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 Radeon VII
- Thread starter eksu
- Start date
- Sort by reaction score
A lot of science doesn't need FP64 - good FP32 and super fast FP16 covers most ML (machine learning) and AI (artificial intelligence) and DS (data science) tasks.
[doublepost=1547757415][/doublepost]
The gaming cards, however, are pathetic at FP64. A $6K Quadro RTX 6000 card is 32x faster at FP32 than FP64. (A Quadro RTX, of course, isn't a gaming card. It's an FP32 science card. All of the Turing GPUs are 1:32 for FP64:FP32 however.)
Last edited:
The point is that it is ugly to waste these chips, when they are more appropriate for Radeon Pro.
Nah...
People doing scientific work use pretty much exclusively CUDA ruling out AMD GPUs. FP32 vs FP64 not necessarily an issue (neither NNs nor genetic algorithms require FP64 for example)
Sad but true...
Last edited:
Right, both AMD and NVIDIA charge a premium for good FP64 performance, because it requires a lot of space in the chip to get that performance. If you have a design where you can turn off those units when there's a defect, then you have two options when you find such a defect:
1) You can throw that GPU in the trash, and lose 100% of the money.
2) Sell it as a different gaming-focused product and make some money from it.
Seems pretty reasonable that AMD is selling these chips as the Radeon VII, which is supposed to be a consumer/gaming focused product, instead of literally throwing them in the trash because they can't sell them as MI50/MI60.
[doublepost=1547774899][/doublepost]
Yeah I didn't want to mention that as folks would come out of the woodwork and accuse me of being a shill or something. It's pretty hard to beat the performance of the tensor cores in recent NVIDIA GPUs for deep learning stuff.
If the chips are faulty, disabling functionality is fine. If they are just crippled, it is an aberration.
I wouldn't expect Radeon 7 that much because AMD is preparing Navi architecture which is a whole new architecture. Until Radeon 7, they've been using the same architecture for 7 years. I'm talking about GCN architecture including Vega and Polaris. Navi architecture will bring massive improvements. The real question is will Mac Pro 2019 use Navi architecture or not? It seems Navi will launch in late 2019 so it may not be possible to use Navi GPU for Mac Pro.
If we actually get the new Mac Pro in 2019, I would expect it would be 7nm Vega. Since Navi is more of Polaris replacement, more likely they will be featured in iMacs and Macbook Pros.
There is no concrete info on whether Navi will be a brand new architecture or another improvement of GCN at this point so I wouldn't declare anything at this point.
Tell me, how better Geometry pipeline will improve compute, and professional workloads?
Navi is Gaming architecture. Compute oriented, Vega 20 replacement with Navi Architecture is coming Mid 2020, made on 7 nm EUV.
Indeed. What's more is the fact that CUDA is really usable. Not necessarily easy to integrate depending on IDE, but CUDA came a long way.
That said, OpenCL or ROCm are nowhere near in that regard.
For now, there's no AMD equivalent that could match with CUDA.
HPC at its current state is quite a mess. I'd love to see AMD come up with a good solution running on all major platforms. That would be awesome ...
Last edited:
I for one really couldn't care if it was team green or team red in my system - I just want to be able to buy a card that, if necessary, costs as much as a 2080ti, has the same power envelope as a 2080ti, AND has the same ability to drive a realtime (immersive) 3d environment at the same resolutions and object density as the 2080ti. I don't really give a toss about compute performance, because I can always wait for a another few minutes for a filter to apply, or a render to occur, but a lower quality realtime 3d workspace is a lower quality experience for every single second, minute and day of my working process.
I wish AMD would stop it with this exclusively "cheaper midrange (compute) card" strategy, and actually make a high end 3d card.
Everything that requires translating code from CUDA, which is Vendor lock-in, to any other, open source platform will be hard to implement, because of the work you have to put in.
Im am, and always were baffled, that people are not able to see this sole reason.
If you are locked to specific vendor, don't expect that everything will be easy to use, when the software has been developed for years. OpenCL was released years ago, and never took off, for this very reason. In psychology this factor is called 'Resistance to Change".
