AMD Dragged to Court over Core Count on "Bulldozer"

[cdawall]

Yea lets add words because that makes consumers happy. I like how you keep calling them cores however.

 

[FordGT90Concept]

Only when there's a preceding word that clarifies it is not traditional. Example: compact SUV or pygmy goat.

 

[cdawall]

Only when there's a preceding word that clarifies it is not traditional. Example: compact SUV or pygmy goat.

Using your own example, would a compact SUV still not be an SUV? Verbatim what you are saying is AMD is provided users 8 cores to their truest of forms.

 

[FordGT90Concept]

A compact SUV is much, much smaller than a typical SUV; likewise, a pygmy goat is much smaller than a typical goat. The former is a "type of" the latter but not the same as. A conjoined core is a type of core but not the same as a (conventional) core. So no, they didn't provide "8 cores," they provided "8 conjoined cores." That's an important distinction between the two.

Additionally, the first of a type usually defines what the normal is. For example, the popularity of the Ford Explorer (mid-sized, five passenger, front engine, rear wheel drive) established it as the class of SUV. Core was defined a decade ago in the x86 world by the Athlon 64 X2.

If you phoned a Ford dealer, told them to send you an SUV, and an Escape (compact SUV) showed up, would you not be disappointed?

 

[cdawall]

No because I asked for an SUV and that is by definition an SUV. This is also why everyone thinks your argument is incorrect.

I would be mad if I called for an SUV and they showed up with a Fiesta, those are two different things. Notice how the analogy works.

 

[FordGT90Concept]

Do note that that Ford Escape (compact SUV) starts at $23,600 where Ford Explorer (mid-sized SUV) starts at $31,660. They cheated you out of $8,060 or 25% of your vehicle. The exact same argument can easily be made against Bulldozer. If you can't see that, I guess we're going to have to agree to disagree on that point.

 

[cdawall]

You didn't ask for a specific SUV you said SUV. Thing is when you purchased an AMD you didn't buy an Intel, so asking for Intel's cores would be, well asinine. You can disagree all you want, but so far the courts agree that an SUV is an SUV.

 

[FordGT90Concept]

When I'm sold "8 cores," I expect the equivalent of two Phenom II X4s, not 10-30% performance loss because of the conjoining.

Ask yourself this: if AMD correctly labeled their Bulldozer products, would they have sold fewer of said products? If yes, AMD defrauded the public. If no, it's water under the bridge.

 

[cdawall]

When I'm sold "8 cores," I expect the equivalent of two Phenom II X4s, not 10-30% performance loss because of the conjoining.

Ask yourself this: if AMD correctly labeled their Bulldozer products, would they have sold fewer of said products? If yes, AMD defrauded the public. If no, it's water under the bridge.

I expect nothing more than what I got because I have the ability to read reviews prior to purchase. The performance loss was also only single core IPC, or poorly threaded applications neither of which has to do with conjoining and does have to do the the 2 ALU/AGU setup, tiny cache and overall piss poor design of the individual cores.

You know what I equate this two? Suing Intel because their 3ghz P4 wasn't as fast as AMD's Athlon 64. Is there a definition of what a ghz should equate to, is a small tire not a tire because it doesn't go on a dump truck? They are independent cores whether you see it that way or not, and if this ends up back in court and not dismissed, because it isn't worth the courts time AMD will end up winning because guess what? It's still a core even without and FPU or do you want everyone to start labeling boxes as 8 integer cores+ 4 floating point units? Hey you know something funny the integer section is called a core and the floating point a unit...Wonder if that's because one of them is required in a core, and the other is not.

 

[FordGT90Concept]

People that go to Best Buy and the like generally don't read reviews. They look at the labels and make their decision based on them.


Hz is a very scientific unit of measure (cycle per second). The difference is in what it does with clocks and that's not something easily differentiated by a product label. Perhaps there should be a standard made to make this clear--akin to horsepower and torque ratings on engines.

A tire on a bicycle shares little in common with a tire on a Caterpillar 797. The distinction is very important.

They are not "independent cores" (even AMD said they never made that claim in their arguments) they are very clearly "conjoined cores"--a very important distinction.

 

[BiggieShady]

I like how you keep calling them cores however.

We should all call them ATFKAC (a thing formerly known as core) after Dickie wins the case

They are not "independent cores" (even AMD said they never made that claim in their arguments) they are very clearly "conjoined cores"--a very important distinction.

