AMD CPU speculation... and expert conjecture

[jed]

CPU test of headline has some bugs to work out. HTT is mucking with the CPU results, like I thought it was.

Look at it logically: If I have 4 cores, and I do the same amount of work in 4 threads, you would expect linear absolute performance gains up to 4 threads, correct?

Taking a look at two 2600(k) runs:

http://www.headline-benchmark.com/results/c309c1f4-6f00-4def-8127-b8d679fe5989
http://www.headline-benchmark.com/results/93a871e9-4a76-42aa-8f91-ef08fedef7dc

Notice the various dropoffs in performance as threads increase? That shouldn't be happening, unless multiple threads are getting stuck on the same core, OR a HTT core is getting used before a "full" core, and the shared frontend is causing massive performance drops.

Compare that to this run:

http://www.headline-benchmark.com/results/359025ad-a71c-4df1-a21e-2dcac265a956/87080cac-179d-4a2c-a2a9-cb1e90667a3f

Where I disabled the HTT cores, and compared against my last benchmarks. Note how the performance dropoffs vanished, and in some cases, performance improved by DISABLING cores?

Now, we still have an issue where performance stalls. What I *think* is going on, is multiple threads are going to the same core, so performance flatlines, then jumps by double the next time a thread gets assigned, maxing at about 5 threads (remember: HTT cores disabled). So in the case of the Integer test, I think the treads are being assigned like this:

Thread 1: Core 0
Thread 2: Core 1
Thread 3: Core 0/1
Thread 4: Core 2
Thread 5: Core 3

Hence why performance stalls on three threads, then increases linearly up to 5 (which doesn't make sense for a quad). Same thing likely happening in the FP test:

Thread 1: Core 0
Thread 2: Core 0
Thread 3: Core 1
Thread 4: Core 1
Thread 5: Core 2

Based on the huge jump going from 4 to 5, I'd suspect one of the other threads gets reassigned to core 3 at some point, but I'd need to break out my low level debug tools to prove this.

Ags1, I'd recommend that you manually assign threads to cores for this test. Start with the Even (Non HTT/CMT) cores, then proceed to the Odd ones. Once you have more threads then cores, don't bother anymore (really doesn't matter at that point). That should remove the Windows scheduler from affecting performance...Right now, due to HTT having a much higher performance cost if used wrong, Intel's numbers are probably grossly understated in your benchmark. [Can't wait to see the response to THAT one.]

[Can't wait to see the response to THAT one.] +1

 

[jed]

Jed, I find those bit-tech results a bit hard to believe. It's all very well to say they did a 'real-world' bench using Gimp, but which parts of Gimp? Seeing how the i3 places relative to the i5 I would say the test is single threaded. They also mention the test results rely on memory and storage performance, which means they are not doing a pure test of the CPU. Which is why I prefer synthetic benchmarks, and of course wrote one myself.

The handbrake result also looked a bit off to me, and I recall Tom's also used handbrake in their FX8350 review, and got totally different results:

http://www.tomshardware.com/reviews/fx-8350-vishera-review,3328-12.html

I notice every thing you post has Intel at a clock disadvantage of about 600MHz or more,
I used the 2500k and 3570k because they cost about the same as the 8350, with just 4
cores and no HT, now all three cpu's top out at 4.8 - 5.0GHz and here on this forum most
of us overclock everything we get, so with that being said no more of the stock clocks for
the Intel cpu, get some benchmark with both overclocked so we can inform people on just
how good these chips perform against one another.

Agreed, but currently all my FX results are OC'd and most of the Intel ones aren't. Looks like a different kind of user buys an FX and they are highly likely to OC. I will post an OC'd 3570 if I ever get one, or a stock 8350, if I ever get one.

Let's look at it this way you gave us a benchmark to perform and some of us ran it,
and post the results.
Now the Intel CPU's from the 2600k to the 4770k sit at the top of your benchmark list,
and the only one you compare the 8350 to are the one's under it, never the one's above.
So the question is why would you continuously do this?

Here's where other's can see what I'm talking about.
http://www.headline-benchmark.com/

 

[etayorius]

Jed, I find those bit-tech results a bit hard to believe. It's all very well to say they did a 'real-world' bench using Gimp, but which parts of Gimp? Seeing how the i3 places relative to the i5 I would say the test is single threaded. They also mention the test results rely on memory and storage performance, which means they are not doing a pure test of the CPU. Which is why I prefer synthetic benchmarks, and of course wrote one myself.

