AMD CPU speculation... and expert conjecture

[juanrga]

So Steamroller. http://. Looks like AMDFXBLOGSPOT has borrowed the data from WCCFtech and posted some comparative numbers between the Kaveri part and Iris Pro. All values given are GPGPU - which are astounding if true. I would like to see some numbers of the CPU alone, but that's not the goal of the Fusion HSA.

I'm hoping that the drivers/codes used are not finished and that we'll see even better performance when measured as PPW/ PPD$/ IPC /FTW!

If Kaveri does have a 13CU version that will be a pleasant surprise. Even if they are clocked slower (600Mhz vs 844Mz) it will still be a good 16% faster than 8C. Slower clocks mean better thermals and power consumption too.

Some tentative CPU alone and CPU+GPU (HSA) numbers are found here with comparison to Sandy Bridge and Ivy Bridge i5s.

About the supposed 13CU version of Kaveri:

During last days some rumours about a new Kaveri APU top model with larger number of graphics cores are taking the web, but the number of CUs reported is even, which seems strange; moreover, it is not known if the rumours correspond to some kind of dual graphics configuration (APU + discrete graphics).

The even number of CUs is interpreted by some as the software failing to report adequately a dual-graphics config. Others believe that one module of the supposed 3M version of Kaveri has been substituted by more iGPU and that the 13CU is a new top model Kaveri APU.

Also to add that the same guy tested an ES of a A10-6800k with the GPU clocked at 600MHz. Therefore his ES Kaveri being clocked at 600MHz GPU means nothing. I continue maintaining that Kaveri GPU will be clocked at ~900MHz.

 

[8350rocks]

is there a filter that I can turn on that automatically hides hafijur's comments or something?

+1

 

[8350rocks]

As I said before this is untrue.

Really, where is ARM's desktop OS? Ubuntu? They're not exactly a flagship OS in terms of volume you know. What else? Android? That's not a desktop OS, nor will it be anytime soon.

As I said before, it is not David vs Goliath thing.

It is Intel vs [ AMD + Nvidia + Samsung + Apple + Qualcomm ... ]

Ok, let's examine the *not* David vs Goliath real quick shall we?

AMD - Has not *yet* shipped an ARM CPU at all.

NVidia - highest selling ARM device is the shield with under 100k units sold. Not a major player

Samsung - they have *some* promise, though they're working out the kinks and relying on Qualcomm for their high end ARM devices.

Apple - entirely proprietary, combined with low DT presence, I doubt they're a serious threat at all. Plus iOS is not viable as a DT OS.

Qualcomm - The 1 major ARM player that *could* potentially make a move at desktop, except they're not...why? They know what you fail to recognize. ARM won't make a dent in the installed base of x86 PCs worldwide.

You realize out of 7,000,000,000 people...there are roughly 2,500,000,000 households. Of those 2,500,000,000 households, 2,250,000,000 have a desktop PC of some sort. Of those households with a PC, roughly 2,000,000,000 of those run windows, which is only on x86 hardware.

You getting my point yet? Windows has roughly 90% market share in desktop. 24% of the world's devices run on windows...and they're 99% desktop computers.

You see why it's futile for ARM to even bother? Linux is *FREE* and it still hasn't overcome Windows.

I hate M$ as much as the next guy for their stagnation, and proprietary software, that keeps the little guys, and innovative software writers, out of the mainstream. However, they have some clout in desktop PCs.

I am not anti AMD, on the contrary, I want them to succeed HUGELY. Though, I can see the forest for the trees and know what AMD pushing ARM is up against. It's a fool's errand.

You can compare it to future still to be released Warsaw, Berlin, Kaveri... if you want, the result is the same: Seattle is "already in the performance envelope" than x86 chips, because ARM64 chips are already in the same performance envelop than x86.

Only to add that the above figure is assuming lower clock for Seattle. AMD said that the chip would run a 2GHz or more. I did the GFLOP computation for 2 GHz. The A57 Chip is prepared to run up to 3Ghz or so.

