CUDA vs OpenCL needs a reality check...

[NTMBK]

With his post history unfortunately it might not far fetched. He's either the most deranged fan of NV, or ... well I guess I shouldn't go there.

However, if you read my post, I said not shills opinions i.e. NV's blog.

(I did tell him to go back in his cave, which is pretty mild compared to what he keeps saying around here)

Ah, you were referring to the NV blog? My bad, I misread that.

 

[Eureka]

Because you have no bias at all, am I right?

Developer's blog is a fine place to look for examples. Just because it's their website doesn't mean they're making up random articles out of nowhere. If you go to AMD's site, you'll get AMD articles mostly, but it doesn't mean those are fake articles either.

In fact, I'd say it's the best place to look if you want good examples, because they want to put their best foot forward. We'd like to see what the environment is fully capable of, not examples of bad code. Unless you're just trying to find things to slander.

 

[wand3r3r]

Because you have no bias at all, am I right?

Developer's blog is a fine place to look for examples. Just because it's their website doesn't mean they're making up random articles out of nowhere. If you go to AMD's site, you'll get AMD articles mostly, but it doesn't mean those are fake articles either.

In fact, I'd say it's the best place to look if you want good examples, because they want to put their best foot forward. We'd like to see what the environment is fully capable of, not examples of bad code. Unless you're just trying to find things to slander.

As for the bias, I try not to be biased. I buy both NV and AMD products and I switch when there is compelling reason to do so ($/performance being my primary concern).

You have valid points, but I am not very interested in marketing materials especially in this particular topic.

I am curious about ATers because there are 300k+ members (I believe) so there are certainly people who do parallel computing, which this thread is about to see their opinions on the topic. There are a few biased people in this thread trying to enforce their notion about how great CUDA is and how bad OpenCL is but they keep pitching their opinion on everything their fav. company is the same way.

I accept CUDA is better, it's logical that if it's well thought out and easy to begin along with great documentation then it probably is nice to work with. On the other hand, the big opinions probably haven't tried OpenCL. I'd like to hear more about it from people with experience, out of curiousity, and not from the people that praise all things from their favorite company.

 

[Eureka]

I am curious about ATers because there are 300k+ members (I believe) so there are certainly people who do parallel computing, which this thread is about to see their opinions on the topic. There are a few biased people in this thread trying to enforce their notion about how great CUDA is and how bad OpenCL is but they keep pitching their opinion on everything their fav. company is the same way.

I accept CUDA is better, it's logical that if it's well thought out and easy to begin along with great documentation then it probably is nice to work with. On the other hand, the big opinions probably haven't tried OpenCL. I'd like to hear more about it from people with experience, out of curiousity, and not from the people that praise all things from their favorite company.

Unfortunately, there are very few developers who are truly "unbiased". In industry, the first thing you learn is generally the last tool you'll ever use. You're probably going to see far more CUDA arguments just because of Nvidia's presence in the industry. Which, based off of presence alone, may make it the better industrial tool (we engineers give less thought to "openness" or "up and coming", but rather better support).

The best way, to extrapolate what is actually better, is to do your own comparison, using the best examples given, and unfortunately, do it yourself. Still, whenever I am looking into a new tool, I'd prefer to have the corporate showroom. They know their tool best, and it can tell me what I can actually get out of it.

 

[parvadomus]

OpenCL will kill CUDA @ mainstream easy, taking with it physx and that kind of garbage.
For HPC it will take a while, but Intel may be more than a threat.

 

[Genx87]

I think it musing AMD is somehow associated with OpenCL. A standard AMD has done their best not to promote in an real sense. If OpenCL becomes the defacto standard Nvidia will run it just fine. Remember, when AMD was tooting their horn about openCL nvidia was delivering a working driver that didnt require an SDK to function.

 

[ams23]

Its not apples to apples, its AMD card (openCL) vs Nvidia (cuda). But it looks like in this case, OpenCL is doing better than CUDA is.

Actually this AMD blog really doesn't give us much insight about differences in performance and implementation between OpenCL vs. CUDA on their own, but rather shows differences in performance between select AMD and NVIDIA GPU's (using pre-release software too). And the marketing graphs from AMD start with a scale less than zero, so at first glance to a casual observer, a 10-20% difference in performance looks like a 2-3x difference in performance! It is deplorable when companies (such as AMD, NVIDIA, or anyone else) do this, especially when the performance differences are fairly small all things considered. As for the GPU utilization graph, that is incomplete without showing CPU utilization too.

 

[Arkadrel]

@Genx87

Ofc Nvidia will run it fine too. Probably not as great as AMD does, because they themselfs optimise their GPU designs for CUDA (rather than opencl), but good enough.

No one loses from OpenCL comeing out on top.

Intel supports it.
ARM supports it.
AMD supports it.
Nvidia supports it.
...
ect
..
.

^ if you have any product from these guys in the future, OpenCL will work for you.
There in lies its strength(s), its good for software guys that make money selling something to many people.

