Asgorath, I appreciate your insight here. Any idea if Metal will bring
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
AMD Confirms RX 480 At $199 USD, Other APU & Polaris Announcements
- Thread starter lowendlinux
- Start date
- Sort by reaction score
Asgorath, I appreciate your insight here. Any idea if Metal will bring
DirectX 12 was also developed for years before Mantle. And you know what is funniest? Microsoft included Mantle, as feature set base in DX12 in the process of developing it.
Vulkan and LiquidVR are to Mantle like Michael Jackson Black vs White - same guy, with same personality, just different surface.
DX12 and Metal to Mantle are like Disco Volante vs Alfa Romeo 8C. Same skeleton, same chassis, same engine but ride is softer, different tune of exhaust, different surface. And different feeling. Makes sense?
And I agree with you. Definitely Metal is the easiest to work with, I think easier than even DX11, despite the years of development, education and experience with that API for many developers.
Nobody has disproven me anything.
Last edited:
You keep saying no one as disproved you because you just ignore what many others have been telling you.
DirectX 12 isn't based on Mantle and neither is Metal. I know that you wish it with all your heart, but you are wrong and have been proved wrong before in similar discussion.
Whether Metal is easy to develop for or not, at the end of the day the big name developers already struggle to make deadlines for the three main platforms.
And if the Mac has mobile GPUs they don't have incentive to develop for that. Customers will just cry like grown up babies that the Mac games don't perform like they do in Windows. And we know for sure Windows will always have better drivers with full feature sets too.
So Metal games will be mostly iOS ports and some stuff from companies like Feral. Nothing major will change when it comes to AAA titles.
I'm interested not in Mac gaming but what will Metal do for pro apps. Nothing materialised there after big hype more than a year ago. I think you still won't see much for about one more year.
And if the Mac has mobile GPUs they don't have incentive to develop for that. Customers will just cry like grown up babies that the Mac games don't perform like they do in Windows. And we know for sure Windows will always have better drivers with full feature sets too.
So Metal games will be mostly iOS ports and some stuff from companies like Feral. Nothing major will change when it comes to AAA titles.
I'm interested not in Mac gaming but what will Metal do for pro apps. Nothing materialised there after big hype more than a year ago. I think you still won't see much for about one more year.
Unfortunately, Metal is still missing some key features when compared with DX11, which is now ~7 years old. So, it makes it really hard for folks who want to port DX11 games from Windows to macOS, because there's stuff that those games are doing that are either impossible to do or really difficult to emulate with Metal. This is confirmed by the fact there are basically no actual games that use Metal, aside from perhaps the new World of Warcraft expansion (Legion). UE4 has a Metal back end, so perhaps when games that use UE4 get released we'll see better Mac versions.
I'm happy to just agree to disagree on this whole Metal vs Mantle thing. Are they similar in some regards? Sure. Does that mean Metal was based on Mantle? I haven't seen any evidence to suggest so. I think Apple likes being in control over their platform, and I think AMD is struggling and might overstate the relevance or importance of Mantle in the grand scheme of things (even Vulkan ended up being pretty different to Mantle, once Khronos started working on it). Did AMD succeed in pushing the other next-gen APIs to be lower level? I would say so. Is that actually a good thing? Based on the number of DX12 titles that exist, especially those that perform better than their DX11 versions, I kind of doubt it. We kind of learned all of this back in the 90s with 3dfx Glide and other low-level APIs (i.e. they're really hard to work with, not future proof, etc).
hay found a some benchmarks of the rx 480 v titan x in DaVinci resolve 12.5 looks like the rx 480 8GB is a good value card.
(+ more cards later on looks like it works as well as or better than the GTX 1070 but it's early days so things may change with driver updates or something)
https://forum.blackmagicdesign.com/viewtopic.php?f=21&t=49713
(also seen reports of resolve 12.5 using much more vram giving people with 2GB cards some problems.)
They seem happy to have titan x performance at such a price
(+ more cards later on looks like it works as well as or better than the GTX 1070 but it's early days so things may change with driver updates or something)
https://forum.blackmagicdesign.com/viewtopic.php?f=21&t=49713
(also seen reports of resolve 12.5 using much more vram giving people with 2GB cards some problems.)
They seem happy to have titan x performance at such a price
Last edited:
Compute performance of both cards is similar. 6.1 TFLOPs vs 5.8 TFLOPs. If your software is properly coded that is what you will see. Similar performance on GPUs that have similar compute performance.
FirePros at Siggraph on Monday, and FireRays:
http://radeon.com/Siggraph/
http://www.amd.com/en-us/press-rele...sionGUID=5435999b-60cd-b47e-38c3-182ee5b3d9cc
Watch the videos...
