AMD faces class action suit over Bulldozer missrepresentation

[Dabombinable]

What software shows is pretty irrelevant don't you think? That's just the way the system manager's programmers chose to have the program say in that situation.

Nope. Windows shows Bulldozer as having 1 module per thread (or ALU), so does Cinebench, so does Linux. So do other programs.

[zMeul]

What software shows is pretty irrelevant don't you think?

it's actually very relevant since consumers don't know what they bought - and in this case, I've had and seen a lot of circumstances where people were confused on what Windows was showing, namely 4 cores

you show them the block diagram of the CPU and you might as well show them string theory

[TheElt]

the point with 4+4 cpus is that they still have 8 cores, but in a benchmark you'll only see half of them in action. That does not mean they don't have 8 cores.

OK, count it as a quadcore when under load, it still beats the scaling of an FX 4100 at 4.5GHz (Taken from the Cinebench thread) with lower multithreaded turbo speeds.

 

IIRC there were some patches to the way Windows scheduling works on FX chips to fix the performance issues when a task was ran on one thread of one module and one thread of a different module, so that if the application uses 2 threads it will be executed on one module instead of spread across 2, and so on

[patrickjp93]

It's sold as an 8 core to the layman;

http://anandtech.com/bench/product/697?vs=1544

When you then look at scaling;

http://anandtech.com/bench/product/697?vs=1544

It doesn't do that. Looking at, for example, the R15 scores. You have;

6600K, Single 169, Multi 645

8350, Single 96, Multi 640

Intel's MT score is 3.87x the ST score. AMD's MT score is 6.66x the ST score. It doesn't scale up to 8 cores when held under scrutiny, Intel's chip is much close to perfect scaling. If you sell an 8 core, it better damn well work like one. It's more like a 6.5 core.

So it's sold as something it isn't, I really don't see how this is any different.

I'm afraid you can't use that scaling factor in its own to make that claim. The reality is it scales beyond a multiplication of 4. How much of the program is parallel will determine the score and scaling just as much as the number of cores you have. It's Amdahl's Law in benchmark form. That's also not taking into account noisy neighbor effects in the L3 Cache.

[Tomsen]

A bulldozer module does have some performance penalty under workloads on both clusters. The bottleneck lies in the decoders, around a 20% performance penalty module wide. It was fixed with steamroller, where they doubled the decoders.

But it does have a significant high scaling still even under load.

IIRC windows used to show them as 2 cores per module.

They should have went with the module and thread terms, as it is much more descriptive.

But it is not a good term marketing wise. the "core" term is much more recognizable for the average consumer.

[harrynowl]

correct me if I'm wrong, but that scheduler "issues" jobs in order, not in parallel and/or out of order!??!

Out of order execution of instructions has been a thing on AMD processors since K5 from the 90s.

 

Each integer pipeline has its own schedule, it just comes from a single fetch unit so the steps between decode and scheduling are quite latency heavy I expect. The integer unit probably sits there forever waiting for something to work on.

 

Here's a diagram that is quite useful:

[KemoKa]

This is the sort of shit that you have to go through to market a server processor to the consumer. Redefine what a "Core" is and then sell.

AMD should never have sold Bulldozer to the consumer. That was just laziness on their part.

[Sauron]

Nope. Windows shows Bulldozer as having 1 module per thread (or ALU), so does Cinebench, so does Linux. So do other programs.

 

So? It's still software... it shows what it was programmed to show. They probably cound fpus as indicators of a core. A core is not defined by what a piece of software says it is, if anything the piece of software should adapt accordingly. Not that it matters in any way since the way it's called doesn't affect performance.