The handbrake result also looked a bit off to me, and I recall Tom's also used handbrake in their FX8350 review, and got totally different results:

http://www.tomshardware.com/reviews/fx-8350-vishera-review,3328-12.html

I notice every thing you post has Intel at a clock disadvantage of about 600MHz or more,
I used the 2500k and 3570k because they cost about the same as the 8350, with just 4
cores and no HT, now all three cpu's top out at 4.8 - 5.0GHz and here on this forum most
of us overclock everything we get, so with that being said no more of the stock clocks for
the Intel cpu, get some benchmark with both overclocked so we can inform people on just
how good these chips perform against one another.

Agreed, but currently all my FX results are OC'd and most of the Intel ones aren't. Looks like a different kind of user buys an FX and they are highly likely to OC. I will post an OC'd 3570 if I ever get one, or a stock 8350, if I ever get one.

Let's look at it this way you gave us a benchmark to perform and some of us ran it,
and post the results.
Now the Intel CPU's from the 2600k to the 4770k sit at the top of your benchmark list,
and the only one you compare the 8350 to are the one's under it, never the one's above.
So the question is why would you continuously do this?

Here's where other's can see what I'm talking about.
http://www.headline-benchmark.com/

Or better yet... what does this has to do with Steamroller?

 

[jed]

now all three cpu's top out at 4.8 - 5.0GHz
No way
8350fx 4.6-4.8 some around 5.1(water or at least should be water)
I7 3770K and If 3570K 4.4Ghz
Air cooling

First of all the i7 3770k wasn't mentioned in the above paragraph not once,
in my experience the 2500k would overclock a little better then the 2600k I
owned, which by the way ran @ 4.7 with air cooling and @ 5.2 with water.
The three cpu's I used was the 2500k, 3570k and 8350 with all of them
using aftermarket cooling, weather it be air or water cooling.

I'm not saying that Haswell and IB can't hit higher clocks, however...

Intels' CPUs have been gradually growing less likely to hit higher clocks. You can find Haswells close to 5ghz if you dig across the internet, and then compare that to SB and then say that they overclock the same.

But if you have 1000 Sandy Bridges, 1000 Ivy Bridges, and 1000 Haswells, you're going to have a lot more Haswells that don't overclock very well compared to how many good clocking Sandy Bridges you'll see.

That is the point we're getting at. I don't think I've ever seen a Piledriver chip that wouldn't go past 4.4ghz no matter how much voltage you threw at it (temps are a different matter, but temps can be fixed while there's nothing you can do about a chip that just wants vcore). You do find Haswells that won't go 4.4ghz and you even find IBs that don't OC very well at all.

Unless I'm mistaken, you're pretty much guaranteed 4.7ghz with a piledriver chip if you have good enough cooling, and if you're on water you're pretty much guaranteed 5ghz. The golden FX 8350s are the ones that hjit 5ghz+ on air, and the golden IBs and Haswells hit high 4ghz range.

I'm quite sure you can find people running Haswell and IB at higher vcore claiming to go fast, but the chips don't last. I knew an Intel fanboy on OCN who had a 5.2ghz delidded SB he ran 1.55v through. He boasted about it for a year and called it the Vishera killer, and then one day he put the whole thing up for sale, parted out, and then downgraded from what he had. In other words, he probably degraded the chip and went to unload it on some poor, unsuspecting OCN member.

So guess I just have very good luck with Intel chips then right?

http://valid.canardpc.com/95b7ds

 

[de5_Roy]

Nvidia holds CES press event at AMD Kaveri hotel
http://www.fudzilla.com/home/item/33560-nvidia-holds-ces-press-event-at-amd-kaveri-hotel
it's on now...

 

[Rum]

If this is true then I wonder what a fully unlocked Card would do against the 780TI?
http://wccftech.com/amd-hawaii-gpu-die-shot-analyzed-fully-unlocked-chip-48-compute-units-3072-stream-processors/

 

[-Fran-]

Nvidia holds CES press event at AMD Kaveri hotel
http://www.fudzilla.com/home/item/33560-nvidia-holds-ces-press-event-at-amd-kaveri-hotel
it's on now...