You can achieve the same things on AMD platform, as you can with CUDA. You just have to put the same amount of money in software development, you had put into software development for CUDA.
If you are not, don't complain that AMD is uncompetitve, then their hardware is competitive, with properly optimised software, outside Tensor Core usage(but FP16 is better, actually for MI50 and MI60 than it is for V100), because ALUs, are just ALUs. What matters here is the throughput of those ALUs, and AMD always was competitive, or had advantage(thats why in pure compute performance only lately Nvidia tied with AMD GPUs) on this front.
P.S. The only way we can get something good, is that CUDA equivalent, but Open Source, that IBM started working lately, on.
Not true. The boilerplate code required is just insane. OpenCL is nowhere as near as efficient in development compared to CUDA.
But that is not my main point. Problem is market fragmentation. Let me explain:
Using nVidia cards and CUDA: u can develop your applications on a Mac and run after compiling on Linux or Windows as well.
Using OpenCL you have to make sacrifices, since OpenCL is only supported up to version 1.2 on the Mac, basically meaning: no dynamic programming. Also, OpenCL is marked deprecated by Apple.
Similarly, nVidia cards only support OpenCL up to version 1.1. No option therefore.
Metal: Mac only. No option therefore.
ROCm: Linux only.
Since most applications run under Linux in some cloud there is another pitfall: the lack of AMD cards in cloud computing. Only Baidu cloud offers a selection of AMD cards. The big three - AWS, GCP and MS Azure currently don't.
This is why I called HPC a mess in some earlier post: there is no platform agnostic solution. CUDA, however, despite being proprietary, is not only the most efficient in therms of development effort required, it also IS the industry standard in GPGPU cloud computing as well as the only framework at least in some sense platform agnostic. Apple's resistance on Mojave drivers therefore contributes even more to described fragmentation.
I am not opposed to AMD cards. Quite the opposite, in fact I'd like to see a real competitor to CUDA. Alas, for now such a thing is nowhere in sight.
Unfortunate, but that's how the current state of affairs is
Last edited:
Navi is not just a gaming architecture lol. AMD can make different GPU with one architecture. How come Vega architecture used for Radeon RX, Pro, and instinct series? Polaris architecture also used for three series. I guess you dont know anything about AMD GPU since they use a single architecture to create 3 different series.
Radeon 7 is repurposed HPC chip... Just because it has ROPs, Geometry Engines, and execution units does not mean it is the same type gaming GPU architecture, as is Navi.
Navi will have two features which are not apparent in previous generations of GPUs. SUPER-SIMD architecture, which allows for higher ALU throughput in situations where ALUs are not utilized fully(Gaming, is the sole definition here). Everything related to Compute fully utilizes ALUs. It may have an effect on compute, and it may also not affect compute at all, depends if AMD has changed the ALU/CU structure. If they have switched from 64 ALUs/CU to 32 ALUs/ CU - the effect will be profound, because those 32 ALU CU will have the same performance as 64 ALU CU, without Super-SIMD architecture. Much lower power consumption, possibly higher clocks, which is designed to affect mostly...
... second Navi feature, which is... Primitive Shaders. Culling the Geometry, done on Hardware level, and not software level, like it was with Vega. It will not have any affect at all on comute performance. Simple as that.
Navi compute GPU, replacement for Vega 20 will come with Navi 20, mid 2020. Navi is gaming GPU architecture - first. Remember that.
Dude, you are the one who is not understanding about AMD GPU. Navi architecture can be used for Radeon Pro and Instinct like Polaris architecture. Can you explain why AMD made Radeon Pro 560 for MacBook Pro? What about iMac with Radeon Pro 580? They were gaming architecture and yet they were customized to workstation GPU.
Radeon Instinct WAS a workstation GPU but now, they gonna release a gaming GPU with Radeon Instinct.
There is no proof that Navi will be a gaming GPU only.
OpenCL is dead, killed by Apple.
