Are AMD cpus true octo-cores?

[AtenRa]

But that's where you're 100% wrong. It's *nearly* the identical thing, just implemented slightly differently, by using more hardware in AMD's case.

Then a dual socket (Multi-chip) or dual core(Multi-core) is the same thing as HT(SMT) and CMT, all of them are multi-threaded

 

[myocardia]

Then a dual socket (Multi-chip) or dual core(Multi-core) is the same thing as HT(SMT) and CMT, all of them are multi-threaded

Talk about a strawman. I think you just invented a new category of strawman, with this post.

edit: Did you need a definition of what the word "nearly" means, by chance?

 

[AtenRa]

edit: Did you need a definition of what the word "nearly" means, by chance?

There are so many differences between SMT and CMT that "nearly" doesn’t apply to it.

 

[myocardia]

There are so many differences between SMT and CMT that "nearly" doesn’t apply to it.

That I agree with. Your problem though is you are arguing that they are NOT two different ways to do the same thing, which they of course are.

 

[AtenRa]

That I agree with. Your problem though is you are arguing that they are NOT two different ways to do the same thing, which they of course are.

SMT was created to maximize the utilization of a single core resources, CMT was created to minimize the die area and power consumption of two cores.

SMT shares the entire core between two threads. CMT only shares the Front-End, Execution is done in two different Execution Units unlike SMT that shares a single Execution Unit.

 

[jaqie]

You are both right... it's opposite approaches to the same problem, lowering TDP and die space needed while upping total compute power of the CPU.

One does this by beefing up a single core, the other does this by slicing off big bits of one core and conjoining it to another full core. Opposite approaches to the same type of goal.

 

[Charles Kozierok]

I think this was the comment that got this particular subdiscussion going:

And yet, although AMD's approach to the exact same thing, while using slightly more hardware to do so, is now somehow magically different?

I'm afraid I have to agree with the others that this somewhat downplays the significance of the difference between Intel's HT and AMD's modules. HT mostly just duplicates architectural state, while an AMD module has significantly more execution resources than a straight single core. I've read estimates that HT increases die space by 5%, while an AMD module is 50% larger than a conventional single core.

Ultimately the goals are the same no matter what approach is taken, but these are pretty different ways of getting there.

What I always wonder about is what the difference is between an AMD module and a single core with extra schedulers and a whole lot of EUs.

 

[parvadomus]

What I always wonder about is what the difference is between an AMD module and a single core with extra schedulers and a whole lot of EUs.

Complexity

 

[myocardia]

SMT was created to maximize the utilization of a single core resources, CMT was created to minimize the die area and power consumption of two cores.

SMT shares the entire core between two threads. CMT only shares the Front-End, Execution is done in two different Execution Units unlike SMT that shares a single Execution Unit.

This is what I've been saying all along: two different approaches to the same problem. They are both getting "extra" performance, from less than complete additional cores. This isn't hard to understand, is it?? Neither Intel nor AMD have 8 complete cores in either of the two CPU's we're discussing, no matter what any "fan" of either company tries to imply.

 

[myocardia]

I've read estimates that HT increases die space by 5%, while an AMD module is 50% larger than a conventional single core.

Ultimately the goals are the same no matter what approach is taken, but these are pretty different ways of getting there.

Did you completely miss the part where I said "AMD uses slightly more hardware to do so"? Fine, I was trying to give AMD some credit, that I guess they don't deserve. They use close to 50% more hardware, and gain a couple of percentage points worth of performance by doing so, in some software. All in all, I still like the design of the AMD chips more. I just wish they were much more efficient. I'd still love, just for fun, to see how much better Intel could fab Visheras, compared to ones fabbed by Global Foundries.

 

[Idontcare]

SMT was created to maximize the utilization of a single core resources, CMT was created to minimize the die area and power consumption of two cores.

SMT shares the entire core between two threads. CMT only shares the Front-End, Execution is done in two different Execution Units unlike SMT that shares a single Execution Unit.

This is what I've been saying all along: two different approaches to the same problem. They are both getting "extra" performance, from less than complete additional cores. This isn't hard to understand, is it?? Neither Intel nor AMD have 8 complete cores in either of the two CPU's we're discussing, no matter what any "fan" of either company tries to imply.

