Discussion AMD Ryzen MegaThread! FAQ and Resources

[8350rocks]

That video provides a lot of good data, although I don't like the set up itself used.

Still, it can be used as proof that testing with nVidia is detrimental to AMD until the drivers are optimized for Zen. Reviewers should take notice on that, just in case, when accounting all data into the conclusions.

Cheers!

It really should be noted pretty heavily that Nvidia drivers are horribly unoptimized for Ryzen.

 

[gamerk316]

Hmm with Ryzen's higher thread count, and good core performance, we might actually get Dev's better adopting higher core/thread usages rates for games. So people who bought those FX 8000s will finally get good use out of their mutli threaded processors. BUT I bet they'll still be screwed because they'll be stuck with slower DDR3 RAM, and also the architecture of the FX line is just not good.

I've been through this for what, half a decade now? Games are a perfect example of a task that doesn't scale well to multiple CPU cores.

Now, DX12 certainly helps due to not requiring a single rendering thread to do all the work, but for CPUs where that thread wasn't itself a bottleneck (Intel CPUs mostly), you don't really get any benefit in performance.

It really should be noted pretty heavily that Nvidia drivers are horribly unoptimized for Ryzen.

Drivers shouldn't care about the CPU. My guess is NVIDIA's driver layer is more CPU intensive, which exposes Ryzen somewhat.

 

[juanrga]

@Juan what would lead you to believe all these given reviewers are "having a dozen of background apps open: three chat sessions, encoding video, playing the game, writing the report, looking for OS upgrades,..."
I can't think to believe all would be doing that. Furthermore if that were true would that be bad? I often have background apps open while I game and I know others who do even more than I, so if this were true and all these reviewers have all these things going on at once (a stretch) it would be another selling point for R7 processors, no? I believe this very comparison was made by AMD with a 1700(X?) Vs a 6700k at one of their earlier events.
The problem with comparing 8c vs 8c in your screenshots is price. For 30 some dollars over 7700k with cooler you get an overclocked 1700 with comparible max frames, lower average, but smoother gameplay, plus the other advantages of the 16 threads.
Thoughts/criticism welcome, cheers all

That they are overloading the quad with background stuff is the logical conclusion when the results of those 'reviews' contradict mainstream professional reviews.

I remember a pre-launch demo from AMD that used a background task to claim better gameplay. I recall some of those youtubers did make explicit he was playing the games and doing some other thing. I also recall reading in some part (reddit? another forum?) some user saying those youtuber reviews were reporting "real gameplay", because he cheated and encoded videos whereas gaming.

Different people has different needs and take different attitudes regarding pricing.

 

[juanrga]

I feel like i should say once again that kabylake+skylake and even haswell benefit from faster memory in gaming to. Digital Foundry proved this several times.

Proof down below

http://www.eurogamer.net/articles/digitalfoundry-2017-intel-core-i5-7600k-review

Ryzen getting better gaming performance with faster memory is nothing special IMHO even more so when its hard to even get even 3000mhz.

Indeed, memory-bound games will benefit from faster ram, albeit RyZen gets slightly higher gains with OC memory because it has a worse memory controller

It also happens with ranks

 

[8350rocks]

Hmm with Ryzen's higher thread count, and good core performance, we might actually get Dev's better adopting higher core/thread usages rates for games. So people who bought those FX 8000s will finally get good use out of their mutli threaded processors. BUT I bet they'll still be screwed because they'll be stuck with slower DDR3 RAM, and also the architecture of the FX line is just not good.

I've been through this for what, half a decade now? Games are a perfect example of a task that doesn't scale well to multiple CPU cores.

Now, DX12 certainly helps due to not requiring a single rendering thread to do all the work, but for CPUs where that thread wasn't itself a bottleneck (Intel CPUs mostly), you don't really get any benefit in performance.

It really should be noted pretty heavily that Nvidia drivers are horribly unoptimized for Ryzen.

Drivers shouldn't care about the CPU. My guess is NVIDIA's driver layer is more CPU intensive, which exposes Ryzen somewhat.