CUDAs strength are in software not sold to many. Uniquely made software, made to run 1 type of research on some super computer.

While it is nice that the AMD cards perform well with this OpenCL implementation, AMD is misleading people with this blog for two reasons: 1) The comparisons they made show performance differences between select AMD and NVIDIA GPU's, rather than OpenCL vs. CUDA on it's own; 2) The marketing graphs start with a scale less than zero and at first glance to a casual observer make a 10-20% difference in performance look like a 2-3x difference in performance. It is deplorable when companies (such as AMD, NVIDIA, or anyone else) do this.

1) AMD cannot test OpenCL vs Cuda on same video card
(with their own cards, because they cant run CUDA).

Are you suggesting they publish OpenCL vs CUDA on nvidia cards? what if nvidia designs run OpenCL poorly by choice? because they prioritise CUDA? is that fair? What would nvidia say if AMD started makeing graphs like that? Say it shows Opencl beating CUDA, on nvidia cards and AMD posted it? what do you think Nvidia would say publicly? would they call out AMD for slandering their name? Discredit the chart/results, claiming it unfair marketing? (its a recipe for disaster for amd if they did)

So they did the next best thing.
They took Nvidia class card costing xxx$ and compaired it to AMD class card same price catagory.
They left the Nvidia card running CUDA, and had theirs running OpenCL.

2)

...at first glance to a casual observer make a 10-20% difference in performance look like a 2-3x difference in performance.

Everyone does that lol (nvidia,intel,amd,...). No one falls for it.

 

[Jaydip]

While it is nice that the AMD cards perform well with this OpenCL implementation, AMD is misleading people with this blog for two reasons: 1) The comparisons they made show performance differences between select AMD and NVIDIA GPU's, rather than OpenCL vs. CUDA on it's own; 2) The marketing graphs start with a scale less than zero and at first glance to a casual observer make a 10-20% difference in performance look like a 2-3x difference in performance. It is deplorable when companies (such as AMD, NVIDIA, or anyone else) do this.

Both of your points are valid but unfortunately point two is abused by both camps.Regarding first yes they compared W9000 (the top model) with the second highest K5000.K5000 replaces Quadro 5000 not 6000 which still remains NV's most potent workstation card.They are going to introduce the successor the Quadro 6000 based on GK110 pretty soon.

 

[sontin]

Both of your points are valid but unfortunately point two is abused by both camps.Regarding first yes they compared W9000 (the top model) with the second highest K5000.K5000 replaces Quadro 5000 not 6000 which still remains NV's most potent workstation card.They are going to introduce the successor the Quadro 6000 based on GK110 pretty soon.

The problem with AMD's numbers is that they comparing cards which are not in the same price level.
The W9000 cost $4999, the K5000 less than half with $2299. Cou can buy K20c cards if you need something to compute which cost less than <$4000.

What AMD did was nothing else than a canned comparison between their and nVidia cards.

And looking at these numbers:
K5000 has the better perf/$ and perf/watt than the W9000. So if you need a Workstation card for Adobe K5000 is the much better deal...

 

[Jaydip]

The problem with AMD's numbers is that they comparing cards which are not in the same price level.
The W9000 cost $4999, the K5000 less than half with $2299. Cou can buy K20c cards if you need something to compute which cost less than <$4000.

What AMD did was nothing else than a canned comparison between their and nVidia cards.

And looking at these numbers:
K5000 has the better perf/$ and perf/watt than the W9000. So if you need a Workstation card for Adobe K5000 is the much better deal...

Isn't that what I said?

 

[sontin]

Hm, yeah, yes...
My mistake.

 

[Imouto]

I really don't know why ppl is bashing OpenCL based on 3 years old news.

As for my experience it works flawless in SmallLuxGPU, the fastest ray tracing renderer for Blender, which is being ported for other 3D design apps like 3DS Max.

CUDA was cool for early adopters and Nvidia did a great job making it a must for researchers and other pros. OpenCL is going to take it over since it's not brand bound, it's royalty free and open. I can't really figure why some ppl is against it when it will run just fine in their machines or even better since it can use CPU + iGPU + GPU. You can even make ASICs to run your OpenCL code faster.

 

[Arkadrel]

@Sontin

on newegg:

K5000 = ~1800 $ (nvidia)
W8000 = ~1450 $ (AMD) (~10% faster than the k5000)
w9000 = ~3,400 $ (AMD) (~20% faster than the k5000)

The W8000 is about 10% faster than the K5000 (in this benchmark they posted).
While costing less.


Go check Newegg prices for those cards.

The problem with AMD's numbers is that they comparing cards which are not in the same price level.
The W9000 cost $4999, the K5000 less than half with $2299. Cou can buy K20c cards if you need something to compute which cost less than <$4000.

Bulls*** prices right there. Your makeing sh*t up.