Well, its a sign that nVidia thinks AMD has something good in their hands with Kaveri, so they won't give them the benefit of getting ahead of them PR wise. AMD does the same thing to Intel and nVidia, so no surprises here.

The weird thing, is the hotel coincidence, haha.

If this is true then I wonder what a fully unlocked Card would do against the 780TI?
http://wccftech.com/amd-hawaii-gpu-die-shot-analyzed-fully-unlocked-chip-48-compute-units-3072-stream-processors/

We might see cherry picked ones later on, since we always see cherry picked ones later, haha. Maybe Sapphire or Asus will get them. Or maybe they're already creating "special edition" ones. It's been done before, so it wouldn't surprise me at all.

Cheers!

 

[etayorius]

nVidia will probably reveal the first details of Maxwell, but i am more interested in Kaveri than Maxwell "Details"

 

[Cazalan]

If this is true then I wonder what a fully unlocked Card would do against the 780TI?
http://wccftech.com/amd-hawaii-gpu-die-shot-analyzed-fully-unlocked-chip-48-compute-units-3072-stream-processors/

We might see cherry picked ones later on, since we always see cherry picked ones later, haha. Maybe Sapphire or Asus will get them. Or maybe they're already creating "special edition" ones. It's been done before, so it wouldn't surprise me at all.

Cheers!

Or they're reserved for the Workstation class video cards. They didn't update the FirePro series yet right?

 

[Cazalan]

nVidia will probably reveal the first details of Maxwell, but i am more interested in Kaveri than Maxwell "Details"

At this point I'm more interested in Maxwell as Kaveri has been beat to death the past year. Assuming Maxwell is 20nm, otherwise meh.

 

[juanrga]

Here goes the update

http://juanrga.com/en/AMD-kaveri-benchmark.html

Now it is time to wait and see

 

[Cazalan]

Here goes the update

http://juanrga.com/en/AMD-kaveri-benchmark.html

Now it is time to wait and see

Why all i5's and no i3's? i3 is it's direct competition.

 

[-Fran-]

Here goes the update

http://juanrga.com/en/AMD-kaveri-benchmark.html

Now it is time to wait and see

One correction. At the end, the RAM debacle, you cannot conclude Kaveri uses slower RAM because the 4GB are reporting "1Mhz". That's just un-recognized. It could be 2133Mhz or 2400Mhz for all we know, but its less (if that isn't badly reported either) than the 8GB for the other one.

Cheers!

 

[de5_Roy]

One correction. At the end, the RAM debacle, you cannot conclude Kaveri uses slower RAM because the 4GB are reporting "1Mhz". That's just un-recognized. It could be 2133Mhz or 2400Mhz for all we know, but its less (if that isn't badly reported either) than the 8GB for the other one.

Cheers!

afaik, amd a10 7850k should run 1.5v ddr3 2133 ram by default. higher speed ram might need setting a.m.p.

 

[etayorius]

Here goes the update

http://juanrga.com/en/AMD-kaveri-benchmark.html

Now it is time to wait and see

Why all i5's and no i3's? i3 is it's direct competition.

Why the hell would the i3 be direct competition? the i3 is a Dual Core with Hyper Threading, Kaveri is a Full 4 Cores and their cores are much stronger than a Phenom/FX Core... and the GPU inside Kaveri is at least 4 times Stronger to any i3 GPU, unless you`re comparing Price point and even so... the i3 looks like Smelly crap compared to Kaveri.

Kaveri seems to be direct competition to the older i5 models.

Even if Kaveri is only 10% faster than Richland, it would still be a heck ton of a better deal that the fastest i3, just NO.

Also, Kaveri using it`s 4 cores will probably laugh at any i3 in Multi Threaded Apps... i3 Single Core performance may be crazy, but it`s Multi Threaded is just Laughable.


http://www.anandtech.com/bench/product/675?vs=677

Even the lowly 5800k pulls ahead against the i3 3220 in the Intel Biased Cinebench in Multi Threading.

 

[Ags1]

CPU test of headline has some bugs to work out. HTT is mucking with the CPU results, like I thought it was.

