.

 

[jaymc]

Another sales pitch for zen or naples is built in hardware security, something that Intel is not offering at the moment:
SME (Secure Memory Encryption), SEV (Secure Encrypted Virtualization) and hardware based SHA, powered by a security co-processor.

This could help them break back into the server market. Along with the fact that they (AMD) can provide a full solution as in CPU/GPU package deal... Zen/Vega... at cheaper rates than if the customer had to purchase both solutions seperately from Intel/Nvidia.

 

[juanrga]

Ok, so I took TechSpot's review, and did the math:

Out of 32 tests (16 games at 1080 and 1440), 15 resulted in improved performance with SMT disabled. A few got worse, and the rest had no effect.

But if you consider they solve the SMT issue, then only those games who had better performance without it will see gains, while the others will not (since they look already optimized, or has no effect). So, I did the difference in each scenario where SMT hurts performance, and found this (sorry for the bad table):

SMT - no SMT
min/avg - min/avg == %min/%avg (game)
56/94 - 59/96 == 5,4/2 (hitman)
50/71 - 53/72 == 6/1 (civ)
47/70 - 52/70 == 10,6/0 (civ)
133/166 - 146/170 == 9,8/2,4 (ow)
95/136 - 98/139 == 3,2/2,2 (ow)
104/126 - 106/131 == 1,9/4 (gears 4)
102/126 - 105/130 == 2,9/3,2 (gears 4)
57/79 - 58/93 == 1,8/17,7 (deus ex)
85/120 - 99/128 == 16,5/6,7 (F1 2016)
87/111 - 90/114 == 3,5/2,7 (F1 2016)
65/78 - 71/88 == 9,2/12,8 (TW Warh)
65/79 - 70/86 == 7,7/8,9 (TW Warh)
87/104 - 96/111 == 10,3/6,7 (GTA V)
60/88 - 65/100 == 8,3/13,6 (FC Primal)
61/86 - 66/93 == 8,2/8,1 (FC Primal)

which resulted in 6,98% average increase in min FPS, and 6,13% for average FPS. Highlights for F1 2016 at 1080p, for a 16.6% increase in minimum FPS, and Deus Ex at 1080p, with a 17.7% increase in average FPS.

Also, this only takes into consideration the games that won't take advantage of SMT instead of being hindered by it, in which case performance would grow even higher.

Solving the SMT problem will bring A LOT to the table. It won't solve the latency issues, but many reviews would have to be redone, and I think conclusions would change. AMD realy should have waited a bit more to launch Ryzen.

Nice work. Note that 6--7% average you got is the result of manually enabling/disabling SMT after testing which is the better option on a individual title per title basis. This is a best case scenario. No one is going to provide a scheduler with specific profiles for each one of the games that exist or will exist in a future. A real scheduler will provide a schedule on the fly less optimal. Thus a supposed future patch will provide an average performance gain of about 5% or less.

Recall what happened with the Bulldozer patches for Windows. Installing them increased performance on average by a ridiculous 2--5%. We are in a similar scenario now.

This supposed 5% extra performance from some future Ryzen patch doesn't solve the ~20% deficit that RyZen has compared to BDW-E.

 

[jdwii]

Interesting read on TR's take on Ryzen gaming justifying it's not "bad" as some make it believe (which I agree with) based on averages and min FPS:

http://techreport.com/review/31546/where-minimum-fps-figures-mislead-frame-time-analysis-shines

Cheers!

Sure its not bad its not like games are unplayable on even a 8150 hell a pentium with HT will game just fine in most cases its getting I3 performance with a 65$ price tag.

The point is its 20% slower in GTA5 in both FPS and frame times in the article you mentioned and in crysis 3 its close enough imo if perhaps we get a SMT patch with Windows 10 maybe things will look a bit better.

