AMD faces class action suit over Bulldozer missrepresentation

[dalekphalm]

Says I.

 

Cores always had symmetrical floating point and integer math capability. If you make a CPU effectively a 4 core for floating point operations, and an 8 core for integer calculations, you have a 4 core. Rule of weakest link.

Right, so you've just confirmed my point for me. There's no official, industry standard definition. Therefore, the term "core" has a certain amount of flexibility to it. It evolves over time.

 

Your opinion is that a core needs to have both FPU and Integer. That's only your opinion. Not a fact, nor a definition.

 

So your saying that software should be changed to show Bulldozer in AMD's favour? Again: ALU are not capable of performing all  of a CPU's functions on their own, therefore the are not cores.

Software should evolve with current technology trends, yes. Not necessarily to show bulldozer in AMD's flavour. I hardly see how that would be the case. Would making that change make AMD's CPU's perform faster or do better in benchmarks? Nope. Therefore, it would not be in AMD's favour.

 

Also, what is a CPU function? That changes all the freaking time. New instructions to do new things are added over time. Old and redundant things are removed. AMD was banking on GPGPU, because a GPU does FPU wayyyyy better than a CPU ever will. Of course, their architecture sucked, the implementation sucked, and to top it all off, the Software Industry was (and still is) very slow in adopting HSA/HUMA, etc.

 

That doesn't mean their cores weren't cores though. It just means they were shitty cores.

 

Frankly, if their architecture had been more effective, I doubt anyone would be claiming that a module was only one core. The concept makes perfect sense. Why is the CPU doing FPU when the GPU does it way more effectively? An on-die GPU that can co-process a task with the CPU makes perfect sense in theory.

 

What this comes down to is simply this: Everyone who thinks AMD is wrong: You're just making up an arbitrary definition of a CPU core. Please feel free to provide an industry accepted definition and prove me wrong.

[Dabombinable]

Right, so you've just confirmed my point for me. There's no official, industry standard definition. Therefore, the term "core" has a certain amount of flexibility to it. It evolves over time.

 

Your opinion is that a core needs to have both FPU and Integer. That's only your opinion. Not a fact, nor a definition.

 

Software should evolve with current technology trends, yes. Not necessarily to show bulldozer in AMD's flavour. I hardly see how that would be the case. Would making that change make AMD's CPU's perform faster or do better in benchmarks? Nope. Therefore, it would not be in AMD's favour.

 

Also, what is a CPU function? That changes all the freaking time. New instructions to do new things are added over time. Old and redundant things are removed. AMD was banking on GPGPU, because a GPU does FPU wayyyyy better than a CPU ever will. Of course, their architecture sucked, the implementation sucked, and to top it all off, the Software Industry was (and still is) very slow in adopting HSA/HUMA, etc.

 

That doesn't mean their cores weren't cores though. It just means they were shitty cores.

 

Frankly, if their architecture had been more effective, I doubt anyone would be claiming that a module was only one core. The concept makes perfect sense. Why is the CPU doing FPU when the GPU does it way more effectively? An on-die GPU that can co-process a task with the CPU makes perfect sense in theory.

 

What this comes down to is simply this: Everyone who thinks AMD is wrong: You're just making up an arbitrary definition of a CPU core. Please feel free to provide an industry accepted definition and prove me wrong.

Ok, please go back and look at the multithreaded scaling of the FX 6300/6350 vs AMD's own Phenom II X6. Bulldozer does not scale any where near as well as K10, so badly in fact that the FX 8350 only has the scaling of 7 cores. I'll quote this since people seem to ahve missed it:

 

Wall of text here, I went overboard and realised too late

 

Actually, we can use scaling in Cinebench to make these claims since it is probably almost perfectly parallel and that's backed up when you look at the scaling of every CPU other than the FX series. Take a look at the Cinebench thread and you'll see 80% or more on average performance increase when you double the amount of cores being used (Cores, not threads).