Where are you extrapolating numbers from? Source?

You keep dodging the question.

 

[esrever]

so whats steamroller been up to?

 

[juanrga]

As I said before this is untrue.

Really, where is ARM's desktop OS? Ubuntu? They're not exactly a flagship OS in terms of volume you know. What else? Android? That's not a desktop OS, nor will it be anytime soon.

As I said before, it is not David vs Goliath thing.

It is Intel vs [ AMD + Nvidia + Samsung + Apple + Qualcomm ... ]

Ok, let's examine the *not* David vs Goliath real quick shall we?

AMD - Has not *yet* shipped an ARM CPU at all.

NVidia - highest selling ARM device is the shield with under 100k units sold. Not a major player

Samsung - they have *some* promise, though they're working out the kinks and relying on Qualcomm for their high end ARM devices.

Apple - entirely proprietary, combined with low DT presence, I doubt they're a serious threat at all. Plus iOS is not viable as a DT OS.

Qualcomm - The 1 major ARM player that *could* potentially make a move at desktop, except they're not...why? They know what you fail to recognize. ARM won't make a dent in the installed base of x86 PCs worldwide.

You realize out of 7,000,000,000 people...there are roughly 2,500,000,000 households. Of those 2,500,000,000 households, 2,250,000,000 have a desktop PC of some sort. Of those households with a PC, roughly 2,000,000,000 of those run windows, which is only on x86 hardware.

You getting my point yet? Windows has roughly 90% market share in desktop. 24% of the world's devices run on windows...and they're 99% desktop computers.

You see why it's futile for ARM to even bother? Linux is *FREE* and it still hasn't overcome Windows.

I hate M$ as much as the next guy for their stagnation, and proprietary software, that keeps the little guys, and innovative software writers, out of the mainstream. However, they have some clout in desktop PCs.

I am not anti AMD, on the contrary, I want them to succeed HUGELY. Though, I can see the forest for the trees and know what AMD pushing ARM is up against. It's a fool's errand.

You can compare it to future still to be released Warsaw, Berlin, Kaveri... if you want, the result is the same: Seattle is "already in the performance envelope" than x86 chips, because ARM64 chips are already in the same performance envelop than x86.

Only to add that the above figure is assuming lower clock for Seattle. AMD said that the chip would run a 2GHz or more. I did the GFLOP computation for 2 GHz. The A57 Chip is prepared to run up to 3Ghz or so.

Where are you extrapolating numbers from? Source?

You keep dodging the question.

Well, you can delete my entire answer to your post and then believe that I didn't answer you, but that will not change anything.

Interesting kaveri cpus or if you want to call it apus come with strong gpu integrated with them it seems :http://wccftech.com/amd-desktop-kaveri-apu-13-cus-enabled-radeon-r5-m200-832-stream-processors-gpu-spotted/

http://wccftech.com/amd-kaveri-apu-arriving-desktop-pcs-q4-2013-kaveri-graphics-demonstrated-virtual-reality-tech/
http://www.bjorn3d.com/2013/06/amd-richland-june-5-2013-1201am-edt-nda/
Juangra one of the people use your website I believe as fact lol. I wonder if the new steamroller 4 core kaveri apus are as fast as a top i3 haswell. I think it will be very close. You can see from here in most stuff current i3 haswell just blow away the a10-6800k richland apu so steamroller even with 30% improvement will struggle to match the i3 haswell. Theres no way a 4 core steamroller has any chance unless clocked to over 5ghz to compete against an i5 overall.
http://www.hardware.fr/articles/901-3/performances-applicatives.html

The French article coincides very much with my findings despite using a different platform. E.g. taking their

and assuming Kaveri will be about a 20% faster than Richland puts Kaveri at the level of the FX-6000 and i5 series, that is just what I found

Of course, this is without any HSA acceleration. With HSA acceleration the Kaveri APU would be 2x-3x faster than the fastest chip tested by them.

