CUDA vs OpenCL needs a reality check...

[BallaTheFeared]

Redemption.

 

[Cerb]

And yet nvidia dominates the pro market - is that because 80% of pro buyers are serious nvidia fanboys?

Anway that's cost for the customer, not the dev. The dev saves money by developing on CUDA because it's easier due to better language support, features, libraries and debug tools. It costs him more to do OpenCL.

1. Most users really don't care. Dell or HP generally make the nV v. AMD decision, except in cases where CUDA is specifically called for. NV has the market share they do, outside of pure compute, due to being able to make good deals. Obviously, AMD couldn't do the same.

2. nV had stable driver OpenCL support before AMD did, so really, the entire framing of the discussion is BS, and has been from the start. As of now, all 3 companies you might use, in the land of x86, have some level of OpenCL support, with Intel's being the weakest (not bad, just Intel being Intel, and tying feature and performance improvements directly to hardware updates).

OpenCL, flaws and all, has gotten many companies willing to support mostly the same thing--Intel, AMD, nV, PowerVR, and others. Great for users. Great for companies like Adobe. Fine either way for nV or AMD. The dev makes money, by having a product that allows increased productivity compared to their previous version, this time getting fair performance benefits on a wider variety of hardware and software platforms. The user thus gets something tangible for their upgrade costs, and remains more likely to stick with said company gladly, rather than grudgingly (upgrading only when they have to, seeking alternatives, etc.).

The only people that left AMD out in the cold, wrt to OpenCL, was AMD, who left it in beta too long, and clearly never got enough people writing software to make their hot air reality, despite all of their hardware expertise.

 

[NTMBK]

1. Most users really don't care. Dell or HP generally make the nV v. AMD decision, except in cases where CUDA is specifically called for. NV has the market share they do, outside of pure compute, due to being able to make good deals. Obviously, AMD couldn't do the same.

2. nV had stable driver OpenCL support before AMD did, so really, the entire framing of the discussion is BS, and has been from the start. As of now, all 3 companies you might use, in the land of x86, have some level of OpenCL support, with Intel's being the weakest (not bad, just Intel being Intel, and tying feature and performance improvements directly to hardware updates).

OpenCL, flaws and all, has gotten many companies willing to support mostly the same thing--Intel, AMD, nV, PowerVR, and others. Great for users. Great for companies like Adobe. Fine either way for nV or AMD. The dev makes money, by having a product that allows increased productivity compared to their previous version, this time getting fair performance benefits on a wider variety of hardware and software platforms. The user thus gets something tangible for their upgrade costs, and remains more likely to stick with said company gladly, rather than grudgingly (upgrading only when they have to, seeking alternatives, etc.).

The only people that left AMD out in the cold, wrt to OpenCL, was AMD, who left it in beta too long, and clearly never got enough people writing software to make their hot air reality, despite all of their hardware expertise.

AMD has OpenCL 1.2 support. Hell, even Intel now supports OpenCL 1.2. Remind me, what version do NVidia support?

 

[notty22]

AMD has OpenCL 1.2 support. Hell, even Intel now supports OpenCL 1.2. Remind me, what version do NVidia support?

Intel Releases OpenCL 1.2 Driver and Tools Update for Ivy Bridge and Haswell
AnandTech ‎- 4 days ago

 

[Arkadrel]

nV had stable driver OpenCL support before AMD did, so really, the entire framing of the discussion is BS, and has been from the start.

So years ago... nvidia had a driver out first?
I guess how their drivers look now doesnt matter then right?

http://www.anandtech.com/show/6886/intel-updates-opencl-driver-and-tools-for-ivy-bridge-and-haswell

...Nvidia provides mature profiling tools for CUDA applications but their OpenCL tooling has been lackluster at best and Nvidia drivers remain at OpenCL 1.1.

The only people that left AMD out in the cold, wrt to OpenCL, was AMD, who left it in beta too long, and clearly never got enough people writing software to make their hot air reality, despite all of their hardware expertise.

Hater gonna Hate ^^
AMD didnt make OpenCL, afaik, nor are they responsible for the state its in.
Also isnt OpenCL getting more program supports now than cuda is? How is that a hot-air-reality?

Lets give it 2-3 years and see where cuda is by then.