[doublepost=1469200909][/doublepost]Holly crap...
http://wccftech.com/nvidia-geforce-gtx-titan-x-pascal-unleashed/
http://radeon.com/Siggraph/
http://www.amd.com/en-us/press-rele...sionGUID=5435999b-60cd-b47e-38c3-182ee5b3d9cc
Watch the videos...
[doublepost=1469200909][/doublepost]Holly crap...
http://wccftech.com/nvidia-geforce-gtx-titan-x-pascal-unleashed/
Last edited:
You keep listing all these raw theoretical maximum performance numbers. If you compare the 480 vs the 1060, the 480 wins in every metric of theoretical performance, often by a considerable margin. However, in real games and applications, the 1060 is usually 10% or more ahead.
So, as usual, take these theoretical maximum numbers with a huge grain of salt. If ever application ran at the theoretical maximum the 480 would wipe the floor with the 1060, but in real life it's the reverse.
It appears that your reading comprehension from context is not working
.---------
Here are the Standard Candle results for the new $199 AMD Radeon RX480 in RED, Nvidia Titan X in GREEN.
Rsolve 12.5 Studio, X99 Motherboard, 12 core Xeon.
Blur:
09 Nodes: RX480:24fpsTitanX:24fps
18 Nodes: RX480:20fps TitanX:16fps
30 Nodes: RX480:13fps TitanX:11fps
66 Nodes: RX480:6fps TitanX:5-6fps
TNR:
1: RX480:24fps TitanX:24fps
2: RX480:18fpsTitanX:17-20fps
4: RX480:9fpsTitanX:11fps
6: RX480:7fpsTitanX:8fps
Big thanks to Peter Richards of fixafilm.com for purchasing the 480 card, as there are no compute based reviews of the card, he thought it important that the card be tested for OPENCL performance. Let me know if you want any other benchmarks run.
--------
This is the post that has been linked in the post that I have responded.
Secondly, what if in DX11 games AMD GPUs are bottlenecked by nature of API, and by CPU overhead? Have you considered this? What happens when you lift CPU overhead and GPUs are not bottlenecked by nature of API like is in DX12 and Vulkan?
And again, What I have posted: Properly coded Application that is not bottlenecking in any way any hardware will show exactly that: 6 TFLOPs GPU will be slightly faster than 5.8 TFLOPs GPU. 8.6 TFLOPs GPU will be MUCH faster than 5.8 TFLOPS GPU.
Whole point, even if you wanted to contradict it, has been proven correct.
It appears that your reading comprehension from context is not working
---------
Here are the Standard Candle results for the new $199 AMD Radeon RX480 in RED, Nvidia Titan X in GREEN.
Rsolve 12.5 Studio, X99 Motherboard, 12 core Xeon.
Blur:
09 Nodes: RX480:24fpsTitanX:24fps
18 Nodes: RX480:20fps TitanX:16fps
30 Nodes: RX480:13fps TitanX:11fps
66 Nodes: RX480:6fps TitanX:5-6fps
TNR:
1: RX480:24fps TitanX:24fps
2: RX480:18fpsTitanX:17-20fps
4: RX480:9fpsTitanX:11fps
6: RX480:7fpsTitanX:8fps
Big thanks to Peter Richards of fixafilm.com for purchasing the 480 card, as there are no compute based reviews of the card, he thought it important that the card be tested for OPENCL performance. Let me know if you want any other benchmarks run.
Its easy to pick out compute tests that favor one vendor over another. Another OpenCL test shows the 980 ti with much better performance over the rx 480.
Could be but don't expect crazy performance breakthroughs. Comparing Directx11 vs 12 on the RX 480 At 1440p, Arstechnica shows that AMD sees a 15% increases in Ashes of the singularity, 3% increase in Hitman and 11% decrease in Rise of the tomb raider.
Additionally with the release of Pascal, Nvidia also has implemented their own limited form of asynch compute meaning that AMD is no longer the sole beneficiary of async compute performance increases from these new APIs.
Nvidia is using Preemption not true hardware level Asynchronous Compute. Games are using Asynchronous Shaders.
Hotel in Luxmark benchmark aside compute performance of GPU, is relying heavily on Rasterising - something which AMD lacks in terms of performance. Unfortunately.
Again, completely false.
http://www.anandtech.com/show/10325/the-nvidia-geforce-gtx-1080-and-1070-founders-edition-review/9
AnandTech explains in great detail how Maxwell and Pascal implement asynchronous compute, and the improvements that are in the Pascal version. Preemption is something completely different and has nothing to do with this.
Regardless of whether its "real" async compute, benchmarks that test this show that Pascal sees increases in performance where maxwell sees no increase in performance.
