Vice President and General Manager of AMD Radeon Gaming accuses NVIDIA GPP of monopolistic and anti competitive practices

[Razor01]

2 minutes ago, mr moose said:

Sounds like a very arbitrary number.   Unless you are spending the money on a free/g sync monitor there is no ecosystem.  And just to confuse the issue, 50% of all gaming monitors sold were freesync while something like only 20% of the market have a freesycn capable card.  Which indicates even if AMD do absolutely nothing,  another 30% of the current market is intending to buy an AMD card in the future or don't understand what they are buying.  

https://www.contextworld.com/display-research-update-09-02-2017

 

The boycotts we are talking about don't even have that kind of detailed planning involved.  Boycotts people are referring to here are not about AMD gaining market share but about punishing Nvidia for perceived injustice.   AMD gaining market share is just the natural consequence of a successful boycott (which I don't believe actually exists or at the very least would be barely measurable).

 

Of course, every business is the same. any market share is good market share.

Not too sure how this relates to my post, we are talking about boycotts not corporate propaganda.

 

Again not too sure what this has to do with boycotts or what I said, but I will say that gameworks has earnt nvidia more market share than bad PR. That makes it a success.  whether a small portion of people hate it or not.

 

 

For a boycott to have any meaningful impact it must meet 4 major criteria:

 

1 Customers must care passionately.  -and enough of them to stick with it

2.The cost of participation must be low. -for most consumers with any type off budget this will be very hard

3.The issues must be easy to understand.  -given the debate over the many Nvidia/AMD topics I'd say most people don't fully understand the issue

4.The mass media is still essential.  -never going to happen with such an isolated product like GPU's, if it were an everyday table item that every households buys then sure.

 

 

 

https://hbr.org/2012/08/when-do-company-boycotts-work

 

 

 

Yep like dolphin killing tuna, that was

 

1) easy to understand

2) most people are passionate about animals

3) cost is very little also alternatives are there too

4) Mass media picked it up fast.

[mr moose]

1 minute ago, Razor01 said:

 

Yep like dolphin killing tuna, that was

 

1) easy to understand

2) most people are passionate about animals

3) cost is very little also alternatives are there too

4) Mass media picked it up fast.

And for every major boycott planned like that that does meet the criteria there are another 30 that fail.  I know I harp on about nestle, but there is also the mcnuggets campaign,  the GMO campaign, the BIG pharma campaign and  the Nike campaign.  All companies who don't even know how many people refuse to buy their product because those figures are lost in the noise of the statistics.

 

 

P.S we don't know if they actually did anything other than label their tins as dolphin friendly tuna.  

[Razor01]

4 minutes ago, mr moose said:

And for every major boycott planned like that that does meet the criteria there are another 30 that fail.  I know I harp on about nestle, but there is also the mcnuggets campaign,  the GMO campaign, the BIG pharma campaign and  the Nike campaign.  All companies who don't even know how many people refuse to buy their product because those figures are lost in the noise of the statistics.

 

 

P.S we don't know if they actually did anything other than label their tins as dolphin friendly tuna.  

We can always hope companies are ethical but in reality......

[laminutederire]

18 hours ago, leadeater said:

Caffe is about the worst thing to use to compare OpenCL vs CUDA considering Caffe was explicitly developed for CUDA, it's native vs non-native comparison so it doesn't take a genius to figure out which will produce better performance.

 

Best to leave this conversation chain alone though, it's mightily off topic.

Yeah I know my point was just a few optimizations led to up to a 5x speedup, leaving the OpenCL to be just non optimized software wise, and the AMD cards and future Intel GPU could be on par with Nvidia's with better optimized software for their respective arch. He could say that titan is 3 times faster, fact remains that in many publications they sped up the runtime by more than 3 times already, so there could be optimizations more to make. That's mostly on AMD and Intel to do them though.

As for the discussions I gave up when he won't even recognize he mistook a fury X for a fury, calling me a liar because I took the figures for the fury, when he was taking the wrong figures himself of the fury X. Nothing more to add if someone do not even want to admit a single trivial mistake.

[Razor01]

6 hours ago, laminutederire said:

Yeah I know my point was just a few optimizations led to up to a 5x speedup, leaving the OpenCL to be just non optimized software wise, and the AMD cards and future Intel GPU could be on par with Nvidia's with better optimized software for their respective arch. He could say that titan is 3 times faster, fact remains that in many publications they sped up the runtime by more than 3 times already, so there could be optimizations more to make. That's mostly on AMD and Intel to do them though.