No, Nvidia drivers are HUGE because they use a software scheduler instead of a hardware scheduler. This is why their DX11 tends to be better, and is also why their DX12 sucks. DX12 overloads their software scheduler with loads of draw calls, while DX12/Vulkan is significantly better suited to AMD's robust hardware scheduler. Which also contributes to their slightly higher power draw. The hardware scheduler draws more power than the software scheduler, but is much more effective with a higher amount of draw calls.

The issue with Nvidia is that AMD can more easily overload their software scheduler in their driver with the significantly higher amount of threads.

Here is a tremendous video about this difference: https://www.youtube.com/watch?v=nIoZB-cnjc0

 

[juanrga]

It really should be noted pretty heavily that Nvidia drivers are horribly unoptimized for Ryzen.

Drivers shouldn't care about the CPU. My guess is NVIDIA's driver layer is more CPU intensive, which exposes Ryzen somewhat.

No, Nvidia drivers are HUGE because they use a software scheduler instead of a hardware scheduler.

And this software scheduler is CPU intensive and affects the CPU with higher latency: aka RyZen.

 

[-Fran-]

And this software scheduler is CPU intensive and affects the CPU with higher latency: aka RyZen.

What is a "CPU with higher latency"? I'm already looking forward for your definition

Cheers!

 

[8350rocks]

It really should be noted pretty heavily that Nvidia drivers are horribly unoptimized for Ryzen.

Drivers shouldn't care about the CPU. My guess is NVIDIA's driver layer is more CPU intensive, which exposes Ryzen somewhat.

No, Nvidia drivers are HUGE because they use a software scheduler instead of a hardware scheduler.

And this software scheduler is CPU intensive and affects the CPU with higher latency: aka RyZen.

Then why is the DX11 difference typically the equivalent to the gap in IPC + the gap in clockspeed, but you see a regression in DX12, where there should be no impact?

 

[juanrga]

It really should be noted pretty heavily that Nvidia drivers are horribly unoptimized for Ryzen.

Drivers shouldn't care about the CPU. My guess is NVIDIA's driver layer is more CPU intensive, which exposes Ryzen somewhat.

No, Nvidia drivers are HUGE because they use a software scheduler instead of a hardware scheduler.

And this software scheduler is CPU intensive and affects the CPU with higher latency: aka RyZen.

Then why is the DX11 difference typically the equivalent to the gap in IPC + the gap in clockspeed, but you see a regression in DX12, where there should be no impact?

I will bold the relevant part

One reason why we see things playing out the way they do is because the goal and performance-improving functions of low-overhead APIs have been misunderstood. It’s been known for years that Nvidia GPUs are often faster with lower-end Intel or AMD CPUs (pre-Ryzen) than AMD’s own GPUs are. Part of the reason for this is because Nvidia’s own DX11 drivers implement multi-threading, whereas AMD’s do not. That’s one reason why, in games like Ashes of the Singularity, AMD’s GPU performance skyrocketed so much in DX12. But fundamentally, DX12, Vulkan, and Mantle are methods of compensating for weak single-threaded performance (or for spreading out a workload more evenly so it isn’t bottlenecked by a single thread).

https://www.extremetech.com/gaming/246377-new-directx-11-vs-directx-12-comparison-shows-uneven-results-limited-improvements

Also read that jdwii wrote at the beginning of page 26.

 

[jaymc]

This is from the same quote just after the highlighted part:
"(or for spreading out a workload more evenly so it isn’t bottlenecked by a single thread)."

If used correctly more performance can be gained from either architecture, Intel or AMD's CPU's or GPU's for that matter.

It's makes sense to use this technology regardless of who's CPU or GPU is using it..

As it is progress no matter what way you look at it

 

[Nope 1151]

Two days till 4/11.

Anyone waiting on R5?

 

[salgado18]

Two days till 4/11.

Anyone waiting on R5?

Nope 😀

I'm waiting for the R3, until then I can get enough money, and upgrade my i3 onto the platform with a cheaper CPU.

 

[jdwii]

Two days till 4/11.

Anyone waiting on R5?

Nope 😀

I'm waiting for the R3, until then I can get enough money, and upgrade my i3 onto the platform with a cheaper CPU.