My main issue is with games that can use 8 cores the 1800X still loses to a 7700K if that isn't over scheduling issues and it might be then i can NOT recommend a Ryzen part over a I7 even more so when it cost the same for a 7700K as a 1700 which OC's terrible on REALISTIC(Sorry someone posted some crazy results and called Ryzen a winner cause it could OC to 5.8 or something with LN2 just about spit water on my monitor) cases like water or air cooling

"Does Ryzen 7 REALLY suck for gamers?" Jayztwocents just in case some people might think he is paid off or some other conspricy theory
https://www.youtube.com/watch?v=8-mMBbWHrwM

 

[jdwii]

Ok I think this settles it quite nicely:
https://www.youtube.com/watch?v=ylvdSnEbL50

Here Adored TV have done an investigation into the idea that low resolution tests = future performance with higher end gpus.

What they have found is: Core i5 2500k was around 10% faster than FX 8350 on average in games at launch at 1080p. At 480p it was 17% faster on average using a GTX 680. Moving forward to the original Titan, the gap at 1080p and 480p remains pretty much the same. Forward to today using Titan Pascal- the gap in the most modern titles is actually *reversed* and the FX8350 is 10% faster on average than the 2500k. It basically proves that even the low ipc 'failure' that is the FX cpu's in fact have improved with age better than the 'best' gaming cpu from Intel of the same geneartion due to having more threads. If you still think a 4 thread i5 is a better investment than Ryzen in games today I don't know what more I can say but basically more cores = better longevity.

Compare a 8150 to a 2500K also the 8150 cost 270$ not 230$ so keep that in mind you should be comparing the 8120 i think to a 2500K and in that case no the 2500K will beat it even more so since the 2500K could OC to 4.4-4.6 for a decent amount of people, Also it took years for piledriver to even look this good and its still trading blows with a I3 today so i can not say its worth it to have lower performance for 3-4 years then again that's up to the builder if it is.

I'd much rather have fantastic performance for 4 years then to have mediocre performance for 4 years and keep up with a 130$ processor the 5th year.

If a SMT patch or driver/bios fix happens and solves these issues i will however recommend the 1700 or R5 to gamers i really think even now a hypothetical R3 would do nice for budget gamers if the price is at 130. My main issue was never with Ryzen but with how Amd and the fanbase was acting.

 

[juanrga]

Hopefully it is for IP reasons that they can't use Intels HTT system an not pure stubbornness.

It's not quite the same though (correct me if I'm wrong here) Where Intel is using shared level 3 cache AMD is using 8mb shared per CCX... this means keeping treads of similar workloads on the same CCX or suffering a cache copy (from ccx to ccx) penelty...

i.e. so maybe some more elaborate scheduling system is required...

Or maybe the design with a partitioned LLC linked via a slow interconnect is just ridiculous to start and no magic scheduler will hide the deficiencies of that. I have been thinking on some other similar design and don't remember anything. Everyone, Intel, ARM, IBM, Sun,... uses unified LLC.

For practical purposes Ryzen is working as two quad-core CPUs glued together in a die.

 

[-Fran-]

Or maybe the design with a partitioned LLC linked via a slow interconnect is just ridiculous to start and no magic scheduler will hide the deficiencies of that. I have been thinking on some other similar design and don't remember anything. Everyone, Intel, ARM, IBM, Sun,... uses unified LLC.

For practical purposes Ryzen is working as two quad-core CPUs glued together in a die.

Q6600 anyone?

In any case, Juan, it doesn't matter how awful or elegant a solution is; if it works, deal with it. Now, Intel has the better and more elegant solution it seems, but if it weren't the better performing, people wouldn't care.

The Phenom 1 (good ol' Barcie) was a "true" core and a very nice solution, but the Q6600 was better (and other Kentsfield successors), so no one really cared it was 2 dual cores slapped together.

I think you're missing the point with the CCX arrangement and it's a very clever one. I won't share my opinion, because it is plenty obvious and you might already even know why.

Cheers!

 

[jaymc]

Ok, so I took TechSpot's review, and did the math:

Out of 32 tests (16 games at 1080 and 1440), 15 resulted in improved performance with SMT disabled. A few got worse, and the rest had no effect.