 

Athlon 64 x2 3800+ - 1 Core 41CB - 2 Cores 71CB - 1.78x scaling

C2D T8100 - 1 Core 60CB - 2 Cores 109CB - 1.82x Scaling

C2Q Q6700/Q6600 (Both at 4GHz) - 1 Core 102CB - 4 Cores 372CB - 3.64x Scaling This one might be a bit out due to different system configs, no idea how true it is

Phenom II x6 1090T (3.96GHz) 1 Core 102CB - 6 Cores 586CB - 5.74x Scaling

 

The scaling improves when you look at the later chips from Intel, and goes down the pan when you look at the FX chips from AMD

 

Core i5 4690k (4.6GHz) - 1 Core 181CB - 4 Cores 692CB - 3.82x Scaling

Core i5 3570k (4.4GHz) - 1 Core 156CB - 4 Cores 595CB - 3.81x Scaling

Core i5 2500k (4.6GHz) - 1 Core 155CB - 4 Cores 579CB - 3.74x Scaling

 

AMD FX 4100 (4.5GHz) - 1 Core 97CB - 4 'Cores'  312CB - 3.22x Scaling

AMD FX 6300 (5.17GHz) - 1 Core 124CB - 6 'Cores' 626CB - 5.05x Scaling Only score I can find done at the same speeds by the same person, but scaling shouldn't break with overclocking

AMD FX 8350 (4.6GHz) - 1 Core 110CB - 8 'Cores' 736CB - 6.69x Scaling

 

AMD has consistently worse scaling than Intel when you look at the FX series of chips, now this isn't because Cinebench isn't very parallel. Sadly there aren't any people benching the 6 and 8 core Intel chips with HT off, so math needs to be done.

If you look around at some benchmarks from here (Would have liked a sample of 3 CPUs, but damn it's hard to come across) HyperThreading gives a performance boost of 22.29% over those 2 CPUs, so we can use that as a rough idea of the benefits of HT.

 

If we look at a Core i7 4790k at 4.8GHz scores 981CB, if we also use an i5 4690k at 4.8GHz as a comparison, that scores 730CB. If you do the math and pull 22.29% off the score of the i7 it comes to 762CB, higher than the i5 but that's expected due to imperfect math, different hardware and the increased cache. So that 22.29% seems like a good enough value for rough calculations

 

Core i7 5820k (4.4GHz) - Single Core 179CB - All Threads 1323CB - 6 Cores (No HT) 1028CB - 7.39x scaling (HT) - 5.74x Scaling (No HT) Falls in line with what we've seen so far

Core i7 5960x  (4.5GHz) - Single Core 176CB - All Threads 1733CB - 8 Cores (No HT) 1347CB - 9.85x scaling (HT) - 7.65x scaling (No HT)

 

Even on rough calculations, we can see that Cinebench scales extremely well the more cores you add, like it should. I would make a point here by doing the same for the E5-2699 v3 but there are no single threaded scores for that CPU, so instead I'll do it for the closest CPU to it, the E5-2630 v2

 

2x 2699 v3 - All Threads 4447CB - 36 Cores 3456CB

2630 v2 (2.964GHz on the sheet) - Single Thread 124CB

 

35.86x Scaling with HT / 27.87x Scaling without HT

 

Now with this many cores the calculations will not be very accurate, including the fact that there is no single threaded 2699 v3 score makes it worse. But even with 36 cores Cinebench scales extremely well.

 

I've spent far too long doing this now that I realise it was a colossal waste of time so it's getting posted anyway, so it can serve to the thread as well. AMD should scale as well as Intel if they did make true 4, 6 and 8 core CPUs. However the scaling is far worse, because they are not true 4, 6 or 8 core CPUs.

 

The FX-6300 scales 0.69x worse than a Phenom II x6, they are both marketed as 6 core CPUs, it's obvious which one actually is a 6 core.

 

[Sauron]

OK, in that case then the ALU are irrelevant and Bulldozer only has 4 cores then, with each module being the core. AMD still lied about the core count. I can see where your coming from with the ALU since it was only after the 486 that they went from being co processor to the modern integrated ALU.

 

No, because the full module is clearly more than a single standard core, it's almost two. Furthermore, the definition of a core is not set in stone. Different architecture have different things that can be considered as cores for lack of a better word. I mentioned multiple times the gpu example - a CUDA core is completely different from an x86 core and is a lot less functional by itself, that doesn't mean that nvidia can't call it a core. Near the beginning of the thread I stated that I would much rather have had an 8 or 10 core phenom II with a die shrink than bulldozer and its derivatives.

 

Ok, please go back and look at the multithreaded scaling of the FX 6300/6350 vs AMD's own Phenom II X6. Bulldozer does not scale any where near as well as K10, so badly in fact that the FX 8350 only has the scaling of 7 cores. I'll quote this since people seem to ahve missed it:

 

Nobody is disputing the cores are bad. What we're saying is that they are effectively 8, and this lawsuit has no basis.

[Dabombinable]

No, because the full module is clearly more than a single standard core, it's almost two. Furthermore, the definition of a core is not set in stone. Different architecture have different things that can be considered as cores for lack of a better word. I mentioned multiple times the gpu example - a CUDA core is completely different from an x86 core and is a lot less functional by itself, that doesn't mean that nvidia can't call it a core. Near the beginning of the thread I stated that I would much rather have had an 8 or 10 core phenom II with a die shrink than bulldozer and its derivatives.

