AMD CPU speculation... and expert conjecture

[noob2222]

P.s. extremetechs flaws stem from benchmarks that dont scale 1:1 with respect to clockspeed.

I.E. does handbrake score 5 pts at 2.5 ghz and with the same exact cpu score 10 at 5 ghz? Or does instead only score 9 and give an inflated performance advantage to the slower cpu if you calculated icp? And yes, handbrake doesnt scale 1:1

 

[-Fran-]

In CISC uArchs, it's up to the CPU how it will tear it appart and then resolve how to solve the full instruction. The term IPC losses a LOT of meaning when current uArchs are doing 2 or 3 instructions at any given time. That's why you get IPC tied to benchmarks: the serialization of how the program (and OS) sends the instructions will be what type of "IPC" you get. That's why you basically have a different IPC with each program. I think palladin already explained this in more details a while ago. It's way better to talk raw performance instead of "IPC".

Cheers!

 

[Cazalan]

If the statement was to combine the best DNA of Jaguar and the best DNA of Bulldozer then it is a hybrid either way. Saying it is more Jaguar like or more Bulldozer like is a moot point. All this back and forth with relatively no real information is just word play.

 

[Bogdan Chivu]

amd skiped 22nm. Why? I dont think they(amd, intel) work to finish say skylane. I think they work 2generation cpu ahead. Amd was on top, now is intel. Amd in 2years will be on top again.

 

[Bogdan Chivu]

Ok intel is the best. but for a $500 rig, amd is best.

 

[blackkstar]

P.s. extremetechs flaws stem from benchmarks that dont scale 1:1 with respect to clockspeed.

I.E. does handbrake score 5 pts at 2.5 ghz and with the same exact cpu score 10 at 5 ghz? Or does instead only score 9 and give an inflated performance advantage to the slower cpu if you calculated icp? And yes, handbrake doesnt scale 1:1

I think you nailed it. IPC assumes that performance scales linearly with frequency. As you said, if something can do X amount of things in Y ticks of the clock (frequency), then it would do aX amount of things in aY ticks of the clock. Think about it. If you do 1 thing in 10 ticks of the clock, then you'd still do 2 things in 20 ticks of the clock. And that's simply not how these benchmarks scale.

If you're not seeing performance scale perfectly with frequency, then you're not measuring instructions per clock tick. You're measuring something else.

 

[jdwii]

jdwill, I apologize for my post asking if you are drunk. You are usually more articulate in what you post.

It seemed to me out of character for you to post that.

Again I apologies for making that assumption.

Ha ha yeah its all good its my own fault.

 

[juanrga]

I find it funny when people cant even read an article properly. Gamerk hit the nail on the head, ipc is a theoretical "maximum" number thats affected by multiple aspects during any given test program.

No. IPC can be theoretical or can be measured; it can be maximum, minimum, or average. It depends of what you are doing and how. In the Extremetech article the IPC they obtained is an average value obtained from measuring the IPC of different benchmarks and taking the average.

Good point tg. One thing ill point out also, IPC is independent of benchmarks, how can processors swap places during certain benchmarks? Benchmarks arent testing IPC. In todays cpu designs, raw IPC is a useless term when talking actual performance.

No. IPC depends of the benchmark used and even depends of the version of the benchmark. A new binary compiled with newest compiler or more aggressive flags will run faster on the same processor and will give higher IPC values.

P.s. extremetechs flaws stem from benchmarks that dont scale 1:1 with respect to clockspeed.

I.E. does handbrake score 5 pts at 2.5 ghz and with the same exact cpu score 10 at 5 ghz? Or does instead only score 9 and give an inflated performance advantage to the slower cpu if you calculated icp? And yes, handbrake doesnt scale 1:1

No. Extremetech used the well-known expression

Performance = IPC * frequency

and applied it correctly to the CPUs at stock.

In fact the above equation explains the two main kind of CPUs that exist. If engineers focus on optimizing the frequency part at cost of IPC the CPU is known as a speed demon, if engineers focus on optimizing the IPC term at cost of frequency the machine will be a brainiac design.

What you are doing is comparing the same CPU at different frequencies and ignoring that what surrounds a CPU will not scale up. That is your mistake.

 

[jdwii]

When people use the term IPC i have to ask what instruction? I myself used to do such a thing but a CPU can perform better using certain instruction sets vs another. This is something i feel even writers on tomshardware fail to realize.

 

[juanrga]

If the statement was to combine the best DNA of Jaguar and the best DNA of Bulldozer then it is a hybrid either way. Saying it is more Jaguar like or more Bulldozer like is a moot point. All this back and forth with relatively no real information is just word play.