But if you consider they solve the SMT issue, then only those games who had better performance without it will see gains, while the others will not (since they look already optimized, or has no effect). So, I did the difference in each scenario where SMT hurts performance, and found this (sorry for the bad table):

SMT - no SMT
min/avg - min/avg == %min/%avg (game)
56/94 - 59/96 == 5,4/2 (hitman)
50/71 - 53/72 == 6/1 (civ)
47/70 - 52/70 == 10,6/0 (civ)
133/166 - 146/170 == 9,8/2,4 (ow)
95/136 - 98/139 == 3,2/2,2 (ow)
104/126 - 106/131 == 1,9/4 (gears 4)
102/126 - 105/130 == 2,9/3,2 (gears 4)
57/79 - 58/93 == 1,8/17,7 (deus ex)
85/120 - 99/128 == 16,5/6,7 (F1 2016)
87/111 - 90/114 == 3,5/2,7 (F1 2016)
65/78 - 71/88 == 9,2/12,8 (TW Warh)
65/79 - 70/86 == 7,7/8,9 (TW Warh)
87/104 - 96/111 == 10,3/6,7 (GTA V)
60/88 - 65/100 == 8,3/13,6 (FC Primal)
61/86 - 66/93 == 8,2/8,1 (FC Primal)

which resulted in 6,98% average increase in min FPS, and 6,13% for average FPS. Highlights for F1 2016 at 1080p, for a 16.6% increase in minimum FPS, and Deus Ex at 1080p, with a 17.7% increase in average FPS.

Also, this only takes into consideration the games that won't take advantage of SMT instead of being hindered by it, in which case performance would grow even higher.

Solving the SMT problem will bring A LOT to the table. It won't solve the latency issues, but many reviews would have to be redone, and I think conclusions would change. AMD realy should have waited a bit more to launch Ryzen.

Nice work. Note that 6--7% average you got is the result of manually enabling/disabling SMT after testing which is the better option on a individual title per title basis. This is a best case scenario. No one is going to provide a scheduler with specific profiles for each one of the games that exist or will exist in a future. A real scheduler will provide a schedule on the fly less optimal. Thus a supposed future patch will provide an average performance gain of about 5% or less.

Recall what happened with the Bulldozer patches for Windows. Installing them increased performance on average by a ridiculous 2--5%. We are in a similar scenario now.

This supposed 5% extra performance from some future Ryzen patch doesn't solve the ~20% deficit that RyZen has compared to BDW-E.

Does Dx12 not allow the Game Engine's themselves to allocate processor resources specifically... on a game by game basis. (instead of windows doing it that is) ??
So each game can uniquely tune the cpu resources to suit each games individual needs ?

 

[jaymc]

Hopefully it is for IP reasons that they can't use Intels HTT system an not pure stubbornness.

It's not quite the same though (correct me if I'm wrong here) Where Intel is using shared level 3 cache AMD is using 8mb shared per CCX... this means keeping treads of similar workloads on the same CCX or suffering a cache copy (from ccx to ccx) penelty...

i.e. so maybe some more elaborate scheduling system is required...

Or maybe the design with a partitioned LLC linked via a slow interconnect is just ridiculous to start and no magic scheduler will hide the deficiencies of that. I have been thinking on some other similar design and don't remember anything. Everyone, Intel, ARM, IBM, Sun,... uses unified LLC.

For practical purposes Ryzen is working as two quad-core CPUs glued together in a die.

Agreed... an maybe this is how windows an the game engines need to view it... as two quad cores that is.

Doesn't sound as good as an octa-core does it

An this is also why simply using intel's cpuid's would not be the most efficient solution...

@yuka
I guess you mean the fact that they can replace one of the ccx's with a vega gpu... it's pretty clever in that sense for sure mate

http://www.pcgamer.com/amds-shows-off-how-crazily-zen-can-scale/

 

[eathdemon1]

Hopefully it is for IP reasons that they can't use Intels HTT system an not pure stubbornness.

It's not quite the same though (correct me if I'm wrong here) Where Intel is using shared level 3 cache AMD is using 8mb shared per CCX... this means keeping treads of similar workloads on the same CCX or suffering a cache copy (from ccx to ccx) penelty...

i.e. so maybe some more elaborate scheduling system is required...