[doublepost=1469213924][/doublepost]
Currently it is not. There is some limited driver support in MacOS Sierra, but as far as I know it either does not fully work or is unstable.
You are absolutely right. Pascal and Maxwell both has Async Compute, both can execute compute workload parallel to graphics workload, without preemption. Maxwell was just very hard to tune to get perf increase
Right, the dynamic scheduling part that was introduced in Pascal is very important for the apps that are currently written to use async compute. I guess it ended up being just too hard to tune for the static scheduling that Maxwell provided.
Zen-Vega will come to nMP.
Polaris to MBP and iMac. Apple will have seen mobile Pascal by now but the evidence is in AMD's favour ATM.
Polaris to MBP and iMac. Apple will have seen mobile Pascal by now but the evidence is in AMD's favour ATM.
The power is no issue, yes the card draws more than spec on the PCIe slot (in most cases) however this lane seems to be shared across all 4 slots at 300W with no slot specific limit/measurement on my 4.1.
My RX480 runs fine on a single 6 pin + PCIe in Windows, with both old and new driver (did *not* test the power saving setting that clocks it down even more).
Considering i stole power from the bay and HDD drives to supply 2 RX270 before just fine a dual RX480 setup *should* work if your PCIe draw otherwise is low (like the usual PCIe SSD and a USB 3.0 card) without any mods.
It is still an unknown at this moment. The manual says total cannot exceed 300W, however, 75W per slot is the normal standard. It's had to tell if Apple use poor wording (because they never says you can install any non-Apple approved PCIe card in there), or really a single slot can deliver up to 300W (in no other card installed).
TBH, it hard to believe that any of those slot can deliver 4x normal standard.
Yes, 300W on a single slot is alone electrically dangerous - however, the RX480 max measured slot consumption is 90-100W, which is not a problem on the pins or the PCB tracks yet. A single one seems to work just fine, without having it set with the new driver to hotfix either.
Best guess for limit would be the Apple approved/official GPUs (The 5770 + 3x Nvidia config is IIRC officially supported) and the RAID adapter, but i have doubts either draw 75W+.
@koyoot if you look at post further down on the black magic forum the GTX1070 has about the same performance as the RX480 (in one app so only matters if your thinking of video grading)
"Using the Blur settings on the original candle test on the reference GTX 1070
9 Nodes - 24 (we managed to go up to 16 with it at 24 and on the 17th node it dropped to 21)
18 Nodes - 19-20
30 Nodes - 12-13
66 Nodes - 6"
but i think this is in openCL mode, dont think the CUDA side has been fully updated but not shore so i may be wrong.
the cards are all new so i assume there will be some speedup as things get updated.
& a lot of this relay comes down to what do you (or 'I') want/need, not if one card is better.
if i need CUDA then ATI is a no go but if i want a cheep openCL card then ATI is worth looking at but it's still early days, + the RX 480 is a cheaper card than most which needs to be mentioned.
(and at the mo there both as much use to me as a toaster for video editing as there's still no osx drivers)
ps thats a retro joke https://en.wikipedia.org/wiki/Video_Toaster
"Using the Blur settings on the original candle test on the reference GTX 1070
9 Nodes - 24 (we managed to go up to 16 with it at 24 and on the 17th node it dropped to 21)
18 Nodes - 19-20
30 Nodes - 12-13
66 Nodes - 6"
but i think this is in openCL mode, dont think the CUDA side has been fully updated but not shore so i may be wrong.
the cards are all new so i assume there will be some speedup as things get updated.
& a lot of this relay comes down to what do you (or 'I') want/need, not if one card is better.
if i need CUDA then ATI is a no go but if i want a cheep openCL card then ATI is worth looking at but it's still early days, + the RX 480 is a cheaper card than most which needs to be mentioned.
(and at the mo there both as much use to me as a toaster for video editing as there's still no osx drivers)
ps thats a retro joke https://en.wikipedia.org/wiki/Video_Toaster
Well it should be similar because it also has similar compute performance to both RX 480 and Titan X.
What I am posting is all about optimising the software to fully utilize the compute performance of GPUs, not how they are better than others. And properly optimised software always will show that 6 TFLOPs GPU is on par with 6 TFLOPs GPU. They are mathematical algorithms. And then you should pick hardware by your pocket, not brand favourism. But that is "ideal" world.
What I am posting is all about optimising the software to fully utilize the compute performance of GPUs, not how they are better than others. And properly optimised software always will show that 6 TFLOPs GPU is on par with 6 TFLOPs GPU. They are mathematical algorithms. And then you should pick hardware by your pocket, not brand favourism. But that is "ideal" world.