SMT and CMT are both on the same continuous curve that exists as a spectrum of non-CMP microarchitectures in which more (or less) hardware resources are shared for the purpose of processing more than one thread on a processor.

SMT is an extreme case of CMT in the limit of sharing nearly all the hardware resources, minimizing the performance gain while maximizing the die-savings.

IMO the only reason why people get up in arms over debating the relevance of the differences between AMD's implementation (which delivers better thread-performance at less die-space savings) versus that of Intel's (which delivers less thread-performance at better die-space savings) is because somewhere in the spectrum between CMP and SMT lies AMD's specific implementation of CMT and AMD's marketing wants to count them as full-fledged cores whereas Intel's marketing is willing to concede the point and not count their extreme-CMT design as two individual cores.

If Intel marketed their i7-3770k at an 8-core CPU then people would be quite down on Intel about their questionable definition of cores, and if AMD marketed their FX8350 as a 4C/8T processor then people would be generous in their applause of AMD's thread-scaling efficiency for the thread-performance from the virtual cores.

Marketing has muddled the discussion because that is why marketing exists. But the difference between the two microarchitectures merely comes down to the degree of resource sharing at the hardware level for supporting multiple threads as neither are CMP and both are CMT (with SMT being an extreme case of CMT).

 

[mrle]

I've read estimates that HT increases die space by 5%, while an AMD module is 50% larger than a conventional single core.

Actually just 12% of module area is "cloned".

What I always wonder about is what the difference is between an AMD module and a single core with extra schedulers and a whole lot of EUs.

Me too, but Intel is adding two more EUs in Haswell, so I guess we'll soon find out if it makes any difference.

 

[ShintaiDK]

Intel adds <5% to the core(It was <5% with P4, most olikely alot less today), while AMD adds 5% to the entire die is quite a difference.

 

[Charles Kozierok]

Actually just 12% of module area is "cloned".

I have no idea where the 12% comes from. The 50% figure I used comes straight from Anand, who got it from AMD back when Bulldozer was announced.

IMO the only reason why people get up in arms over debating the relevance of the differences between AMD's implementation (which delivers better thread-performance at less die-space savings) versus that of Intel's (which delivers less thread-performance at better die-space savings) is because somewhere in the spectrum between CMP and SMT lies AMD's specific implementation of CMT and AMD's marketing wants to count them as full-fledged cores whereas Intel's marketing is willing to concede the point and not count their extreme-CMT design as two individual cores.

If Intel marketed their i7-3770k at an 8-core CPU then people would be quite down on Intel about their questionable definition of cores, and if AMD marketed their FX8350 as a 4C/8T processor then people would be generous in their applause of AMD's thread-scaling efficiency for the thread-performance from the virtual cores.

/thread

Those two paragraphs are the entire issue.

 

[mrmt]

I have no idea where the 12% comes from. The 50% figure I used comes straight from Anand, who got it from AMD back when Bulldozer was announced.

The 12% number was posted by John "IPC won't go down" Fruehe in his AMD blog. I'd recommend gargantuan quantities of salt with this number.

 

[SPBHM]

Intel adds <5% to the core(It was <5% with P4, most olikely alot less today), while AMD adds 5% to the entire die is quite a difference.

indeed but it's for 4 cores, I also got a bit confused with this, it was apparently 5% for the original Northwood P4,
but now with the huge cache, IGP taking a lot of die space, HT probably adds almost nothing in terms of die size, with no clear sacrifice to anything else, it looks like a lot of "free" performance...

 

[aigomorla]

So has the real definition of cores been hashed out yet?
Or we going back to dowsing rods looking for cores and then counting what they say?

You know whats funny... 90% of us here probably dont know how to really answer this question yet we can get the answer correct because we "feel" it has this many cores.
^ hence my dowsing rod example

 

[MightyMalus]

A single core is whatever the consumer thinks that runs the computer. Times 2, 4, 6 or 8. Depending on the Company and amount of $$$Dough$$$ that they spent on Marketing.