Every execution engine needs a frontend and cache/mem in all architectures. It's not a huge paradigm shift to have two execution engines need only a single frontend.
It works like having two frontends at half clockspeed (yes, uops are dispatched left, right, left, right ... alternating both cores every other cycle).
Knowing that, how can you argue dependency?
Hint: you could argue performance, but not dependency.
It's completely possible that you are using adjective "independent of something" as "will work without something", and not referring to execution of uops at all.
If that's the case, let's do something stupid ... let's make, say, a pin #917 part of every core because none of the cores will work without pin #917 ... because one depends on the other

 

[cdawall]

People that go to Best Buy and the like generally don't read reviews. They look at the labels and make their decision based on them.

People who shop at best buy don't care and wouldn't know one was faster than the other unless you handed them a graph.

Hz is a very scientific unit of measure (cycle per second). The difference is in what it does with clocks and that's not something easily differentiated by a product label.

Oh yea where on the label does it state that a 3ghz pentium 4 is slower than a 3ghz athlon 64?

Perhaps there should be a standard made to make this clear--akin to horsepower and torque ratings on engines.

There would be no way to standardize what a ghz can do. It isn't HP/TQ which is an actual measurement of work done. That would be something akin to gflops, you know a measurement of work not speed. Ghz would be like measuring MPH, some take more HP/TQ to hit the same speed.

A tire on a bicycle shares little in common with a tire on a Caterpillar 797. The distinction is very important.

Or is it assumed that they are not the same because people aren't oblivious to the world.

They are not "independent cores" (even AMD said they never made that claim in their arguments) they are very clearly "conjoined cores"--a very important distinction.

Prove it? They are clocked independently, perform instructions independently etc.

 

[Recon-UK]

WOW i never seen this thread before.

I really dislike the FX series (i'm a huge fan of AMD though) but taking them to court over what they wanted to call cores?

REALLY you are going to take a top tier CPU maker to court on a subjective term?


LMFAO

 

[FordGT90Concept]

People who shop at best buy don't care and wouldn't know one was faster than the other unless you handed them a graph.

Dickey proves this wrong.

Oh yea where on the label does it state that a 3ghz pentium 4 is slower than a 3ghz athlon 64?

Never said it did. I said they should.

There would be no way to standardize what a ghz can do. It isn't HP/TQ which is an actual measurement of work done. That would be something akin to gflops, you know a measurement of work not speed. Ghz would be like measuring MPH, some take more HP/TQ to hit the same speed.

I think an IEEE standard for measuring instructions/second would suffice (would include a mix of standard instructions) not unlike how SAE measures towing and payload capacities now.

Or is it assumed that they are not the same because people aren't oblivious to the world.

You clearly have too much faith in the common consumer. Remember, the population we're talking about are the type that buy Beats Audio products just because some celebrity was paid to say it's good.

Prove it? They are clocked independently, perform instructions independently etc.

@BiggieShady's linked IEEE paper + AMD's published core diagrams prove it. They practically copied Kumar, Jouppi, & Tullsen's 2004 design with some tweaks (many of which hurt performance). That same paper explains why two "conjoined" cores will never perform on par with two "conventional" cores.

 

[BiggieShady]

... many of which hurt performance ...

It's mentioned as how little performance is hurt compared on how much die space is saved. You are interpreting it as a bad thing.

 

[FordGT90Concept]

What I was alluding to was suffocating it of L1 data cache. Cutting back on cache always saves die space but it also reduces performance--not unique to Bulldozer.

Yes, saving die space is the main advantage of conjoined cores. That's pretty much the only reason why any chip manufacturer would do it.

 

[cdawall]

Dickey proves this wrong.

He sighted Tom's Hardware as his only source of evidence proving it wasn't an 8 core CPU, if you would like I can post a core mentioning how it is an 8 core CPU and we can site that instead?

Never said it did. I said they should.

And there in lies the problem. AMD should mention that it uses a module design as opposed to a standard design of old, but you know what if every technology company had to write a dissertation or face criminal persecution every single time they attempted to innovate then guess what you will stop seeing. Remember Hyperthreading was an innovation saved from an atrocious netburst pile of shit design and currently exists in every single high end chip Intel sells almost as a status quo.

I think an IEEE standard for measuring instructions/second would suffice (would include a mix of standard instructions) not unlike how SAE measures towing and payload capacities now.