How can you fault Nvidia for not supporting the latest version of a dead standard?
It is not dead. What the Apple toy company does now is irrelevant.
It is dead. The benefit of OpenCL on Mac was being able to use the CPU/GPU and being cross platform. Apple propped up OpenCL on their platform. But, now they are only pushing Metal. For cross platform developers why would you target OpenCL now. It's wasted money on development since the api is basically deprecated on Mac and more of a headache to support in the long run. And if you are just developing on Linux – you'd be a fool not to just go with Nvidia.
These discussions veer off into pointless arguments. Nvidia hasn’t released a supported macOS product or final non-beta driver for years so until that happens it’s futile circular arguments and dead boring.
We can only talk about what is supported and does work properly. We can let OpenGL/CL/CUDA go and need to focus on Metal performance on AMD and ARM chips.
We are at the end of the road and starting a new path.
Next year all apps currently using CUDA, GL or CL on macOS will have full Metal support. 2020 is when we can FINALLY. start to focus on the future of Metal performance on macOS. The new Mac Pro will probably be released with 10.15 ready for that new era.
We can only talk about what is supported and does work properly. We can let OpenGL/CL/CUDA go and need to focus on Metal performance on AMD and ARM chips.
We are at the end of the road and starting a new path.
Next year all apps currently using CUDA, GL or CL on macOS will have full Metal support. 2020 is when we can FINALLY. start to focus on the future of Metal performance on macOS. The new Mac Pro will probably be released with 10.15 ready for that new era.
Perhaps you are correct. However, as a developer I have no other choice than to ignore who's fault it is since there is nothing I could do about it.
All we can do is draw conclusions, pick what is best for our immediate purposes and continue from there
Well, its kind of dead.
Apple deprecated it on its platform.
nVidia does not wholehartedly support it; it clearly favours CUDA.
AMD - assumingly because of the above - apparently chose to go ahead anyway and came up with ROCm. I guess not a bad move, albeit ROCm is not relevant at present. It might in the future, but not for now
[doublepost=1547848949][/doublepost]
I couldn't disagree more. While this may be true for users of applications running on mentioned platforms, the said ignores the plethora of developers using Macs to develop cloud or AI applications.
If Apple continues to block nVidia drivers it forces pretty much ANY AI developer or scientist to use competing platforms. Metal is literally irrelevant in this rapidly developing fields.
ML, cNNs/rNNs, TensorFlow... all blocked along with nVidia drivers. Bad, bad choice
Last edited:
Officially dead. After MacOS Mojave, they stop supporting OpenCL and OpenGL.
This is valid for Apple platforms only.
Since OpenCL is a platform independent standard, the fact that niche manufacturer Apple deprecated it is not that important.
“Blocked drivers” comes back to the circular argument. If Nvidia won’t releases a product that says ‘macOS supported’ on the retail box and won’t supply final non beta drivers then Apple has every right to not approve them.
The big issue here is cost. I’m speaking as an Nvidia share holder who has owned many Nvidia cards and used them with macOS. See my posting history for the last 4 years and you will see all the cards I owned, tested and all the bugs I reported.
When those bugs appeared because of Nvidia’s beta drivers users didn’t go to Nvidia for help. They always initially assume the apps are buggy, whether it is Apple or Adobe or other apps. Users then go to Apple or Adobe asking for tech support. That has cost Apple and Adobe a lot of money in wasted time and effort.
Just google and you will see these frustrated users on Apple and Adobe message boards (and here too). Some of them aren’t even using Macs. They have a Hackintosh and are wasting Apple’s time and money with support messages.
Add the cost up. It’s a lot more than you think. Apple is happy to carry the costs of problems related to AMD drivers but then when you add Nvidia bugs on top of that it’s a major annoyance.
That’s why Nvidia has to take responsibility, state which OS is supported on the retail boxes, finalize the drivers, and offer proper tech support.
I don’t want to discuss this any further. I don’t expect everyone to take off their blinders and see what is obviously a cost problem here.