 

[The Q6660 Inside]

He got a "CPUs authority" Badge here at Toms. Some of you would start considering his posts seriously now... I will start laughing at badges.

I think if you really believe lets say the kaveri 4 core steamroller cpu is level with an i5 2500k it would literally be a miracle and 50% better then what amd predict with steamroller if both at same clock speed. On anything truly multithreaded the richland cpu gets destroyed by the 2500k:
http://www.anandtech.com/bench/product/288?vs=675

The 4 core kaveri steamroller cpu competitor is the i3 haswell. The i5 is in a different league unless the kaveri cpu is clocked at 4.5ghz it won't come as close to the i5 2500k as you predict overall. The 2500k is faster then an fx6300:
http://www.anandtech.com/bench/product/288?vs=699

Heck the i5 2500k is close to an fx8320 overall:
http://www.anandtech.com/bench/product/288?vs=698

Seems pretty good for the 8320, $50 cheaper after all... You do realize that a 30% increase in single core performance may very well put it @K10 or even SB per core right?

 

[palladin9479]

I agree about a lot of the limitations, however, it does provide some nice features for ports to PC.

Gamer I believe this is the point of Mantle. Games developed for the next generation of Consoles will support MANTLE natively anyway. When their brought to PC they will still have that support along with the additional DX path put in. Games developed exclusively on PC most likely won't have Mantle support. I think that's AMD's game plan here. To setup the situation where they get an advantage in PC gaming by supporting a console API on a PC platform.

If whatever comes after the XBONE / PS4 doesn't have MANTLE then that is when it'll die off.

 

[palladin9479]

Ok guys stop using GFLOP's when comparing CPU's as it's a very bad metric to use. Floating operations are actually a very small minority of the code that gets run on CPU's but it's highly scalable code. You can run that code on an integer unit but it's incredibly slow, instead it's better to run on a specialized vector co-processor. Otherwise known as a GPU. General purpose processors are instead measured by their capability to perform integer operations which compromise the vast majority of code. Integer OPs are not just 1+1 but logic compares and memory operations (get A, get B, if A > V jump D ect..). The difference in integer performance explains the huge ST difference between the AMD design and the current Intel design (2 ALU's per AMD unit vs 3 ALU's per Intel unit). Getting high integer performance requires a scalar design with multiple ALU's and the ability to rapidly predict, dispatch, execute and return the code along with a very good caching system to prevent stalling.

So yeah in the game of Integer operations, x86-64 is much stronger then ARM.

 

[griptwister]

It's a sign guys :lol:

http://www.tomshardware.com/news/bitfenix-phenom-m-mini-itx-micro-atx,24440.html

I really hope Stereoscopic 3d gaming is done... I mean, I like where the Oculus rift is headed, but 3d gaming on a panel, IMO, not so great.

 

[palladin9479]

It's a sign guys :lol:

http://www.tomshardware.com/news/bitfenix-phenom-m-mini-itx-micro-atx,24440.html

I really hope Stereoscopic 3d gaming is done... I mean, I like where the Oculus rift is headed, but 3d gaming on a panel, IMO, not so great.

I was doing stereoscopic gaming back in 2001 and I'm still doing stereoscopic gaming today in 2013. Its not "going away" anytime soon, modern 120hz panels work just find for it. The real difference is quality of glass's, you need expensive high quality glass's otherwise you'll get bad ghosting and improper image separation.

From my own experiences introducing people to it, the biggest problem is initial adjustment and calibration. You've spent your entire life using the stereoscopic vision built into your own brain, your convergence is based on your natural eye separation. When using artificial stereoscopic vision you will experience quite a bit of vertigo for awhile as your brain has to get used to different convergence and spacing. The break in time isn't 15m or even 60m, but actually several days of consistent use. After that the headaches go away and it becomes quite natural to use.