The statement was more concrete and mentioned what part of Jaguar DNA "small dense core" and what part of Bulldozer DNA "high frequency" would be mixed, and suggests that Zen will be a minicore that can hit frequencies close to 4GHz.

Moreover, we have info/leaks from Fottemberg that Zen will be a "mini-core" "Jaguar-Puma style".

It also makes sense that Zen is a a mini-core Jaguar-Puma style because Zen is specifically targeting the semicustom/embedded market towards which AMD is pivoting.

There is an idiom "The devil is in the details" I suggest you learn It.

That refers to important details that make a difference, not to the well-known concept of "irrelevant details" that add nothing except noise. We can ignore you pretension to discuss if the measured 2.69% gap is better characterized as "less than 3%" or as "less than 10%", specially when we are here discussing possible increases in IPC of ~40% and the error on the estimation will be superior to what you want discuss.

 

[gamerk316]

I can't tell the difference between a FX8320 and a A6-3620 either without benchmarks. The difference between CPUs is so small now I doubt anyone can really tell an i7 from a pentium without running benchmarks. Even in most games, if you have a powerful enough GPU, there really isn't that much difference.

cute. i had this r7 260x on an athlon ii x3. worlds apart compared to when i put it on my i5. so i'm sorry but what you're saying is BS.

ps. we own a net cafe.

Really now? Well, I guess we might do different stuff on our computers. I have had no problem on mine and it runs everything outside of gaming well compared to pretty much every computer I used in the past 3 years. I can't tell the nanoseconds the computer is able to load a webpage or open ms word since the bottlenecks are all elsewhere. In gaming, I don't see the difference between 60 FPS and 80 FPS since my monitor is only 60hz anyways.

And you don't notice the differing latencies, less stuttering, and the like? I doubt it. Even going from my QX9650 to a 2600k was a significant difference to me.

 

[gamerk316]

When people use the term IPC i have to ask what instruction? I myself used to do such a thing but a CPU can perform better using certain instruction sets vs another. This is something i feel even writers on tomshardware fail to realize.

That's why it needs to be measured on a per-benchmark basis. It's really impossible to do, which I why I always say the best you can do in practice is compare the % difference between two different chips, rather then directly measure IPC.

 

[szatkus]

Joel Hruska Admin szatkus • a year ago

1). Jaguar's single-threaded IPC is between 10-20% lower [than Piledriver] depending on the application, yes.

So Jaguar's ST performance is actually worse than PD. MT is very similar because modules. Agree?

Also remember that Jaguar were designed with lower clocks in mind than BD family. For example they used HDL before Carrizo and pipeline is shorter.

Anyway "the best from Cat and Bulldozer family" sounds good

 

[8350rocks]

"I don't see the difference between 60 FPS and 80 FPS since my monitor is only 60hz anyways."

NEWS everyone a Athlon II x3 can run farcry 4 and many other games at 60fps stable all the time. And if it can't its over the game being boring or crappy or some other excuse i come up with.

Sorry just sick of hearing such BS from people.

Just FYI, not everyone plays Far Cry 4.

Like me...I am not buying another ubisoft game for a while now. Sadly...their PC ports are such crap, along with a string of other issues.

 

[juanrga]

I am glad to see the great STILT in full agreement with me on that those leaked "Kaveri refresh" codename Godavari are "utter bullcrap"

http://www.xtremesystems.org/forums/showthread.php?289027-next-big-core-AMD-late-2015-or-Q1-Q2-2016-info-speculation&p=5245514&viewfull=1#post5245514

And he provides a ROM screenshot showing that codename Godavari is already in use (as I said).

It seems my source was right and that Anton Shylov, Swecklocers, WCCFTECH, and the "engineers" asked by 8350 were not...

 

[-Fran-]

He says that it's the codename of the GPU component as well. Does that mean that this Kaveri refresh will sport Tonga-level GCN? Or even a more advanced GCN version?

That's not bad by itself. Like the guy also mentions, this could be Richland to Trinity, so it's a good thing anyway. He did not dismiss anything else, Juan; none of your other red flags: clocks and... I can't remember the other stuff you said.

Cheers!

 

[blackkstar]

From what Stilt said, it sounds like the Kaveri refresh is just a clock speed bump because the process has improved. Doesn't look too exciting for me. Maybe they went back and rebinned spare inventory that hasn't been packaged yet?

 

[juanrga]

He says that it's the codename of the GPU component as well. Does that mean that this Kaveri refresh will sport Tonga-level GCN? Or even a more advanced GCN version?