Or maybe the design with a partitioned LLC linked via a slow interconnect is just ridiculous to start and no magic scheduler will hide the deficiencies of that. I have been thinking on some other similar design and don't remember anything. Everyone, Intel, ARM, IBM, Sun,... uses unified LLC.

For practical purposes Ryzen is working as two quad-core CPUs glued together in a die.

Agreed... an maybe this is how windows an the game engines need to view it... as two quad cores that is.

Doesn't sound as good as an octa-core does it

An this is also why simply using intel's cpuid's would not be the most efficient solution...

@yuka
I guess you mean the fact that they can replace one of the ccx's with a vega gpu... it's pretty clever in that sense for sure mate

http://www.pcgamer.com/amds-shows-off-how-crazily-zen-can-scale/

for that to happen, ms would have to add it to the home version of windows 10. right now multi cpu support is a pro version locked.

 

[salgado18]

I guess you mean the fact that they can replace one of the ccx's with a vega gpu... it's pretty clever in that sense for sure mate

http://www.pcgamer.com/amds-shows-off-how-crazily-zen-can-scale/

I so love this part

Running a seismic analysis workload that involves computationally intensive 3D wave equations, which taxes the entire system—cores, memory, and IO—Naples basically destroyed Intel's Xeon server.

Could this mean, and I'm just dreaming here, that it's possible a 12c/24t desktop 1900X exists in the future?

Also, how big of a gpu can fit in that space?

 

[juanrga]

Or maybe the design with a partitioned LLC linked via a slow interconnect is just ridiculous to start and no magic scheduler will hide the deficiencies of that. I have been thinking on some other similar design and don't remember anything. Everyone, Intel, ARM, IBM, Sun,... uses unified LLC.

For practical purposes Ryzen is working as two quad-core CPUs glued together in a die.

Q6600 anyone?

In any case, Juan, it doesn't matter how awful or elegant a solution is; if it works, deal with it. Now, Intel has the better and more elegant solution it seems, but if it weren't the better performing, people wouldn't care.

The Phenom 1 (good ol' Barcie) was a "true" core and a very nice solution, but the Q6600 was better (and other Kentsfield successors), so no one really cared it was 2 dual cores slapped together.

I think you're missing the point with the CCX arrangement and it's a very clever one. I won't share my opinion, because it is plenty obvious and you might already even know why.

My point is not about "elegance", but about the performance deficits. Note as well that I was the first guy that predicted the quad-core clusters... I did this so early as 2014 when the codename Zen was not public still. I predicted that correctly because I was able to infer AMD modular plans as a way to reduce R&D costs. What I couldn't imagine is that they would do a quad-core modular approach with a split LLC and then get trapped into the difficulties they are experiencing now.

 

[juanrga]

At least publicly AMD is setting the expectations for Naples low and realistic

http://www.eetimes.com/document.asp?doc_id=1331444

I don't doubt that Naples can be a formidable competitor in some specific niche (probably the memory bound ones) but as a general server chip I don't expect miracles. Competition is too strong.

 

[jaymc]

I guess you mean the fact that they can replace one of the ccx's with a vega gpu... it's pretty clever in that sense for sure mate

http://www.pcgamer.com/amds-shows-off-how-crazily-zen-can-scale/

I so love this part

Running a seismic analysis workload that involves computationally intensive 3D wave equations, which taxes the entire system—cores, memory, and IO—Naples basically destroyed Intel's Xeon server.

Could this mean, and I'm just dreaming here, that it's possible a 12c/24t desktop 1900X exists in the future?

Also, how big of a gpu can fit in that space?

I dunno.. good question how many ccx's can fit into an AM4 slot... are we maxed out at two ??

 

[010TheMaster010]

1 Ryzen die -> ~194mm^2
1 CCX -> 44mm^2

 

[jaymc]

1 Ryzen die -> ~194mm^2
1 CCX -> 44mm^2

Does that mean they could get 4 into the same space or 3 an a GPU into an AM4 ????

The mind boggles.... Whaaaaat !