 

[palladin9479]

The Q6660 Inside
30 September 2013 02:09:41

Why don't people get sandy bridge is 80% better per core then piledriver. I find it funny that people think i5's are even compared to low end 4 core apus lol, when they are similar to amd flagship 8 core in performance. Gimp and superpi show this. Do you really think people would get a core i5 cpu if it was as slow as you think. Just remember intels 4 cores compete against amd 8 cores.

Seems your back to selling Intel parts again. We've been over this already. Per-core there is an approximately 40% advantage towards Intel at the same clocks. This is largely due to Intel having 3 ALU's per core vs AMD's 2 ALU's per core followed by Intel's superior prediction, scheduling and caching system. AMD mitigates some of this by clocking their CPUs higher and offering more total ALU's per die. When comparing CPU's you compare them on similar price vs performance not on absolute performance.

 

[palladin9479]

Where the hell are you getting 70~80% from? I'm seeing 40% aggregate at same clock and less depending on software.

You keep repeating the same dishonest cherry picked benchmarks to either sell more intel CPU's or stir up more sh!t.

 

[palladin9479]

Superpi is x87 which has largely been depreciated. AMD's BD design has really sh!tty x87 performance as it's emulated in the SIMD arrays.

The 4670 is a $240 USD CPU with 12 ALU's (4x3) while the fx8350 is a $200 USD CPU with 16 ALU's (8x2). The 4670 is 3.4Ghz while the 8350 is 4.0Ghz. The Intel CPU technically has 50% more raw integer power (3/2) yet due to code limitations you won't always be capable of using all three ALU's on the same thread. This is why the i7's tend to have higher SMP benchmarks then the i5's even through they are the exact same chip. Those extra set of HW registers allows the additional ALU to be used more often. Intel also has a much better instruction caching / prediction system, so any data set that can't fit inside L1/L2 cache will see a significant performance drop in AMD CPU's (L3 can help sometimes).

Now stop using Anandtech's benchmark table's as those are user submitted. You have no idea what MB, memory, HDD or background application's were being used. They average the results of every submitted mark.

Using site reviews benchmarks, like Toms you can easily see the ~40% difference in capability at the same clocks. I use "~" because the exact different can swing dramatically depending on the mix of integer / SIMD operations and how well threaded the application was. Handbrake are x264 are pretty much the industry standard as their fairly vender agnostic though care must be taken about compiler options, there is a world of difference between i386 and i686 as targets (x87 vs SSE).

 

[noob2222]

look unless every multithreaded software uses icc and somehow sabotages amd cpus then why do all the tests prove the same thing. Anyway amd sabotaged themself with modular design old memory controller and platform so unless game developers manage to make pc games physics work on very slow cores intel will have the lead. Even if steamroller was at 6ghz quad an i5 haswell at 3.5ghz would still beat it at games if you go by amds estimation. I would struggle to use such slow per core performance thats what netbooks or mobile phones are for. An fx8350 is like a core 2 cpu at 2.4ghz but with 8 cores.

If I am using dishonest benchmarks then explain how on handbrake an i5 wins against the fx8350 and with the little help of memory speeds and per clock performance.

Bit tech is an inel shill website just like you are, wich is why you love it soo much.

http://www.bit-tech.net/hardware/2013/06/01/intel-core-i7-4770k-cpu-review/4

know what the difference is there? Handbrake is open source, with means you can compile it however you want. One guess what compiler Bi(ntel)-tech used
The fact that Bit-Tech is nearly the only website out there that has the 8350 getting beat by ANY I5 cpu should speak for itself. go ahead, check them. https://www.google.com/#q=8350+handbrake+review

omg AMD is 50% of Intel.

How many times do you have to be told that Stupor PI IS OLD and obsolete?

 

[juanrga]

I think if you really believe lets say the kaveri 4 core steamroller cpu is level with an i5 2500k it would literally be a miracle and 50% better then what amd predict with steamroller if both at same clock speed.

Juangra is comparing an i5 2500k to a low end 4 core hypothetical kaveri cpu when that is probably on par wih an i3 haswell. I am not trying to sell intel parts all I am saying how way off the mark some people think i5's are so low end.