That's not bad by itself. Like the guy also mentions, this could be Richland to Trinity, so it's a good thing anyway. He did not dismiss anything else, Juan; none of your other red flags: clocks and... I can't remember the other stuff you said.

Cheers!

Unfortunately, you got it all wrong, my friend.

He says that the leak doesn't makes sense (he used the term "utter bullcrap") because as I said before the codename is already in use

http://www.tomshardware.co.uk/forum/352312-28-steamroller-speculation-expert-conjecture/page-352#15152459

First red flag: "Godavari" is the codename of the iGPU on Mullins.

He adds that the new APUs "are still 7K series"; there is no Tonga-level GCN, there is no more advanced GCN version, just the same GPUs that in Kaveri. Again confirming what I said before about the 'refresh':

My last info is that "Kaveri refresh" now consists of new models with essentially nothing new beyond minor changes such as firmware fixes.

And finally, contrary to your claims, he says that Kaveri refresh is not Richland-like but something minor:

While "Kaveri Refresh" is generally true I personally wouldn't call it as a refresh.
To me refresh is something similar what Trinity went through when it morphed into Richland.
"Kaveri Refresh" will still be Kaveri as they are simply new models with minor tweaks to the clocks.

 

[juanrga]

From what Stilt said, it sounds like the Kaveri refresh is just a clock speed bump because the process has improved. Doesn't look too exciting for me. Maybe they went back and rebinned spare inventory that hasn't been packaged yet?

He said exactly "minor tweaks to the clocks". He didn't say "clock speed bump". If you pay attention to the slide given before almost all the refresh models have lower clocks than the current Kaveri models. The supposed new 850 has lower clocks than current 840. The supposed new 870K has lower clocks than current 860K. The supposed new A6-8550K has lower clocks than current A6-7400K and so on.

I cannot confirm/deny exact frequencies for each model. But my source said me that the current 7850K model will remain as the higher clocked model ever.

 

[szatkus]

He says that it's the codename of the GPU component as well. Does that mean that this Kaveri refresh will sport Tonga-level GCN? Or even a more advanced GCN version?

That's not bad by itself. Like the guy also mentions, this could be Richland to Trinity, so it's a good thing anyway. He did not dismiss anything else, Juan; none of your other red flags: clocks and... I can't remember the other stuff you said.

Cheers!

GPU at A10-8850K is labeled 130F. The same as in A10-7850K. It's just a refresh...

 

[8350rocks]

From what Stilt said, it sounds like the Kaveri refresh is just a clock speed bump because the process has improved. Doesn't look too exciting for me. Maybe they went back and rebinned spare inventory that hasn't been packaged yet?

He said exactly "minor tweaks to the clocks". He didn't say "clock speed bump". If you pay attention to the slide given before almost all the refresh models have lower clocks than the current Kaveri models. The supposed new 850 has lower clocks than current 840. The supposed new 870K has lower clocks than current 860K. The supposed new A6-8550K has lower clocks than current A6-7400K and so on.

I cannot confirm/deny exact frequencies for each model. But my source said me that the current 7850K model will remain as the higher clocked model ever.

Juan, your source must be out of whack.

Why on earth would a new design, that is a refresh, have lower clocks on a more mature process???

I think you are counting chickens a bit early.

 

[esrever]

I can't tell the difference between a FX8320 and a A6-3620 either without benchmarks. The difference between CPUs is so small now I doubt anyone can really tell an i7 from a pentium without running benchmarks. Even in most games, if you have a powerful enough GPU, there really isn't that much difference.

cute. i had this r7 260x on an athlon ii x3. worlds apart compared to when i put it on my i5. so i'm sorry but what you're saying is BS.

ps. we own a net cafe.

Really now? Well, I guess we might do different stuff on our computers. I have had no problem on mine and it runs everything outside of gaming well compared to pretty much every computer I used in the past 3 years. I can't tell the nanoseconds the computer is able to load a webpage or open ms word since the bottlenecks are all elsewhere. In gaming, I don't see the difference between 60 FPS and 80 FPS since my monitor is only 60hz anyways.

And you don't notice the differing latencies, less stuttering, and the like? I doubt it. Even going from my QX9650 to a 2600k was a significant difference to me.

I guess I don't play enough high demand games then. I doubt my frametimes and such are bad with my current setup for the games I play. CS:GO runs fine and so does hearthstone. Last demanding game I ran was ryze and it ran perfectly smooth on low settings for me. .