 

 

Nobody is disputing the cores are bad. What we're saying is that they are effectively 8, and this lawsuit has no basis.

I'll keep on saying this until you start to actually read everything: ALU are not cores. They can not function as a full Central Processing Unit on their own. ALU never have and never will be cores. Remove 1 ALU from Bulldozer and suddenly AMD would be calling it for example with the FX 8350 a quad core-and ALU aren't cores.

[Prysin]

Yes they were.

 

First of all, "compute" just means calculating. All GPUs can do that. What you mean when you say "compute" is GPGPU, and yes even the APUs that used VLIW4 could do that. Llano APUs had support for OpenCL (which like I said before is GPGPU). Hardware acceleration was supported in Photoshop back in CS6.

read my previous post.

 

Hardware acceleration was supported for Adobe products long before CS6...

depending on the application, you had CUDA support.

[LAwLz]

read my previous post.

Hardware acceleration was supported for Adobe products long before CS6...

depending on the application, you had CUDA support.

Then why did you say the VLIW4 based APUs didn't support it? They clearly did. That's why I posted the CS6 benchmark. To prove that you were wrong.

[Prysin]

Then why did you say the VLIW4 based APUs didn't support it? They clearly did. That's why I posted the CS6 benchmark. To prove that you were wrong.

again, READ my previous previous post.

[Kobz360]

Its nice to see all of you guys coming in with these complex arguments, really is good to read.

However, at the end of the day, there's no clear cut definition of a 'core'and combined with the fact that the legal system isn't full of computing geniuses, i'd have to bet AMD would win this suit quite easily.

You can't just base a 'core' off what Intel (for example) believes one is. Bulldozer was definitely misleading to an extent, but I think they did enough to ensure people that not all cores are equal. How retailers and whatnot deal with this is another issue.

[TheElt]

No, because the full module is clearly more than a single standard core, it's almost two. Furthermore, the definition of a core is not set in stone. Different architecture have different things that can be considered as cores for lack of a better word. I mentioned multiple times the gpu example - a CUDA core is completely different from an x86 core and is a lot less functional by itself, that doesn't mean that nvidia can't call it a core. Near the beginning of the thread I stated that I would much rather have had an 8 or 10 core phenom II with a die shrink than bulldozer and its derivatives.

 

 

Nobody is disputing the cores are bad. What we're saying is that they are effectively 8, and this lawsuit has no basis.

Effectively and actually are two very different terms. An i7 effectively has 8 cores, but it doesn't

 

And the scaling point still stands, if you take one core and add another, you see a near enough 80%+ scale in performance regardless of architecture or whatever as long as the clocks are the same, if you add another core and you're only getting 60% or less scaling, then either something is wrong with the core, the clocks or, and it might just be this, it's not actually a core

 

The numbers do not add up, if it is a true 8 core then where is that 20%+ scaling going missing?

[LAwLz]

again, READ my previous previous post.

Can you link me to the specific part you are referring to?

 

 

Oh never mind. Just noticed that you have made massive edits in your posts. It would be nice if you could make new posts instead of editing your old ones. Makes it a lot easier for people to follow the conversation (including me).

It's also kind of a dick move to edit your false statement and then pretend like you were right all along. You were wrong.

[Prysin]

Can you link me to the specific part you are referring to?

 

 

Oh never mind. Just noticed that you have made massive edits in your posts. It would be nice if you could make new posts instead of editing your old ones. Makes it a lot easier for people to follow the conversation (including me).

It's also kind of a dick move to edit your false statement and then pretend like you were right all along. You were wrong.

i edited that post as you were writing your reply to it. Not after.

[Dabombinable]

Ok, I'll say this for context. My 387 maths coprocessor, or ALU, was never sold as a CPU because it isn't one and can not function as one. Modern multi core were originally invented by AMD themselves to remove or reduce the need for motherboards with multiple CPU sockets, thus saving space and greatly reducing latency between each Central Processing Unit and therefore increasing the multi threaded performance.

 

AMD marketing bulldozer as having 4, 6 and 8 cores was an outright lie, the same as their advertising of APU with 12 Compute Cores.