Standard instructions would show that AMD chip in a favorable light. Remember AVX isn't a standard instruction, most would consider it hardly used actually. Also how often do you update instructions sets? Every new generation of CPU typically has a new instruction set or two. A simple GFlops listing would suffice even if it is theoretical as they do with GPU's. It isn't AMD or Intels fault if current software doesn't utilize the chips.

You clearly have too much faith in the common consumer. Remember, the population we're talking about are the type that buy Beats Audio products just because some celebrity was paid to say it's good.

I actually have a simple view on this if the consumer is too stupid to realize what they are buying they really have no place to sue. Companies aren't there to inform ignorant people of how life works, if they were Microsoft would have been put out of business for Windows 10 upgrades. Quite a few stupid people just clicked next without reading what they did.

@BiggieShady's linked IEEE paper + AMD's published core diagrams prove it. They practically copied Kumar, Jouppi, & Tullsen's 2004 design with some tweaks (many of which hurt performance). That same paper explains why two "conjoined" cores will never perform on par with two "conventional" cores.

Quite a few HPC articles show near perfect scaling on the design. Remember these were designed for an HPC environment and then loosely adapted to be used in a consumer level product. The cores show more scaling than any multi thread design ever could because they are cores not just the ability to process another instruction.

 

[FordGT90Concept]

Quite a few HPC articles show near perfect scaling on the design. Remember these were designed for an HPC environment and then loosely adapted to be used in a consumer level product. The cores show more scaling than any multi thread design ever could because they are cores not just the ability to process another instruction.

It's not 100% and never will be which Kumar et. al. showed. Benchmarks show the same.

 

[cdawall]

It's not 100% and never will be which Kumar et. al. showed.

There isn't 100% scaling with two cores on any CPU. Software overhead will always prevent that.

 

[FordGT90Concept]

Software overhead is pretty consistent and easy to adjust for.

 

[cdawall]

Software overhead is pretty consistent and easy to adjust for.

Constant or not it is still not 100% scaling, adjusting numbers to compensate for that is actually telling one of those lie things. When you do that it can get you sued by consumers.

 

[64K]

I actually have a simple view on this if the consumer is too stupid to realize what they are buying they really have no place to sue. Companies aren't there to inform ignorant people of how life works, if they were Microsoft would have been put out of business for Windows 10 upgrades. Quite a few stupid people just clicked next without reading what they did.

I always due my research before buying too but this guy may not be responsible for his mistake if he can convince the jury that AMD failed at "truth in adverting" and that burden is on AMD and not the customer to advertise their CPUs honestly and clearly.

 

[FordGT90Concept]

Constant or not it is still not 100% scaling, adjusting numbers to compensate for that is actually telling one of those lie things. When you do that it can get you sued by consumers.

You're missing (avoiding?) the point. Bulldozer shows a decline that is not consistent with "conventional" cores.

-Yellow line represents ideal circumstances (800% for 8 threads, 700% for 7 threads, and so on)
-Maroon line represents software overhead. Most of that comes from the main thread (UI updates) which created a minor conflict with four threads and repeated at eight threads (two worker threads on the same core as the main thread).
-Orange line is: time of one thread / time of thread (e.g. 339% for 4 threads)

Note how far orange deviates from maroon after 4 threads, that is the result of SMT--a performance gain over a processor without it but a far cry from the throughput of eight conventional cores. I fully expect Bulldozer to land between those two lines.

 

[Aquinus]

You're missing (avoiding?) the point. Bulldozer shows a decline that is not consistent with "conventional" cores.

Have you used a CPU with more than 4 "real cores" by your own definition to show that your application is actually capable of speeding up to a reasonable extent past 4 threads? Most applications don't speed up very well and start hitting some form of limitation when it comes to concurrent processing of data, even more so if it's just making the full calculation run in parallel and not using different threads to handle different stages of the task. I suspect that if the trend for the times for both the 8350 and the 6700K are the same, that the way it's written merely doesn't speed up past so many cores. This doesn't make AMD's CPU not have real cores, it's just part of the reality that is designing software.

If you don't believe me than maybe you should send the binary to @cdawall to run on his Opterons to see if it scales past 4 or 5 threads and maybe another member who has a 6c or 8c Intel CPU to generate some numbers for us. This is a claim that can be validated, so it should be because not all software scales and unless it has been tested on a machine fitting your "real core" criteria with more than 4 of them, I would say that you have insufficient data to assert that your benchmark is even capable of showing such optimistic speed up with additional cores.

 

[Recon-UK]

PCSX2 can use an infinite amount of threads using GSDX Software mode, however it won't really scale passed 4 threads at all even with HTT.