As for the discussions I gave up when he won't even recognize he mistook a fury X for a fury, calling me a liar because I took the figures for the fury, when he was taking the wrong figures himself of the fury X. Nothing more to add if someone do not even want to admit a single trivial mistake.

AMD and Intel MUST MAKE THE HARDWARE that can actually use those types of extensions first.

 

Its not optimization of the libraries, its actual changes to the API extensions (or additional extensions).

 

What do those extensions do?  They expose hardware features.  Its been 2 years since CUDA 8 released with those features of nV hardware.  And some of those features were there in CUDA 7 and Maxwell, so we are actually talking about 4 years in some cases.  Open CL hasn't exposed those features in AMD or Intel hardware.  If it was so easy to get that type of performance, they would have gotten those types of extensions into Open CL.  Since they haven't, most likely there are things in the hardware that aren't there.

 

Again do you know what the differences are between the ISA of nV and AMD GPU's?  That will determine the API differences.  Just like nV's Maxwell couldn't do concurrent kernel executions in repeated cycles, most likely Intel, AMD, didn't put those features in silicon.

 

Don't make assumptions about things you don't seem to have a firm grasp on.

 

You don't think Apple would have liked a 3 or 4 or 5 time speed up in some cases for their softwares?  Yeah so you have the BIGGEST company in the world, who by the by is also the BIGGEST PROPONENT of Open CL, and the biggest semiconductor business in the world not making any effort to promote something that would actually make them competitive with the up and comer?  So they just hand over the entire HPC GPU market on a silver platter? At this point AMD is just a garnishing and left overs because if these companies don't see the benefit, why would AMD right?  Right that makes business sense man when they can just add a few things to the API and optimize software for it *sarcasm*.  IF this was a three step program, they haven't even taken step one and its been 4 years!

[laminutederire]

12 minutes ago, Razor01 said:

 

Sure like and cards can't do matrix computation or FFTs. What are your credentials as a deep learning specialist to basically ignore whatever everyone said?

Personally I won't respond to you starting now.

I made my case about the fact that the biggest deep learning optimizations are closely related to memory layouts. Amd cards can basically do everything faster than Nvidias if there isn't a memory bottleneck. One part can be fixed by amd by changing their memory layout, but another can be done in software by handling data differently.

Now if you want to discuss with people, start by admitting you were wrong on the tflops for instance where you were blatantly obviously wrong about and then we'll see that you're maybe here to discuss instead of ignoring everybody's response. (Yes I mean everybody's).

Oh and by the way, go work and you'll see that business don't necessarily dictates optimizations. Optimizing is not cheap and that's why it isn't being done. It's cheaper for Apple and AMD to market to a different crowd rather than optimizing everything so far. AMD is starting to change on that now that they actually have a bit of money.

 

[Razor01]

6 minutes ago, laminutederire said:

Sure like and cards can't do matrix computation or FFTs. What are your credentials as a deep learning specialist to basically ignore whatever everyone said?

Personally I won't respond to you starting now.

I made my case about the fact that the biggest deep learning optimizations are closely related to memory layouts. Amd cards can basically do everything faster than Nvidias if there isn't a memory bottleneck. One part can be fixed by amd by changing their memory layout, but another can be done in software by handling data differently.

Now if you want to discuss with people, start by admitting you were wrong on the tflops for instance where you were blatantly obviously wrong about and then we'll see that you're maybe here to discuss instead of ignoring everybody's response. (Yes I mean everybody's).

 

DL, not much, AI and CUDA and Open CL quite a bit, I worked in banking, Credit Suisse, where we built an AI based trading system on CUDA in the mid 2000's.  One of the first fully automated trading systems based on AI.  I was a project manager and producer for the project.  But GPU architecture, and programming over 20 years of experience so there ya go.

 

What experience do you have?  Probably nothing.

 

Optimizing is not cheap?  These companies won't be optimizing SHIT, the software companies will do that. 

 

You don't optimize API's, there is no such thing as API optimization.  Just add in the features that expose the hardware functionality and the software guys can do the rest.

 

What you want freakin rewrite how these things work now?  Come on.

 

You might think 90% of the people here are retarded and don't know how these things work, but man they do.