 

[juanrga]

From what Stilt said, it sounds like the Kaveri refresh is just a clock speed bump because the process has improved. Doesn't look too exciting for me. Maybe they went back and rebinned spare inventory that hasn't been packaged yet?

He said exactly "minor tweaks to the clocks". He didn't say "clock speed bump". If you pay attention to the slide given before almost all the refresh models have lower clocks than the current Kaveri models. The supposed new 850 has lower clocks than current 840. The supposed new 870K has lower clocks than current 860K. The supposed new A6-8550K has lower clocks than current A6-7400K and so on.

I cannot confirm/deny exact frequencies for each model. But my source said me that the current 7850K model will remain as the higher clocked model ever.

Juan, your source must be out of whack.

Why on earth would a new design, that is a refresh, have lower clocks on a more mature process???

I think you are counting chickens a bit early.

My source is not the origin of the claim about lower clocks. The source is this slide

http://twimages.vr-zone.net/2015/01/godavari.jpg

That slide from vr-zone is the origin of the sweclocker leak, and of further reproduction by Anton Shilov and WCCFTECH.

My source said me what I quoted in a former post. Read it.

It is funny that you are evaluating my source, instead evaluating yours. First, I recall what you said about the leak

The codename for next desktop part is finally leaked...godavari.

That is, indeed, the next desktop series cpu until Zen.

Also, wccftech must have gotten some good info from sweclockers, a lot of that lines up well with what I know.

You would ask yourself why your 'engineers' are unaware that Godavari is the codename of the iGPU on Mullins:

My source, myself, and The Stilt know what Godavari is...

http://www.tomshardware.co.uk/forum/352312-28-steamroller-speculation-expert-conjecture/page-352#15152459

http://www.xtremesystems.org/forums/showthread.php?289027-next-big-core-AMD-late-2015-or-Q1-Q2-2016-info-speculation&p=5245514&viewfull=1#post5245514

 

[8350rocks]

From what Stilt said, it sounds like the Kaveri refresh is just a clock speed bump because the process has improved. Doesn't look too exciting for me. Maybe they went back and rebinned spare inventory that hasn't been packaged yet?

He said exactly "minor tweaks to the clocks". He didn't say "clock speed bump". If you pay attention to the slide given before almost all the refresh models have lower clocks than the current Kaveri models. The supposed new 850 has lower clocks than current 840. The supposed new 870K has lower clocks than current 860K. The supposed new A6-8550K has lower clocks than current A6-7400K and so on.

I cannot confirm/deny exact frequencies for each model. But my source said me that the current 7850K model will remain as the higher clocked model ever.

Juan, your source must be out of whack.

Why on earth would a new design, that is a refresh, have lower clocks on a more mature process???

I think you are counting chickens a bit early.

My source is not the origin of the claim about lower clocks. The source is this slide

http://twimages.vr-zone.net/2015/01/godavari.jpg

That slide from vr-zone is the origin of the sweclocker leak, and of further reproduction by Anton Shilov and WCCFTECH.

My source said me what I quoted in a former post. Read it.

It is funny that you are evaluating my source, instead evaluating yours. First, I recall what you said about the leak

The codename for next desktop part is finally leaked...godavari.

That is, indeed, the next desktop series cpu until Zen.

Also, wccftech must have gotten some good info from sweclockers, a lot of that lines up well with what I know.

You would ask yourself why your 'engineers' are unaware that Godavari is the codename of the iGPU on Mullins:

My source, myself, and The Stilt know what Godavari is...

http://www.tomshardware.co.uk/forum/352312-28-steamroller-speculation-expert-conjecture/page-352#15152459

http://www.xtremesystems.org/forums/showthread.php?289027-next-big-core-AMD-late-2015-or-Q1-Q2-2016-info-speculation&p=5245514&viewfull=1#post5245514

My engineers knew they were reusing a codename. However, since one is GPU and the other CPU...they were not concerned about it. Considering all GPU designs are executed in Toronto, and all CPU designs are executed in Austin, it is fairly simple to keep them mostly separated.

 

[juanrga]

8350rocks, do your 'engineers' think that Kaveri refresh Godavari is a "CPU" and that is why reuse a codename from an existing GPU on an existing APU? Your answer is fascinating!

I can imagine those are the same 'engineers' that pretend that Zen will be faster than Skylake and the same that believed found a bug on ARM... And I believe are the same 'engineers' that did release that unfinished thing named Kaveri with broken power management (broken & misconfigured), broken NB (frequency limitation and memory latency) and buggy firmware (there are more things which are broken and incomplete than actually working) and that is why are now releasing a "refresh" to fix stuff...