I'm really looking forward to the R3 too. I can see the R3 being in 400-600$ Gaming PC's. Entry level spot that i enjoy building so much. I enjoy building budget rigs more then high-end ones.

 

[Nope 1151]

Well the lower end R3 are 4c 8t no?\
Edit:https://www.reddit.com/r/Amd/comments/64e3oe/ryzen_5_1600_cinebench/
Another Cbench leak. Pinch of salt recommended.

 

[gamerk316]

This is from the same quote just after the highlighted part:
"(or for spreading out a workload more evenly so it isn’t bottlenecked by a single thread)."

If used correctly more performance can be gained from either architecture, Intel or AMD's CPU's or GPU's for that matter.

It's makes sense to use this technology regardless of who's CPU or GPU is using it..

As it is progress no matter what way you look at it

There's only a finite amount of performance benefit you can gain by increasing threading; Amdhals Law remains in force, low level API or no. Also remember threading incurs it's own overhead; if NVIDIA already optimized it's DX11 driver path to mitigate low IPC, then I'd EXPECT the DX12 driver path to be ever so slightly slower.

 

[salgado18]

Well the lower end R3 are 4c 8t no?\
Edit:https://www.reddit.com/r/Amd/comments/64e3oe/ryzen_5_1600_cinebench/
Another Cbench leak. Pinch of salt recommended.

Nope 😀 (can do this all day)

The R3 are all 4c/4t. With same clocks as the other Ryzen, also unlocked, good IPC, and much lower price than i5's, i'd call them a huge win.

 

[jwcrellin]

I'm still waiting for valid, real world Ryzen 5 performance figs.

 

[jaymc]

R5 1500x 4 core, 8 thread, 2+2 part's. Still have 8mb level 3 cache per ccx.
https://www.game-debate.com/cpu/index.php?pid=2464&pid2=2223&compare=ryzen-r5-1500-vs-core-2-quad-q9705-3-16ghz

This should help gain some ground on the 7700k... what yis reckon ?

Also check this out:
In Photoshop: 8-Core MacPro vs AMD Ryzen 1700 PC
https://www.youtube.com/watch?v=No7eZZb3jdA

"Apple apologized for its Mac Pro line this week and promised to release “something great” in 2018 for professional creatives. Based on the results of this Tech Guy test, it seems Apple has quite a bit of catching up to do."
https://petapixel.com/2017/04/08/5660-mac-pro-crushed-photoshop-test-1530-pc-amd-ryzen/

 

[jdwii]

Well the lower end R3 are 4c 8t no?\
Edit:https://www.reddit.com/r/Amd/comments/64e3oe/ryzen_5_1600_cinebench/
Another Cbench leak. Pinch of salt recommended.

Nope 😀 (can do this all day)

The R3 are all 4c/4t. With same clocks as the other Ryzen, also unlocked, good IPC, and much lower price than i5's, i'd call them a huge win.

Something tells me we might not have L3 cache or as much as the 1400 does. I even wonder if the R3 is nothing more then the APU but with no iGPU.

 

[8350rocks]

Spoiler

It really should be noted pretty heavily that Nvidia drivers are horribly unoptimized for Ryzen.

Drivers shouldn't care about the CPU. My guess is NVIDIA's driver layer is more CPU intensive, which exposes Ryzen somewhat.

No, Nvidia drivers are HUGE because they use a software scheduler instead of a hardware scheduler.

And this software scheduler is CPU intensive and affects the CPU with higher latency: aka RyZen.

Then why is the DX11 difference typically the equivalent to the gap in IPC + the gap in clockspeed, but you see a regression in DX12, where there should be no impact?

I will bold the relevant part

One reason why we see things playing out the way they do is because the goal and performance-improving functions of low-overhead APIs have been misunderstood. It’s been known for years that Nvidia GPUs are often faster with lower-end Intel or AMD CPUs (pre-Ryzen) than AMD’s own GPUs are. Part of the reason for this is because Nvidia’s own DX11 drivers implement multi-threading, whereas AMD’s do not. That’s one reason why, in games like Ashes of the Singularity, AMD’s GPU performance skyrocketed so much in DX12. But fundamentally, DX12, Vulkan, and Mantle are methods of compensating for weak single-threaded performance (or for spreading out a workload more evenly so it isn’t bottlenecked by a single thread).

https://www.extremetech.com/gaming/246377-new-directx-11-vs-directx-12-comparison-shows-uneven-results-limited-improvements

Also read that jdwii wrote at the beginning of page 26.

