AMD: Success of small cores vs. big cores?

[BrightCandle]

I could very well be wrong but I think hyperthreading is going to be dropped soon. It seems like a good way to get an additional 20% performance, but as the number of real cores increases the value of having virtual cores that slow down the primary ones reduces for the grand majority of cases. Makes a lot of sense on low core counts (1 especially) but at 4 cores its already questionable and at 8 cores its actually often a hindrance especially in the server world.

I also wonder if we aren't going to continue to see a gradual increase but a massive jump to smaller cores in the coming years. Rather than going 8 cores to 16 cores to 32 etc we might just see 64 small cores instead as an alternative to 8. Denser compute performance but less single thread performance.

 

[mrmt]

I could very well be wrong but I think hyperthreading is going to be dropped soon. It seems like a good way to get an additional 20% performance, but as the number of real cores increases the value of having virtual cores that slow down the primary ones reduces for the grand majority of cases. Makes a lot of sense on low core counts (1 especially) but at 4 cores its already questionable and at 8 cores its actually often a hindrance especially in the server world.

If Intel keeps improving their core designs the gains from SMT will be much smaller and confined to a smaller set of tasks that it might not be worth the trade off to implement it. The first Core didn't have it, Silvermont didn't. Or Intel might decide to beef up the number of resources in the core and then SMT becomes an interesting proposition again.

The thing is, SMT is already a proven concept. There were successful products on the market that made use of it: Intel Core, Intel Atom, Intel Itanium, IBM POWER, SUN UltraSPARC. You may question the validity of SMT for a given design, but not the value of SMT as a concept.

 

[Fjodor2001]

Not true.

So give company A and B exactly X amount of R&D money each, and they will always come up with the a product performing the same? And if you give company A more R&D money than B, company A will always design a better product? No matter what people work there, what strategy the company pursues, what it focuses on, and so on? Come on...

 

[mrmt]

So give company A and B exactly X amount of R&D money, and they will always come up with the same product? And if you give company A more R&D money than B, company A will always design a better product? No matter what people work there, what strategy the company pursues, what it focuses on, and so on? Come on...

As IDC once said, in R&D you don't necessarily get what you pay for, but you surely won't get what you did NOT pay for. A bigger budget allows you to have much more aggressive targets than a smaller budget. Whether you can execute to reach these targets is a different matter, and is what causes what you are describing (achieving more with less). But you just can't target some parameters if you don't have the right amount of R&D budget. Try this and you'll have to compromise on time or give up some other part of your scope.

AMD did as well as it did with K7 and K8 because of sound execution, but also because they leveraged on the results of Alpha's much bigger R&D budget and from the entire MPU design team they poached once the company shut down. Unless Qualcomm or SUN MPU business take the same route, don't expect AMD to put another K7. AMD simply cannot aim as high as Intel, Qualcomm and others.

 

[senseamp]

If AMD exits the big core market, Intel could run into anti-trust issues in the server market, and be forced to license x86 to other players. Then it's lights out for their margins.

 

[CHADBOGA]

Denser compute performance but less single thread performance.

Single thread performance will always matter for desktop usage.

Most people would agree that a full 8 core CPU would be overkill on the desktop for the vast majority of users, so imagine how misplaced the notion of 16 or 32 cores at the expense of single threaded performance would be.

 

[Fox5]

As IDC once said, in R&D you don't necessarily get what you pay for, but you surely won't get what you did NOT pay for. A bigger budget allows you to have much more aggressive targets than a smaller budget. Whether you can execute to reach these targets is a different matter, and is what causes what you are describing (achieving more with less). But you just can't target some parameters if you don't have the right amount of R&D budget. Try this and you'll have to compromise on time or give up some other part of your scope.

AMD did as well as it did with K7 and K8 because of sound execution, but also because they leveraged on the results of Alpha's much bigger R&D budget and from the entire MPU design team they poached once the company shut down. Unless Qualcomm or SUN MPU business take the same route, don't expect AMD to put another K7. AMD simply cannot aim as high as Intel, Qualcomm and others.