[Dabombinable]

This article on CMT from 2012 is an interesting read as well: https://scalibq.wordpress.com/2012/02/14/the-myth-of-cmt-cluster-based-multithreading/

[Deletive]

Ok, I'll say this for context. My 387 maths coprocessor, or ALU, was never sold as a CPU because it isn't one and can not function as one. Modern multi core were originally invented by AMD themselves to remove or reduce the need for motherboards with multiple CPU sockets, thus saving space and greatly reducing latency between each Central Processing Unit and therefore increasing the multi threaded performance.

 

AMD marketing bulldozer as having 4, 6 and 8 cores was an outright lie, the same as their advertising of APU with 12 Compute Cores.

Their APU's outright lies with their advertising as 12 Compute Cores, because that's what each compute core can do that. But there is a problem of whether or not the software can actually take advantage of the "12" cores. So it's not so much an advertising problem. However, HP rewording those compute cores into being actual cores while thats where you can say it's an outright lie. There are a lot of technical stuff in the bulldozer architecture so I'm not going to comment on the architecture until I have enough time.

[TechGod]

Their APU's outright lies with their advertising as 12 Compute Cores, because that's what each compute core can do that. But there is a problem of whether or not the software can actually take advantage of the "12" cores. So it's not so much an advertising problem. However, HP rewording those compute cores into being actual cores while thats where you can say it's an outright lie. There are a lot of technical stuff in the bulldozer architecture so I'm not going to comment on the architecture until I have enough time.

No. 12 "compute cores" is not a lie. 

[Dabombinable]

No. 12 "compute cores" is not a lie. 

Its extremely misleading however and few programs will see the APU as such. Also, most people buying computers with the APU will read it as being 12 cores since they don't know the difference.

[TechGod]

Its extremely misleading however and few programs will see the APU as such. Also, most people buying computers with the APU will read it as being 12 cores since they don't know the difference.

Oh please. They mention 4 CPU cores and 8 GPU cores. 

[Dabombinable]

Oh please. They mention 4 CPU cores and 8 GPU cores. 

And they get sold with them added up and called compute cores. It is misleading.

[Prysin]

Its extremely misleading however and few programs will see the APU as such. Also, most people buying computers with the APU will read it as being 12 cores since they don't know the difference.

i own a A10 7870k APU, i have yet to find a single fucking program reading it as 12 cores.

 

it always says 2 module/4 threads

 

in the GPU section is ALWAYS says R7 Graphics. It doesnt even mention there is 8 GPU "cores".

[Dabombinable]

i own a A10 7870k APU, i have yet to find a single fucking program reading it as 12 cores.

 

it always says 2 module/4 threads

 

in the GPU section is ALWAYS says R7 Graphics. It doesnt even mention there is 8 GPU "cores".

That's a program-not what it gets advertised as.

[Prysin]

That's a program-not what it gets advertised as.

 

Its extremely misleading however and few programs will see the APU as such. Also, most people buying computers with the APU will read it as being 12 cores since they don't know the difference.

 

your own words. Served to you on a plate.

 

 

here is some filler in info:

 

Whenever AMD (or partners) are to promote the new APUs, AMD tells us clearly that two sets of numbers should be quoted in reference to the Compute Cores – the total, and the breakdown of CPU/GPU on the APU. Thus this would mean that the A10-7850K APU would be marketed at a “12 Compute Core” device, with “(4 CPU + 8 GPU)” following immediately after. I applaud AMD's decision to not obfuscate the internal configuration of its APUs.

from Anandtechs article on the 7850k. In short, if you see it marketed as "12 core" on a website, you should actually report the website to AMD for misrepresenting the specifications of their products.

[That Norwegian Guy]

snip

 

You just let the courts decide that.

[bcredeur97]

so that explains why when you want to disable cores on an FX CPU, you have to disable two. 


Not that you should do this for any reasonable reason other than extreme overclocking...

[Dabombinable]

your own words. Served to you on a plate.

 

 

here is some filler in info:

from Anandtechs article on the 7850k. In short, if you see it marketed as "12 core" on a website, you should actually report the website to AMD for misrepresenting the specifications of their products.

So? Few programs would see, as in use, the iGPU and CPU as "12 compute cores".

They still get advertised as having 12 compute cores, on websites and in stores, its still misleading.

 

While AMD explains what the 12 compute cores are, they still call them 12 compute cores:

http://www.amd.com/en-us/products/processors/desktop/a-series-apu

 

The very term itself is misleading, and if AMD didn't want people to think that their APU were 12 cores they shouldn't even be using the term "12 compute cores". They know that most people buying their products wont understand the terminology at all even with an explanation as to what they mean by compute cores.

[dalekphalm]

You just let the courts decide that.

I think you need to take your own advice. Considering there are no industry standard definitions, I look forward to the courts setting a precedence on this (one way or the other).