This quote does not answer the question.

The quote insinuates a hardware fault/design fault is the root issue.

The reality is much deeper than that. Nvidia DX12 driver show virtually no performance gain on Intel versus their DX11 drivers. They show a regression with AMD versus their DX11 drivers. There is a consistency there from the DX11 drivers that show a gap roughly equal to IPC gap + clockspeed gap in average frame rates.

So, if Nvidia drivers for DX12 show slightly exaggerated gaps from AMD CPUs, and no real gains for Intel...how can you postulate that a hardware design fault is the issue? Especially when DX12 drivers show regression, even for Intel CPUs?

https://www.pcper.com/category/tags/dx12

They also compare the two cards in DX11 and DX12 mode, with both cards using a Skylake-based Core i5 CPU. In this test, AMD's card noticed a nice increase in frame rate when switching to DirectX 12, while NVIDIA had a performance regression in the new API.

Now, let us consider that AMD GPUs show performance increase in DX12 of roughly similar levels on both architectures.

So, this is our dataset:

GPU A:

DX11 exhibits ~10-15% gap between CPU uarch A and B
DX12 exhibits ~12-18% gap between CPU uarch A and B
DX12 exhibits performance regression at times on both CPU uarchs A and B

GPU B:

DX11 exhibits ~6-11% gap between CPU uarch A and B
DX12 exhibits ~(-3)-8% gap between CPU uarch A and B
DX12 exhibits gains all the time on both CPU uarch A and B

What realistic conclusion can you draw when we divorce brands from this and objectively view the data set?

 

[gamerk316]

I again point out: When CPU performance is not an issue, multi-threading is going to be some percent slower due to the additional overhead involved.

Everyone is forgetting the primary advantage of DX12 is alleviating CPU bottlenecks on CPUs that can not handle a single large rendering thread. Assuming a perfectly optimized DX11 code path, I'd expect DX12 to see performance gain on lower tier CPUs [Pentiums, i3's, and FX CPUs come to mind], and a slight performance loss on more powerful CPUs [most i5's/i7's]. Which coincidentally, is EXACTLY what we see with NVIDIA.

 

[salgado18]

Could it be possible for AMD to disable one CCX on idle, to save on power draw? I just saw that R5's idle power consumption is just as high as R7, and a lot higher than i7's. And you would still get 3c/6t on idle clocks.

 

[James Mason]

Could it be possible for AMD to disable one CCX on idle, to save on power draw? I just saw that R5's idle power consumption is just as high as R7, and a lot higher than i7's. And you would still get 3c/6t on idle clocks.

It's extremely hard to actually measure that stuff accurately, and AMD even reports "goals" for CPU wattage rather than accurate wattage.

 

[-Fran-]

Too bad Toms didn't have the 1500X numbers in the mix, but other review sites do have them.

There seems to be a *real* advantage to having 6C over 4C, since in everything where the 1500X is close to the i5, the 1600X makes it onto par or better in some games that actually go wide (Crysis 3 being really surprising) and for productivity stuff, the 6C is a steal; hands down.

I thought the 1500X was going to be the shiny new recommended star from AMD's lineup, but the 1600X seems to be making its case *very* strongly. And like it was mentioned in Toms review, makes the 1700 a tougher sell (ironically enough) and the i7 6800K (E line) completely irrelevant; although that was expected, kind of.

And making myself sound like a broken record, I am really enthusiastic about the APUs, haha. Hype train about to leave the station! All aboard! XD

Cheers!

EDIT: Fixed idea.

 

[XBloodyR]

I dont even understand why you would get the 1600x over 1600 since the x costs more in of itself and additionally the 1600 comes with the cooler which is enough for overclocks to about 3.8-4 Ghz thats for any of the x versions. Only reason would be extreme overclocks.