More so, I'd say the fall of AMD gives a chance for Apple or one of the other ARM players to poach their engineers and do something amazing.

 

[Fjodor2001]

Current iGPUs are also overkill for most users. So then I assume there's no need to improve performance on desktop CPUs. They are completed.

 

[Fjodor2001]

As IDC once said, in R&D you don't necessarily get what you pay for, but you surely won't get what you did NOT pay for. A bigger budget allows you to have much more aggressive targets than a smaller budget. Whether you can execute to reach these targets is a different matter, and is what causes what you are describing (achieving more with less). But you just can't target some parameters if you don't have the right amount of R&D budget. Try this and you'll have to compromise on time or give up some other part of your scope.

AMD did as well as it did with K7 and K8 because of sound execution, but also because they leveraged on the results of Alpha's much bigger R&D budget and from the entire MPU design team they poached once the company shut down. Unless Qualcomm or SUN MPU business take the same route, don't expect AMD to put another K7. AMD simply cannot aim as high as Intel, Qualcomm and others.

With all the billions of R&D dollars that Intel spends, you have to say it's really crappy of them to still not be able to produce a better iGPU than AMD, or a better mobile CPU than ARM. Especially when that's where the R&D focus has been for Intel the last couple of years. How is that even possible given the logic you described?

 

[mrmt]

With all the billions of R&D dollars that Intel spends, you have to say it's really crappy of them to still not be able to produce a better iGPU than AMD, or a better mobile CPU than ARM. Especially when that's where the R&D focus has been for Intel the last couple of years.

How much does Intel actually spend in R&D for iGPUs?

And if something went in the R&D pipeline two years ago it didn't see the light of the day yet.

 

[mrmt]

More so, I'd say the fall of AMD gives a chance for Apple or one of the other ARM players to poach their engineers and do something amazing.

No need for them to go bankrupt. Samsung already poached a team of MPU designers from AMD.

 

[Fjodor2001]

How much does Intel actually spend in R&D for iGPUs?

And if something went in the R&D pipeline two years ago it didn't see the light of the day yet.

Likely quite a lot, since iGPUs and mobile is where Intel's R&D focus has been for quite some time. They've been focusing on it for a lot more than two years.

Just look at mobile - quite recently Intel announced that their mobile group lost $929 million in 2014Q1 alone (i.e. $3.5 billion on a yearly basis). So they have to be spending some serious R&D money on it.

 

[mrmt]

Likely quite a lot, since iGPUs and mobile is where Intel's R&D focus has been for quite some time. They've been focusing on it for a lot more than two years.

Just look at mobile - quite recently Intel announced that their mobile group lost $929 million in 2014Q1 alone (i.e. $3.5 billion on a yearly basis). So they have to be spending some serious R&D money on it.

Mobile also includes the LTE modem, digital radio, the crashing of the core line up, Sofia, Quark, and whatever is inside Intel R&D pipeline that wasn't disclosed, and all these unrelated to iGPU R&D. The loss you mentioned also includes the contra-revenue for Bay Trail tablets and phones. You can't really say that Intel is inefficient in their GPU R&D because we can't break down their R&D that small. It's fairly safe to say that they spend proportionally more than AMD for the target they are trying to reach, because they are developing a set of skills they didn't have, but we cannot say whether they are spending slightly more or a lot more than AMD on this.

And you can't judge their performance by the investment they are making since two years ago, because it takes 4 years for the results of this surge in R&D to reach the market.

 

[PPB]

What I did question is:the fact that AMD could not fix its CMT implementation

This is as baseless as you can get. Firstly, what is to be fixed and what not? On PD, AMD adressed a lot of the low hanging fruit in the bulldozer design (CMT wasnt touched), on SR, AMD recovered the MT penalty of CMT without incurring in a CMP design (so, still CMT for our standards).