So we could be looking at 12c/16t up to even 16c/32t theoretically speaking. Or even a 12c/16t with a GPU ? Would the Pin Grid Array deal with the throughput ?

So how many CCX's can an AM4 chipset actually support ??

AM4 Supports PCIe 3.0, up to 24 lanes

@salgado18 While were dreaming we might as well dream big...

 

[Nope 1151]

Could someone explain to me L3 and L4 (Memory card RAM) and the latency? Would it be possible for AMD to cancel CCX copying and skip straight to L4 cache?

Also: https://cdn.arstechnica.net/wp-content/uploads/2017/03/Forrest_Naples_Embargoed-Until-3_7_17-19-980x551.jpeg
Shots fire at Intel....

 

[jaymc]

"40%+ More Pins Than AM3+ & New Mounting Hole Spacing
Just like previous generation AMD sockets AM4 is a Pin Grid Array socket with zero insertion force. It maintains the same locking mechanism as the previous generation and takes up exactly the same area as AM3+, measuring in at 40x40mm. One of the most significant changes however is the number of pins. AM4 supports 1331 pins which is a very significant increase over the 942 of AM3+ and quite a bit more than Intel’s LGA1151. In fact this makes it the first PGA socket that we’ve seen from AMD to accommodate more contact pins than Intel’s LGA."

Can't help wondering if all these extra pins were to accommodate the increased throughput of adding extra CCX's... maybe the Ryzen 1900x is a 12c/24t part.

 

[gamerk316]

Ok, so I took TechSpot's review, and did the math:

Out of 32 tests (16 games at 1080 and 1440), 15 resulted in improved performance with SMT disabled. A few got worse, and the rest had no effect.

But if you consider they solve the SMT issue, then only those games who had better performance without it will see gains, while the others will not (since they look already optimized, or has no effect). So, I did the difference in each scenario where SMT hurts performance, and found this (sorry for the bad table):

SMT - no SMT
min/avg - min/avg == %min/%avg (game)
56/94 - 59/96 == 5,4/2 (hitman)
50/71 - 53/72 == 6/1 (civ)
47/70 - 52/70 == 10,6/0 (civ)
133/166 - 146/170 == 9,8/2,4 (ow)
95/136 - 98/139 == 3,2/2,2 (ow)
104/126 - 106/131 == 1,9/4 (gears 4)
102/126 - 105/130 == 2,9/3,2 (gears 4)
57/79 - 58/93 == 1,8/17,7 (deus ex)
85/120 - 99/128 == 16,5/6,7 (F1 2016)
87/111 - 90/114 == 3,5/2,7 (F1 2016)
65/78 - 71/88 == 9,2/12,8 (TW Warh)
65/79 - 70/86 == 7,7/8,9 (TW Warh)
87/104 - 96/111 == 10,3/6,7 (GTA V)
60/88 - 65/100 == 8,3/13,6 (FC Primal)
61/86 - 66/93 == 8,2/8,1 (FC Primal)

which resulted in 6,98% average increase in min FPS, and 6,13% for average FPS. Highlights for F1 2016 at 1080p, for a 16.6% increase in minimum FPS, and Deus Ex at 1080p, with a 17.7% increase in average FPS.

Also, this only takes into consideration the games that won't take advantage of SMT instead of being hindered by it, in which case performance would grow even higher.

Solving the SMT problem will bring A LOT to the table. It won't solve the latency issues, but many reviews would have to be redone, and I think conclusions would change. AMD realy should have waited a bit more to launch Ryzen.

Nice work. Note that 6--7% average you got is the result of manually enabling/disabling SMT after testing which is the better option on a individual title per title basis. This is a best case scenario. No one is going to provide a scheduler with specific profiles for each one of the games that exist or will exist in a future. A real scheduler will provide a schedule on the fly less optimal. Thus a supposed future patch will provide an average performance gain of about 5% or less.

Recall what happened with the Bulldozer patches for Windows. Installing them increased performance on average by a ridiculous 2--5%. We are in a similar scenario now.

This supposed 5% extra performance from some future Ryzen patch doesn't solve the ~20% deficit that RyZen has compared to BDW-E.