I am using the same French site that you used here. Assuming Steamroller will be a 20% faster than Piledriver, the benchmark quoted above shows that Kaveri 4C will be at the level of i5-3330, i5-2300, and i5-2500k. The difference with the i5-2500k in that French benchmark is of a 6%, which is within the margin of error. Their findings coincide very well with my findings.

 

[mayankleoboy1]

You keep repeating the same dishonest cherry picked benchmarks to either sell more intel CPU's or stir up more sh!t.

just like some others here keep doing the same for AMD.

 

[de5_Roy]

we have TWO forum threads for intel cpus, one for ivy bridge and another for the rest, opposed to just this one (afaik) for amd cpus. yet all the off-topic intel posts, that have nothing to do with steamroller, are being posted here. moderators?

 

[gamerk316]

we have TWO forum threads for intel cpus, one for ivy bridge and another for the rest, opposed to just this one (afaik) for amd cpus. yet all the off-topic intel posts, that have nothing to do with steamroller, are being posted here. moderators?

To be fair, it IS valid to compare AMD to Intel CPU's when discussing performance. As long as performance keeps up, there are going to be comparisons.

 

[de5_Roy]

we have TWO forum threads for intel cpus, one for ivy bridge and another for the rest, opposed to just this one (afaik) for amd cpus. yet all the off-topic intel posts, that have nothing to do with steamroller, are being posted here. moderators?

To be fair, it IS valid to compare AMD to Intel CPU's when discussing performance. As long as performance keeps up, there are going to be comparisons.

for the sake of speculation, if intel is being compared to steamroller, or even jaguar (on a really slow news day), it may be fine... but comparing sandy bridge and ivy bridge with piledriver that's 2011-2012 stuff...
intel will phase out most of sb and ivb by the time amd launches kaveri and the rest of steamroller products. intel will compete with haswell, ivy bridge-e, haswell-e and silvermont.

 

[8350rocks]

Name one relevant benchmark where an Intel in a similar price bracket is anywhere near 2x faster than an AMD anything...

You can't do it.

EDIT: There may be comparisons; but it shouldn't be at the level it has been, and they have been all too repetitive misrepresentations at that.

 

[palladin9479]

we have TWO forum threads for intel cpus, one for ivy bridge and another for the rest, opposed to just this one (afaik) for amd cpus. yet all the off-topic intel posts, that have nothing to do with steamroller, are being posted here. moderators?

To be fair, it IS valid to compare AMD to Intel CPU's when discussing performance. As long as performance keeps up, there are going to be comparisons.

Comparisons are fine as long as it's within the scope of the topic and retains some level of sanity.

 

[Cazalan]

The even number of CUs is interpreted by some as the software failing to report adequately a dual-graphics config. Others believe that one module of the supposed 3M version of Kaveri has been substituted by more iGPU and that the 13CU is a new top model Kaveri APU.

Anything is possible. 13CU may be odd but they've released odd numbers of CU before. The A6-5400K has 3CU (192 shaders). For yield reasons it may be a 14CU and they disable the worse performing one. As devices get more transistors they're relying on more die harvesting techniques. Even the Xenon (XB 360) had a spare CPU core.

 

[blackkstar]

look unless every multithreaded software uses icc and somehow sabotages amd cpus then why do all the tests prove the same thing. Anyway amd sabotaged themself with modular design old memory controller and platform so unless game developers manage to make pc games physics work on very slow cores intel will have the lead. Even if steamroller was at 6ghz quad an i5 haswell at 3.5ghz would still beat it at games if you go by amds estimation. I would struggle to use such slow per core performance thats what netbooks or mobile phones are for. An fx8350 is like a core 2 cpu at 2.4ghz but with 8 cores.

If I am using dishonest benchmarks then explain how on handbrake an i5 wins against the fx8350 and with the little help of memory speeds and per clock performance.