To fix something, you have to distinguish if its broken in the first place. A lot of people here can tell you a lot of things going wrong on the BD design, CMT is tangent to those issues at best (most of them are pandemic in AMD's designs, like the laughable slow higher level caches).

the fact that the other player that tried CMT also got a bad product, and all the other players with money shunned the approach. Coincidence?

Intel dropped SMT starting from Yonah until Nehalem, SMT suddenly became a bad choice because of that? Let's not incur into fallacies please.

And there were other successful products on the market using the SMT concept. You can't say the same about CMT.

Not having something to compare it to actually adds to my point that first you need another example to put CMT into value. SMT at first was a total hit from Intel, sadly because of how inefficient netburst was (it also was the correct choice when it was added to the dreadfully slow first Atom core). Now we can tell the obvious limitations lying in the SMT when it is used in a really efficient uarch like Haswell.

If you lurk deeper you can see a lot of these cases happening. SB re-coupled the last level cache clock to the core clock, and Haswell decoupled it, again. No context = no proper comparison can be made.

 

[mrmt]

To fix something, you have to distinguish if its broken in the first place. A lot of people here can tell you a lot of things going wrong on the BD design, CMT is tangent to those issues at best (most of them are pandemic in AMD's designs, like the laughable slow higher level caches).

What does CMT brings of positive in a design, and what's the trade off for adopting it?

 

[raghu78]

What does CMT brings of positive in a design, and what's the trade off for adopting it?

CMT brings better scaling than SMT. SMT can get you 10 - 25% while CMT can get you 50 - 60% scaling. The biggest problem with Bulldozer was a very unbalanced design where there were not enough execution units driving single thread perf.

The decision to go with just 2 integer ALUs per cluster was a horrible one. It meant AMD regressed on single thread perf. AMD went down from 3 integer ALUs in K10 (Phenom I & II) to 2 in Bulldozer. AMD desperately needed atleast 3 integer ALUs or maybe even 4. If AMD delivered atleast 30 - 40% better single thread performance than Phenom II , then the module concept would have made sense.

 

[mrmt]

CMT brings better scaling than SMT. SMT can get you 10 - 25% while CMT can get you 50 - 60% scaling.

What's the trade off?

 

[Enigmoid]

What's the trade off?

More complex and takes longer and more money to design and validate.

 

[mrmt]

More complex and takes longer and more money to design and validate.

What about die area? Do you save die area or do you end up using more die area than with conventional designs? And what about power consumption? Does the added complexity for the CMT design adds up power consumption? And what about power gating? Does the added complexity hinders power gating of the core or CMT is neutral on the subject? And cache and memory management? Does CMT adds complexity for memory and cache management or is CMT neutral here?

 

[monstercameron]

What about die area? Do you save die area or do you end up using more die area than with conventional designs? And what about power consumption? Does the added complexity for the CMT design adds up power consumption? And what about power gating? Does the added complexity hinders power gating of the core or CMT is neutral on the subject? And cache and memory management? Does CMT adds complexity for memory and cache management or is CMT neutral here?

are you implying that those issues are limited to CMT designs?

 

[mrmt]

are you implying that those issues are limited to CMT designs?

No. How did you reach this conclusion?

 

[monstercameron]

No. How did you reach this conclusion?

just based on how hard you have been railing on AMD's CMT design, meanwhile you praise intels SMT design...

 

[mrmt]

just based on how hard you have been railing on AMD's CMT design, meanwhile you praise intels SMT design...

Oh, I got what you mean now.

Yes, I'm implying that CMT is negative to all these parameters I mentioned. But no, CMT is not the only way to screw up a design.

 

[Enigmoid]

just based on how hard you have been railing on AMD's CMT design, meanwhile you praise intels SMT design...

AMD's CMT design really hasn't worked well at all. That is probably why he keeps criticizing it.

 

[ShintaiDK]

There havent been a successful CMT design yet, unlike SMT.