 

it's actually very relevant since consumers don't know what they bought - and in this case, I've had and seen a lot of circumstances where people were confused on what Windows was showing, namely 4 cores

you show them the block diagram of the CPU and you might as well show them string theory

 

I've seen a couple of threads of people asking why their computers contain 8 i7s after seeing the performance tab in task manager. People clearly don't know what they buy regardless. Do you expect them to know the difference between a core, a module, a thread and an fpu? Don you think they'd have known any better if it had been advertised as a "quad module" cpu? What they can understand is a performance chart, and those haven't been manipulated in any way. But the thing is. these people don't even bother looking for those, even if a 5 minute google search would have provided them with all the information they needed to make an informed decision. The reality of the matter is they bought the first random prebuilt that said "gaming" on it or had a shiny red case and fit their arbitrarily set budget. It could have contained fairy dust and a bowl of hopes and dreams for all they knew. So bear with me when I say that not only do they have no basis for a class action lawsuit, they pretty much got what was coming to them.

[PlayStation 2]

This lawsuit has so much shit resting in it it's not even funny

Technically they're cores but eh, we'll see how it pans out

[Majestic]

I'm afraid you can't use that scaling factor in its own to make that claim. The reality is it scales beyond a multiplication of 4. How much of the program is parallel will determine the score just as much as the number of cores you have. It's Amdahl's Law in benchmark form. That's also not taking into account noisy neighbor effects in the L3 Cache/

 

Ok fair enough, but at the very least you can address consumer expectations by altering the product description. People don't expect 8 core scaling on a 4770K, since they know it's 4C/8T. Same logic should be applied to the FX series.

[PlayStation 2]

So when AMD sells 3.5 cores there is no massive shitstorm on the forums?

 

Gotta love those double standards

These processors aren't 3 module 7 thread CPUs. The GTX 970 also did have 4GB of RAM and people who actually READ the specs crucified Nvidia for flat-out lying about the amount of ROPs and shit like that

[DuckDodgers]

Referring the main argument in the case, Bulldozer indeed is capable of executing 8 instructions/clock, since there are eight x86-compatible integer blocks, each one completed with it's own instruction pointer and instruction retiring queue and capable of executing at least one instruction per clock. From this PoV, both the OS and system's firmware treats the CPU as 8-core. What matter in the end is the (relative) performance, but that isn't part of the claimant's argument anyway. As for the shared FPU -- this is a separate block from the main pipeline and AMD had made it very clear, many times, in their marketing about BD in the past.

[KemoKa]

Okay, technically there are eight cores, but they're pretty sorry excuses for cores if they're beaten by quad-core i5s.

[zMeul]

People clearly don't know what they buy regardless. Do you expect them to know the difference between a core, a module, a thread and an fpu?

I expect them to? probably not

but it's AMD's "job" to market and present their product accordingly

not just long ago nVidia did a no-no and marketed the GTX970 as a 4GB VRAM card when the reality was quite different - the exact same standard is expected from AMD

do you recall when AMD made fun of this whole "VRAM gate" - own it!

[zMeul]

Okay, technically there are eight cores, but they're pretty sorry excuses for cores if they're beaten by quad-core i5s.

if I recall some Anand benches, Intel's i3 stomped the shit out of them 

[Tomsen]

Okay, technically there are eight cores, but they're pretty sorry excuses for cores if they're beaten by quad-core i5s.

Different cores have different performance.

Not an apples to apples comparison

[Sauron]

I expect them to? probably not

but it's AMD's "job" to market and present their product accordingly

not just long ago nVidia did a no-nbo and marketed the GTX970 as a 4GB VRAM card when the reality was quite different - the exact same standard is expected from AMD

do you recall when AMD made fun of this whole "VRAM gate" - own it!

 

Go look at the specs. The gtx 970 has exactly 4gb of vram, not a single byte less. That is not what they... "misrepresented". What they misrepresented was the gpu core itself. According to the specs the reviewers got, the gtx 970 could handle 4gb of vram at full speed. In reality, the last 500mb block needed to be accessed through a single channel instead of 2, which means the bandwidth gets about halved while accessing it. The architectural graphs were wrong. AMD did no such thing; the full, correct specs were available to everybody from day one. It is a completely different situation.