W9000 = 3399 $
http://www.newegg.com/Product/Product.aspx?Item=N82E16814195116&Tpk=w9000&IsVirtualParent=1

w8000 = 1449$
http://www.newegg.com/Product/Product.aspx?Item=N82E16814195117

k5000 = 1799$
http://www.newegg.com/Product/Product.aspx?Item=N82E16814133468

 

[Jaydip]

@Sontin

on newegg:

K5000 = ~1800 $ (nvidia)
W8000 = ~1450 $ (AMD) (~10% faster than the k5000)
w9000 = ~3,400 $ (AMD) (~20% faster than the k5000)

The W8000 is about 10% faster than the K5000 (in this benchmark they posted).
While costing less.


Go check Newegg prices for those cards.

If this is the only benchmark people care about then sure but W9000 was not able to beat a 3 year old quadro 6000 across multiple benchmarks.Check THG,Legit(crap but a source nonetheless)

 

[Arkadrel]

If this is the only benchmark people care about

People that do this stuff professionally usually buy the PC that suits their 1 and only need (in this case Adobe® Premiere Pro). A person that used that program, would buy AMD reguardless of how much faster supposedly a k6000 was in other programs (that he might never need to use).

 

[Jaydip]

People that do this stuff professionally usually buy the PC that suits their 1 and only need (in this case Adobe® Premiere Pro). A person that used that program, would buy AMD reguardless of how much faster supposedly a k6000 was in other programs (that he might never need to use).

Fair enough.I was just not sure how popular this Adobe® Premiere Pro is.Also K6000 isn't a reality still

 

[ams23]

Are you suggesting they publish OpenCL vs CUDA on nvidia cards?

There is no reason to be obtuse here. The differences in performance in Adobe Premiere Pro (using pre-release software) are likely due to physical differences in the GPU hardware (coupled with differences in the GPU software drivers) rather than any inherent differences in OpenCL vs. CUDA. Think logically about it. If an AMD GPU was slightly faster using DirectX software compared to a completely different NVIDIA GPU using OpenGL software, would it make sense to conclude that DirectX is inherently faster? Of course not.

Everyone does that lol (nvidia,intel,amd,...). No one falls for it.

Whether or not other people do it doesn't make AMD's marketing graph depiction and scale any less deplorable in this case.

 

[BallaTheFeared]

Leader Nvidia continues to reap the bulk of the market's rewards, nudging its market share up to 81.7%, with number two AMD at 17.7%.

http://jonpeddie.com/publications/workstation_report/

It's nearly as lopsided as AMD vs Intel in CPUs...

 

[96Firebird]

OpenCL will kill CUDA @ mainstream easy, taking with it physx and that kind of garbage.
For HPC it will take a while, but Intel may be more than a threat.

What is "mainstream"?

 

[Ibra]

Why the negative tone, the implications, and name calling?
Why do people feel the need to mock others on this forum?

Im gonna go out on a limb and guess, hes innocent of what you accuse him off, because most people usually are (and its the right thing to do, until proved otherwise).

Im guessing wand3r3r doesnt even have a account there.
Neither do I for that matter, before you jump on me.

^ No I don't have an account with NV.



You seem well versed on NV marketing, is it a coincidence?
Now go back into your cave.

I am curious about developers actually dealing with the topic, not shills opinions.

So why do cry that Lonbjerg find out AMD's marketing BS?

 

[Jaydip]

http://jonpeddie.com/publications/workstation_report/

It's nearly as lopsided as AMD vs Intel in CPUs...

Thanks for sharing the links Balla.It is really weird when NV is basically winning with their old hardware.

 

[wand3r3r]

So why do cry that Lonbjerg find out AMD's marketing BS?

Why you don't English?

Put your Internet here, you lost: https://forums.geforce.com/

 

[Cloudfire777]

@Sontin

on newegg:

K5000 = ~1800 $ (nvidia)
W8000 = ~1450 $ (AMD) (~10% faster than the k5000)
w9000 = ~3,400 $ (AMD) (~20% faster than the k5000)

The W8000 is about 10% faster than the K5000 (in this benchmark they posted).
While costing less.


Go check Newegg prices for those cards.



Bulls*** prices right there. Your makeing sh*t up.

W9000 = 3399 $
http://www.newegg.com/Product/Produc...irtualParent=1

w8000 = 1449$
http://www.newegg.com/Product/Produc...82E16814195117

k5000 = 1799$
http://www.newegg.com/Product/Produc...82E16814133468

This means very little for many since they are deeply invested in CUDA and their code is CUDA oriented. Seriously, price/performance is not everything. However it would be very nice if OpenCL was used by everyone in projects all around the world. If that happens, I`m gonna join you guys in the negative campaign against Nvidia and their high prices.

Why you don't English?

Put your Internet here, you lost: https://forums.geforce.com/

Why haven`t this guy been banned yet? He does nothing but troll this forum

 

[Will Robinson]

Why haven`t this guy been banned yet? He does nothing but troll this forum

wand3r3r : Join Date_ May 2008

Clownfire777: Join Date_ a couple of weeks ago...