On topic: I think AMD 8350 has 8 cores. I'm not sure why, but you know, I feel it.
http://www.youtube.com/watch?v=2P5qbcRAXVk <- click, takes 2 seconds...
<.< >.>

 

[Ajay]

Wow, this debate in just so pointless. What really matters is does it satisfy end user requirements. Why are we so hung up on pigeonholing everything?? All we accomplish when we do that is to pigeonhole our own thinking.

 

[Cerb]

So has the real definition of cores been hashed out yet?

A single general-purpose processor has the following basic requirements:
1. Increment or decrement a number.
2. Compare if a number is less than or equal to zero.
3. Change the program counter based on the result of requisite 2.
4. Read or write from or to an arbitrary memory address* (usually set using functionality from requisites 1 to 3 to set sources and destinations).
5. Terminate computation based on the results of requisite 2 (usually implemented by requisites 2 and 3).

A "core" needs to be able to do all of that, independently from other, "cores." So, a post-RISC x86 CPU, "core," must have its own ALU, AGU, and LSU. If a processor blurs the lines, what is important is that they define their, "core," and remain consistent with their definition.

The term, "core," largely exists because, "multiprocessor," historically refers to what we call a single processor, so they needed to call them something else. It's a name to help categorize and differentiate integrated components.

* this may include stacks, registers, special local stores, etc., though with registers or stacks, more than one such memory must be available, for any practical program.

 

[Ajay]

Core used to refer to Memory! If one goes back far enough. I suddenly find it scary that I know that. Core has become a marketing term to a large extent, which is why AMD stopped using Modules and switched to cores since no one knew what they were talking about (heck, some enthusiasts didn't get it at first either).

You are correct to an extent though Cerb as designs for computers moved from multiple discrete components to an increasingly more integrated single component: http://en.wikipedia.org/wiki/Microprocessor

I'm not sure if your list comprises exactly what "Turing-Complete" is, but I assume that's what you are getting at.

 

[Sleepingforest]

Wow, this debate in just so pointless. What really matters is does it satisfy end user requirements. Why are we so hung up on pigeonholing everything?? All we accomplish when we do that is to pigeonhole our own thinking.

Ironically, there's someone complaining in GPU about uninformed consumers who don't care enough about the technical details. Looks like we can't make everyone happy.

 

[cytg111]

There is no good answer to this, because an 8-core (as AMD calls it) CPU is neither 4 cores nor 8 but something in between.

After thinking about this for a long while, and changing my mind several times, I eventually decided to call a four core Intel with HT a "4 core, 8 thread" CPU, and an 8-core Family 15h AMD chip an "8 integer core, 4 FPU" chip. I think those are the fairest terms that make clear what the chips are.

I suppose, but does it really matter what we call them? Facts about the setups as i've come to understand them

1 AMD core under load = 1x performance
2 AMD cores under load, same module = 0.8x performance + 0.8x performance

1 Intel core under load = 1y performance
2 Intel threads (same core, 1 real, 1 hyper) = 1y performance + 0.3y performance

While you get more uniform performance level across cores with AMD, those cores WILL choke a bit when the same module gets the second thread. Intel on the other hand will have more consistent level on the one real core, unless you move into cache thrashing territory.
I like my single threaded performance very much. having 4 bastard childs on the side is just a bonus

 

[Accord99]

1 Intel core under load = 1y performance
2 Intel threads (same core, 1 real, 1 hyper) = 1y performance + 0.3y performance

If you bench the two threads through a Hyperthreading enabled core, the actual performance would be more around 0.6-0.65y per thread; each thread having equal access to the computing resources of the core.

 

[galego]

There is no good answer to this, because an 8-core (as AMD calls it) CPU is neither 4 cores nor 8 but something in between.

After thinking about this for a long while, and changing my mind several times, I eventually decided to call a four core Intel with HT a "4 core, 8 thread" CPU, and an 8-core Family 15h AMD chip an "8 integer core, 4 FPU" chip. I think those are the fairest terms that make clear what the chips are.

Bulldozer cores are based on the CIC (Clustered Integer Core) architecture developed by DEC in 1996.

An alternative notation used in some review sites is 4C/8T for Intel and 4M/8T for AMD.