[Sauron]

Okay, technically there are eight cores, but they're pretty sorry excuses for cores if they're beaten by quad-core i5s.

 

nobody says quad core atoms are dual cores just because they get beaten by desktop i5s.

[zMeul]

Go look at the specs. The gtx 970 has exactly 4gb of vram, not a single byte less. That is not what they... "misrepresented". What they misrepresented was the gpu core itself. According to the specs the reviewers got, the gtx 970 could handle 4gb of vram at full speed. In reality, the last 500mb block needed to be accessed through a single channel instead of 2, which means the bandwidth gets about halved while accessing it. The architectural graphs were wrong. AMD did no such thing; the full, correct specs were available to everybody from day one. It is a completely different situation.

and again, the "problem" is how it's used - because those 3.5GB and 0.5 cannot be used simultaneously - at the time, Anand had a great article explaining the situation

[Sauron]

and again, the "problem" is how it's used - because those 3.5GB and 0.5 cannot be used simultaneously - at the time, Anand had a great article explaining the situation

 

No - the problem is that nvidia (intentionally or not) hid the way it worked. It is correct to say the card has 4gb of vram. If they had been clear on how it worked since the beginning nobody would have been upset about it. So the problem with that situation is not the card """""defect""""" itself, but rather the fact that it was hidden. AMD didn't hide anything. There is no problem.

[Prysin]

Phone SoC produces are not pulling the same shit (Except Motorola with the X8).

There is a difference between saying your GPU is a quad core (which it might very well be), and counting all the cores together and only advertising that.

 

What AMD is doing (was doing?) with their APUs is exactly as if Intel had advertised the i5-2500K as a 16-core chip. That's bullshit and only there to trick consumers. You should list the CPU cores and GPU cores separately. You can't just add them together and advertise that number.

actually NO

 

an APU CAN be 12 cores depending on THE APPLICATION YOU USE IT WITH.

 

As long as it is HSA compliant, which Kaveri is. It can under OpenGL 2.0, Vulcan and DX12 by ALL technicality use ALL their cores, even the GPU cores, for other tasks then GPU tasks.

 

yes the workload MUST be parrallellisised to do so.

yes it is a specific use case scenario

yes such use cases are NOT a consumer thing, and probably wont be for years to come.

[Inkz]

Well I have heard some people refer to a regular i7 as an 8 because of HT. People should just do their research before buying.

[KarateHottie93]

Just some guy wanting easy money.

 

The whole 4 cores cut in half thing was exposed a long time ago. What defines a core anyways?

[KemoKa]

Different cores have different performance.

Not an apples to apples comparison

 

nobody says quad core atoms are dual cores just because they get beaten by desktop i5s.

Obviously different cores are built for different applications. Atoms weren't even part of the argument. I'm talking about how pathetic FX processors are in relation to their direct rivals, the i5 chips they were built and sold to compete with. They tried convincing people that more cores are better, because bigger numbers are better, right?

And they still lost. Here we are almost 5 years after Bulldozer was first launched, and Intel has dominated in IPC by leaps and bounds. And AMD has released little to nothing.

[Prysin]

Ok fair enough, but at the very least you can address consumer expectations by altering the product description. People don't expect 8 core scaling on a 4770K, since they know it's 4C/8T. Same logic should be applied to the FX series.

just gonna add to your scaling thingy earlier...

each module performs about 1.5-1.75 cores worth of output.... If you use cinebench R15, then before doing the test run you disable Core 0, 2, 4, 6 in taskmanager you get a score . If you disable core 1,3,5,7 you get a lower score...

 

tested it with my own FX (by mistake, then redid it to confirm wtf was going on...)..

my FX consistently scored 9 points higher if i forced it to use core 1 for single thread rather then core 0....

(Core 0 = 105, Core 1 = 114. Chip was OCd to 4.52GHz when i found this out....It behaves similarily at stock too)

 

a CPU should not score higher or lower based on the core you select. They should have equal output. In terms of FX, it is not like this.