 

[Keysplayr]

So years ago... nvidia had a driver out first?
I guess how their drivers look now doesnt matter then right?

http://www.anandtech.com/show/6886/intel-updates-opencl-driver-and-tools-for-ivy-bridge-and-haswell





Hater gonna Hate ^^
AMD didnt make OpenCL, afaik, nor are they responsible for the state its in.
Also isnt OpenCL getting more program supports now than cuda is? How is that a hot-air-reality?

Lets give it 2-3 years and see where cuda is by then.

You might want to start thinking about if you would be posting in such a way if the roles were reversed. I don't think you would be. Do you really care if CUDA obliterates OpenCL or if OpenCL obliterates CUDA? No, of course you don't. It's merely because CUDA belongs to Nvidia.

 

[railven]

You might want to start thinking about if you would be posting in such a way if the roles were reversed. I don't think you would be. Do you really care if CUDA obliterates OpenCL or if OpenCL obliterates CUDA? No, of course you don't. It's merely because CUDA belongs to Nvidia.

Keys, this statement is ironic. Don't you think?

 

[Keysplayr]

Keys, this statement is ironic. Don't you think?

Why? I haven't said anything about OpenCL or CUDA in a while. Lets call things what they are. Bout time.

 

[railven]

Why? I haven't said anything about OpenCL or CUDA in a while. Lets call things what they are. Bout time.

No, you just came into this thread challenging opinions of member's positions which is ironic considering yours.

And since I'm now doing the same, I'm bowing out. Just wanted to point out the irony in your statement.

 

[Keysplayr]

No, you just came into this thread challenging opinions of member's positions which is ironic considering yours.

And since I'm now doing the same, I'm bowing out. Just wanted to point out the irony in your statement.

There isn't any irony when I am transparent and forthcoming with my preferences and my favoritism. It's all out there on my sleeve for all to see. When others posts are bias driven, or appear to be, I like to make sure it's acknowledged. These people don't care about CUDA or OpenCL. They care who has what and it's dumb.

 

[NIGELG]

There isn't any irony when I am transparent and forthcoming with my preferences and my FANBOYISM. It's all out there on my sleeve for all to see. When others posts are bias driven, or appear to be, I like to make sure it's acknowledged. These people don't care about CUDA or OpenCL. They care who has what and it's dumb.

Fixed()

 

[Arkadrel]

You might want to start thinking about if you would be posting in such a way if the roles were reversed. I don't think you would be. Do you really care if CUDA obliterates OpenCL or if OpenCL obliterates CUDA? No, of course you don't. It's merely because CUDA belongs to Nvidia.

Its because its Proprietary.
Its because its slowing down another Language that isnt.

OpenCL benefits everyone.
Intel CPUs with iGPUs.
AMD's APU & GPUs
Nvidia's GPU (though not as much a 100% monopoly via cuda would)
ARM smartphones
ect.

 

[Will Robinson]

I'll enjoy watching you "acknowledge" this then...

...Nvidia provides mature profiling tools for CUDA applications but their OpenCL tooling has been lackluster at best and Nvidia drivers remain at OpenCL 1.1.

Ouch :whiste:

 

[Arkadrel]

You might want to start thinking about if you would be posting in such a way if the roles were reversed. I don't think you would be. Do you really care if CUDA obliterates OpenCL or if OpenCL obliterates CUDA? No, of course you don't. It's merely because CUDA belongs to Nvidia.

Did you notice that a Nvidia fanboy made the thread, in a AMD vs Nvidia way?

Did you call him out on it?

Did you read my reply? the 2nd post:

TS says: Soon we ll see! CUDA vs OpenCL.... bla bla bla CUDA best.
I reply:

I suspect it ll be just as fast.... however even if its slower, by vitue that Intel IGP, AMD APU/GPU's and Nvidia GPUs can use it, instead of only nvidia GPU's, it ll be a win in my book.

I dont use a nvidia card atm.... so no cuda for me.
Why is it weird that Id like to see more programs make use of OpenCL?



This was about Adobe Premiere Pro running 6x as fast by use of OpenCL (vs CPU),
as it did without (for render in/out times).
Any program that can see a that kinda increase in performance, should be a great thing.