Look at it logically: If I have 4 cores, and I do the same amount of work in 4 threads, you would expect linear absolute performance gains up to 4 threads, correct?

Taking a look at two 2600(k) runs:

http://www.headline-benchmark.com/results/c309c1f4-6f00-4def-8127-b8d679fe5989
http://www.headline-benchmark.com/results/93a871e9-4a76-42aa-8f91-ef08fedef7dc

Notice the various dropoffs in performance as threads increase? That shouldn't be happening, unless multiple threads are getting stuck on the same core, OR a HTT core is getting used before a "full" core, and the shared frontend is causing massive performance drops.

Compare that to this run:

http://www.headline-benchmark.com/results/359025ad-a71c-4df1-a21e-2dcac265a956/87080cac-179d-4a2c-a2a9-cb1e90667a3f

Where I disabled the HTT cores, and compared against my last benchmarks. Note how the performance dropoffs vanished, and in some cases, performance improved by DISABLING cores?

Now, we still have an issue where performance stalls. What I *think* is going on, is multiple threads are going to the same core, so performance flatlines, then jumps by double the next time a thread gets assigned, maxing at about 5 threads (remember: HTT cores disabled). So in the case of the Integer test, I think the treads are being assigned like this:

Thread 1: Core 0
Thread 2: Core 1
Thread 3: Core 0/1
Thread 4: Core 2
Thread 5: Core 3

Hence why performance stalls on three threads, then increases linearly up to 5 (which doesn't make sense for a quad). Same thing likely happening in the FP test:

Thread 1: Core 0
Thread 2: Core 0
Thread 3: Core 1
Thread 4: Core 1
Thread 5: Core 2

Based on the huge jump going from 4 to 5, I'd suspect one of the other threads gets reassigned to core 3 at some point, but I'd need to break out my low level debug tools to prove this.

Ags1, I'd recommend that you manually assign threads to cores for this test. Start with the Even (Non HTT/CMT) cores, then proceed to the Odd ones. Once you have more threads then cores, don't bother anymore (really doesn't matter at that point). That should remove the Windows scheduler from affecting performance...Right now, due to HTT having a much higher performance cost if used wrong, Intel's numbers are probably grossly understated in your benchmark. [Can't wait to see the response to THAT one.]

[Can't wait to see the response to THAT one.] +1

It's a limitation of Java that I can't assign threads to specific cores. I have to rely on the default OS scheduling. Trust me, I've tried. Still, it's interesting that Microsoft gets it pretty wrong. And relevant to the 90% of users stuck with the Windows scheduler.

The issue does not affect any scores however, as I only count the single-threaded and maximum performance data points. I used to include intermediate threading results, but stopped that and recalculated everything over a month ago after feedback from Yuka and gamerk.

You are however seeing what I think is a real effect, where the assymetric complexity of hyperthreading is difficult for the OS to schedule optimally, whereas the FX chip presents 8 roughly equivalent cores, which is easier to schedule. I.e. This may be a real advantage of the FX architecture that you can't see simply by looking at max theoeretical performance.

Regarding the number of threads, I go to logical processors + 4 because on Windows this squeezes a bit more performance from the CPU. I think what happens is the extra threads squeeze the CPU time of background processes. On linux, performance tends to peak at threads = cores.

Jed, Regarding any bias in my postings, the results come in at different times and not all results are available at the same time. But then, I have an actual website that shows to the whole world that Intel is on top, so what are you complaining about? Also, all CPUs on the site are shown compared to the top (i.e. Intel) CPU. How is that biased to AMD? Also, this an AMD thread, so I only tend to post AMD-relevant results.

 

[anxiousinfusion]

A Chinese site has posted benchmarks comparing Kaveri 7850K to Richland 6800K. And the integrated graphics gains, according to their charts, are actually pretty awful. Often beating Richland by only 1-4 FPS.

http://tieba.baidu.com/p/2795083406

EDIT: Kaveri 7800, rather a 65W part.

 

[etayorius]

A Chinese site has posted benchmarks comparing Kaveri 7850K to Richland 6800K. And the integrated graphics gains, according to their charts, are actually pretty awful. Often beating Richland by only 1-4 FPS.

http://tieba.baidu.com/p/2795083406

If that`s Kaveri performance then it looks like crap, good thing it`s 2 more weeks to release.