Does Dx12 not allow the Game Engine's themselves to allocate processor resources specifically... on a game by game basis. (instead of windows doing it that is) ??
So each game can uniquely tune the cpu resources to suit each games individual needs ?

In theory, yes. But then developers are basically overriding the OS Scheduler, which can lead to a LOT of problems as new CPUs get released that aren't accounted for in code. It's also a lot of work to confirm that your custom partitioning is doing a better job then the OS scheduler is. I'd say in 99.9% of use cases, it's not worth the effort to go out of your way to optimize over the scheduler. Scheduling is inherently an OS function, and unless there's a very specific reason to, it's generally best to let the scheduler handle scheduling threads.

For example, Ubisoft has a tendency to manually lock threads to cores. This is bad design practice, as they are making assumptions about CPUs that may not be true in the future, which could affect performance or even outright prevent titles from running. Case in point, some of their games lock a major thread to the fourth processor core (Core 3). Depending on CPU topology, this could actually be a logical core, rather then a physical one, and this approach could easily kill performance if the topology for a SMT capable CPU is in the form of 00112233, rather then 01230123. Like Ryzen might be. Whoops.

 

[jaymc]

Could someone explain to me L3 and L4 (Memory card RAM) and the latency? Would it be possible for AMD to cancel CCX copying and skip straight to L4 cache?

Also: https://cdn.arstechnica.net/wp-content/uploads/2017/03/Forrest_Naples_Embargoed-Until-3_7_17-19-980x551.jpeg
Shots fire at Intel....

Ok, so I took TechSpot's review, and did the math:

Out of 32 tests (16 games at 1080 and 1440), 15 resulted in improved performance with SMT disabled. A few got worse, and the rest had no effect.

But if you consider they solve the SMT issue, then only those games who had better performance without it will see gains, while the others will not (since they look already optimized, or has no effect). So, I did the difference in each scenario where SMT hurts performance, and found this (sorry for the bad table):

SMT - no SMT
min/avg - min/avg == %min/%avg (game)
56/94 - 59/96 == 5,4/2 (hitman)
50/71 - 53/72 == 6/1 (civ)
47/70 - 52/70 == 10,6/0 (civ)
133/166 - 146/170 == 9,8/2,4 (ow)
95/136 - 98/139 == 3,2/2,2 (ow)
104/126 - 106/131 == 1,9/4 (gears 4)
102/126 - 105/130 == 2,9/3,2 (gears 4)
57/79 - 58/93 == 1,8/17,7 (deus ex)
85/120 - 99/128 == 16,5/6,7 (F1 2016)
87/111 - 90/114 == 3,5/2,7 (F1 2016)
65/78 - 71/88 == 9,2/12,8 (TW Warh)
65/79 - 70/86 == 7,7/8,9 (TW Warh)
87/104 - 96/111 == 10,3/6,7 (GTA V)
60/88 - 65/100 == 8,3/13,6 (FC Primal)
61/86 - 66/93 == 8,2/8,1 (FC Primal)

which resulted in 6,98% average increase in min FPS, and 6,13% for average FPS. Highlights for F1 2016 at 1080p, for a 16.6% increase in minimum FPS, and Deus Ex at 1080p, with a 17.7% increase in average FPS.

Also, this only takes into consideration the games that won't take advantage of SMT instead of being hindered by it, in which case performance would grow even higher.

Solving the SMT problem will bring A LOT to the table. It won't solve the latency issues, but many reviews would have to be redone, and I think conclusions would change. AMD realy should have waited a bit more to launch Ryzen.

Nice work. Note that 6--7% average you got is the result of manually enabling/disabling SMT after testing which is the better option on a individual title per title basis. This is a best case scenario. No one is going to provide a scheduler with specific profiles for each one of the games that exist or will exist in a future. A real scheduler will provide a schedule on the fly less optimal. Thus a supposed future patch will provide an average performance gain of about 5% or less.

Recall what happened with the Bulldozer patches for Windows. Installing them increased performance on average by a ridiculous 2--5%. We are in a similar scenario now.

This supposed 5% extra performance from some future Ryzen patch doesn't solve the ~20% deficit that RyZen has compared to BDW-E.