Bit tech is an inel shill website just like you are, wich is why you love it soo much.

http://www.bit-tech.net/hardware/2013/06/01/intel-core-i7-4770k-cpu-review/4

know what the difference is there? Handbrake is open source, with means you can compile it however you want. One guess what compiler Bi(ntel)-tech used
The fact that Bit-Tech is nearly the only website out there that has the 8350 getting beat by ANY I5 cpu should speak for itself. go ahead, check them. https://www.google.com/#q=8350+handbrake+review

omg AMD is 50% of Intel.

How many times do you have to be told that Stupor PI IS OLD and obsolete?

http://forums.anandtech.com/showthread.php?t=2297052

FX-8350 4GHz DDR3-1866 60.89s
i5 3570K @ 4.020Ghz 29.68s

Gimp a single thread image editing app twice as fast:
http://www.bit-tech.net/hardware/2013/06/01/intel-core-i7-4770k-cpu-review/4
Compare the fx8350 to lets say to an i5 3570k overall similar multithread but twice as fast single thread superpi:
http://www.notebookcheck.net/Mobile-Processors-Benchmarklist.2436.0.html

Now thats why intel ib 4 core at same clock speed perform the same as piledriver with all 8 cores at same clock speed. Do you really think intel would sell 4 core cpus if £90 fx6300 with no igpu was beating the i5's when the i5's are generally about 25% faster then them and similar to an fx8320 at stock and the higher end i5's similar to the fx8350 in most multithreaded stuff.

AT uses the user submissions, have you learned nothing? There is no baseline setup tested for those results...

Additionally...no one uses Gimp, or SuperPi anymore...

Look up adobe photoshop, or Corel studio, or Autodesk products, or some of the other billion programs that are at least from the last 2-3 years.

What you're doing is like me saying..."LOOK!!! This 1960s Computer handles FORTRAN and COBOL better than any new Intel machine ever could! New Intel machines can't even take the punch cards!!! How ridiculous is that?? This 1960s computer has like 1000x IPC over Intel running punch card instructions!!!!"

Gimp actually performs quite well when compiled with bdver2 CFLAGS for me.

Just because something is FOSS does not mean it is compiled well.

For instance, LAME is open source and I saw an over 60% speed up by compiling it properly for piledriver in Gentoo as opposed to the numbers I got in Windows with the official download..

Sometimes you find compiler optimizations where you least expect them. I did the crazy thing of custom compiling a java runtime environment and it didn't do a damn thing for performance as it seems to be jre I used was compiled well. That was a waste of a lot of time, lol.

 

[palladin9479]

Gimp actually performs quite well when compiled with bdver2 CFLAGS for me.

Just because something is FOSS does not mean it is compiled well.

For instance, LAME is open source and I saw an over 60% speed up by compiling it properly for piledriver in Gentoo as opposed to the numbers I got in Windows with the official download..

Sometimes you find compiler optimizations where you least expect them. I did the crazy thing of custom compiling a java runtime environment and it didn't do a damn thing for performance as it seems to be jre I used was compiled well. That was a waste of a lot of time, lol.

That is the big difference between i386 and i686 targets, before going into the more advanced compiler flags. i586 to i686 isn't nearly as big a jump though.

 

[jdwii]

The Q6660 Inside
30 September 2013 02:09:41

Why don't people get sandy bridge is 80% better per core then piledriver. I find it funny that people think i5's are even compared to low end 4 core apus lol, when they are similar to amd flagship 8 core in performance. Gimp and superpi show this. Do you really think people would get a core i5 cpu if it was as slow as you think. Just remember intels 4 cores compete against amd 8 cores. To put this into perspective my i5 3210m is faster then my q9300 at video encoding and my i5 4430 is over twice as fast at encoding compared to my q9300.

Sandy bridge is NOT 80% better per core their is no way that is anywhere close to being true SuperPI is also a outdated crappy piece of software The Intel Haswell is around 40% better per clock per core then piledriver, What you keep getting confused on is the 8 core is barley used look at programs that use it to its full potential not programs that are only work for very few cores.