 

[wand3r3r]

There isn't any irony when I am transparent and forthcoming with my preferences and my favoritism. It's all out there on my sleeve for all to see. When others posts are bias driven, or appear to be, I like to make sure it's acknowledged. These people don't care about CUDA or OpenCL. They care who has what and it's dumb.

Are you missing a little "biased towards AMD" or "biased against NV" in there because I didn't notice it, nor have I seen you call out the abundant NVolunteers ... ever.

But hey, it's great you are here to call out "biased" posts and all.

/OT

On topic, I don't think the OP realized that his thread title goes both ways.

 

[Lonyo]

There isn't any irony when I am transparent and forthcoming with my preferences and my favoritism. It's all out there on my sleeve for all to see. When others posts are bias driven, or appear to be, I like to make sure it's acknowledged. These people don't care about CUDA or OpenCL. They care who has what and it's dumb.

We pretty much all have Intel, but some "biased" people don't care about that, as can be seen from some replies in this thread.

 

[Keysplayr]

Are you missing a little "biased towards AMD" or "biased against NV" in there because I didn't notice it, nor have I seen you call out the abundant NVolunteers ... ever.

But hey, it's great you are here to call out "biased" posts and all.

/OT

On topic, I don't think the OP realized that his thread title goes both ways.

You're very welcome. My pleasure. :thumbsup:

 

[Cerb]

So years ago... nvidia had a driver out first?
I guess how their drivers look now doesnt matter then right?

It's working. It's been working. So, they need to get 1.2 support in there, sure. But, that's not going to stop any practical software adoption. If they don't bother at all, it will hurt them.

http://www.anandtech.com/show/6886/intel-updates-opencl-driver-and-tools-for-ivy-bridge-and-haswell

Exactly what I said: "Intel being Intel, and tying feature and performance improvements directly to hardware updates" Where does that leave Sandy Bridge users, FI? Why didn't they make HW that could be more flexible? nV and AMD have gotten support that applied to buyers of HW before the software support was there. Intel has the fastest CPUs, so they don't have to, but it's not exactly endearing.

AMD didnt make OpenCL, afaik, nor are they responsible for the state its in.

No, but they pushed it in talk only, for their own use, even, while others beat them to the punch.

Also isnt OpenCL getting more program supports now than cuda is? How is that a hot-air-reality?

Seen AMD's financial state, recently? That's how it's not reality. They had two basic choices: fatten themselves and let the company go, or take the reigns and innovate. They chose to try to look like they were doing the latter, hyping it up for their HW, while mostly doing the former.

Lets give it 2-3 years and see where cuda is by then.

HPC clusters, mostly. Will AMD be liquidating assets, by then? Or, maybe get the chance to "buy" a company that want to take them over? CUDA had to exist for nV's sake, long before OpenCL came about, and is still used for more advanced software on nV GPUs (their software and GPU tech move fast, while standards bodies don;t). OpenCL is what the masses will mostly make use of. There hasn't been an A versus B situation, except created by fanboys.

 

[MrrVlad]

I've been using CUDA for some time at work. Recently I had to investigate if next project should use OpenCL for easier portability and I found that even top of the line cards are not well supported - you can't use float atomics, warp shuffle, pointers to global memory in LDS, instruction reordering is not good resulting in same simple code executing up to 20% slower on NVidia hardware (I had not made a port to Mantle yet).

This would cause considerable changes to algorithms and would easily result in 2X perf difference.

So, since perf is important in production environment, at this point OpenCL is a no-go. I'll see where I can got to with Mantle.

 

[Jodell88]

I've been using CUDA for some time at work. Recently I had to investigate if next project should use OpenCL for easier portability and I found that even top of the line cards are not well supported - you can't use float atomics, warp shuffle, pointers to global memory in LDS, instruction reordering is not good resulting in same simple code executing up to 20% slower on NVidia hardware (I had not made a port to Mantle yet).

This would cause considerable changes to algorithms and would easily result in 2X perf difference.

So, since perf is important in production environment, at this point OpenCL is a no-go. I'll see where I can got to with Mantle.

Two things:-
Mantle is for graphics only AFAIK.
Nvidia limits OpenCL performance in their drivers.

 

[Atreidin]

Rise from your grave!

 

[stahlhart]

Closing necro thread.
-- stahlhart