Does Dx12 not allow the Game Engine's themselves to allocate processor resources specifically... on a game by game basis. (instead of windows doing it that is) ??
So each game can uniquely tune the cpu resources to suit each games individual needs ?

In theory, yes. But then developers are basically overriding the OS Scheduler, which can lead to a LOT of problems as new CPUs get released that aren't accounted for in code. It's also a lot of work to confirm that your custom partitioning is doing a better job then the OS scheduler is. I'd say in 99.9% of use cases, it's not worth the effort to go out of your way to optimize over the scheduler. Scheduling is inherently an OS function, and unless there's a very specific reason to, it's generally best to let the scheduler handle scheduling threads.

For example, Ubisoft has a tendency to manually lock threads to cores. This is bad design practice, as they are making assumptions about CPUs that may not be true in the future, which could affect performance or even outright prevent titles from running. Case in point, some of their games lock a major thread to the fourth processor core (Core 3). Depending on CPU topology, this could actually be a logical core, rather then a physical one, and this approach could easily kill performance if the topology for a SMT capable CPU is in the form of 00112233, rather then 01230123. Like Ryzen might be. Whoops.

Thanks for that interesting stuff..
When AMD said they were working with developers directly to show them how to get the most from Ryzen.. I just presumed this is what they meant. I presumed software would identify the CPU in question and use an if/then to allocate treads.

Can't help wondering what optimizations AMD are talking about when they speak of working directly with game developers to achieve... Or are they simply trying to get everyone to develop games to make use of more cores...

 

[gamerk316]

[Thanks for that interesting stuff..
When AMD said they were working with developers directly to show them how to get the most from Ryzen.. I just presumed this is what they meant. I presumed software would identify the CPU in question and use an if/then to allocate treads. I wonder what they are doing with these game developers in that case ?

I've always been in the camp of, barring some serious mismanagement on the part of the OS, that the OS scheduler should handle the scheduling details. There's simply too many conditions out there where wrong assumptions can kill performance.

Now, that isn't to say you might have some code path specific for a CPU architecture to go down a different computing path for one reason or another. For example, if I were doing AES workloads I would probably have a dedicated path for Ryzen to take advantage of it's AES engine. But as far as scheduling goes, I leave it to the scheduler to manage.

 

[jaymc]

Do you think these optimizations they speaking of will have much of an impact on fps in games running on Ryzen ?

I know thats the million dollor question... I know they are trying to get people to code for more cores. Trying to get developers to use dx12 better and vulken. But coding for more cores will benefit Intel multi core CPU's just as much. Can't help wondering what these optimizations are that they speak of when it comes to game developers.

 

[Nope 1151]

I think it depends on what AMD meant by "optimizations".
If it is a-la faildozer as in, "Code for moar cores" its probably not going to happen yet.
If it is a scheduling bug, maybe.

Side note: Has anyone seen the price drops on previous AM3+ board and mobos? TIME TO FIX THAT x6 1045 HERE I COME! (DDR3 ram too )

 

[gamerk316]

Do you think these optimizations they speaking of will have much of an impact on fps in games running on Ryzen ?

I know thats the million dollor question... I know they are trying to get people to code for more cores. Trying to get developers to use dx12 better and vulken. But coding for more cores will benefit Intel multi core CPU's just as much. Can't help wondering what these optimizations are that they speak of when it comes to game developers.

From my perspective, if the CPU is doing enough where the GPU is constantly loaded, there really isn't much more you can do. Point being, if four cores is enough to max the GPU, using more cores won't have any real benefit by itself.

Now, if other things are going on in the system, having more threads doing less work is a good thing, as you reduce the chance of any individual core getting overloaded. But for a *single* application, if the CPU is doing enough, then there really isn't much point to complicate things for little to no performance benefit.

The *real* advantage in DX12 is the fact you no longer have a strict GPU pipeline; you can use unused resources in one stage to start the next, which could drastically improve performance. But all the other low level stuff won't amount to much while adding a LOT of difficulty to code, and will likely fall out of favor over time.