I dont understand the benchmarks completely, but Kaveri iGPU shouldnt be way faster in 3D Mark 11 performance? it seems in those benchmarks Kaveri 512 GCN GPU barely did better than Richland 384 WLVI4 GPU... quite odd.

 

[juanrga]

Yuka Ok, but look at bandwidth.

de5_Roy But the benchmarked sample is not a stock A10-7850k, but a 'capped' sample with lower base frequency, lower turbo, lower...

anxiousinfusion, they are comparing the 65W Kaveri against 100W Richland. Also they are not using final drivers and they had throttling issues.

The CPU benchmarks are more interesting. The CB R11.5 score shows a 12% gain. The wPrime score shows a 30% gain. And the Sisoft Sandra where Kaveri is 5x faster than Richland is showing HSA in action ;-)

 

[Master-flaw]

A Chinese site has posted benchmarks comparing Kaveri 7850K to Richland 6800K. And the integrated graphics gains, according to their charts, are actually pretty awful. Often beating Richland by only 1-4 FPS.

http://tieba.baidu.com/p/2795083406

Be ridiculous if they upgraded there APU's to only have a stronger CPU. But I'm waiting for official benchmarkswhen the drivers for these things are released.

 

[Master-flaw]

poor amd. they still believe that casual gamers are people who play triple A titles at crap-to-low settings... casual gamers play with their tablets and their smarthphones and these silly apus are completely useless.

my celeron SB ($30) + gts 450 ($60) , beats any AMD apu setup out there.

Keep telling yourself that,

Huge market for people who want a cheap OEM with the ability to play a game or two.

 

[Cazalan]

Why the hell would the i3 be direct competition? the i3 is a Dual Core with Hyper Threading, Kaveri is a Full 4 Cores and their cores are much stronger than a Phenom/FX Core... and the GPU inside Kaveri is at least 4 times Stronger to any i3 GPU, unless you`re comparing Price point and even so... the i3 looks like Smelly crap compared to Kaveri.

Kaveri seems to be direct competition to the older i5 models.

Even if Kaveri is only 10% faster than Richland, it would still be a heck ton of a better deal that the fastest i3, just NO.

Also, Kaveri using it`s 4 cores will probably laugh at any i3 in Multi Threaded Apps... i3 Single Core performance may be crazy, but it`s Multi Threaded is just Laughable.


http://www.anandtech.com/bench/product/675?vs=677

Even the lowly 5800k pulls ahead against the i3 3220 in the Intel Biased Cinebench in Multi Threading.

Competition is based on price of course. You don't compare luxury cars vs economy cars do you? You also don't compare much older models vs new. It's rather irrelevant as the supply of i5-2500k has dropped to used parts only.

The Haswell i3 has 8 ALU (2 x 4 ALU), which is the same as Kaveri (4 x 2 ALU). Haswell cores have much wider pipes. Graphics is of course the AMD advantage.

i3-4340 is $149, Kaveri 7850K is rumored $159. That is it's direct competition.

 

[Master-flaw]

"huge" yet pc sales are declining.

between playing games at low to medium on PC, i'd stick with an xbox360 instead (case in point: skyrim). see, the problem with these APUS is that it's fixed per segment, you get the lowest CPU+GPU for the cheapest, fastest CPU+gpu for the most expensive one.

give me a 4 to 5 ghz single module / 384 (gpu)core APU chip at half the price of the 6800k and you probably have a winner. design a mother board that doesnt sport a pcie slot make it BGA but sell the board at a massive discount make it micro ITX and you have winner. put it all together sell it at $120 (call this a5-6xxxk)), and that makes sense. this way you can have a disposable steambox that'll last 2 years to play previous gen steam games. bank on mantle and it'll probably last 3.

another cool idea is to design a single module / 768 gpu core chip APU chip. it doesn't have to be top of the line x86 core + top of the line gpu core.

i just don't get it with AMD. i just don't.

A lot of people don't want a 360...they want to play Dota or Starcraft after checking their Email...There is a market for it whether it's declining or not...yes portability's taking over, but these chips are also in laptops.

Also, It probably costs the same to produce and ship a single module APU as a quad-core. I'm sure AMD would be all over that market if it provided a significant value for development. but keeping specs as high as they can while keeping costs level is the best way to develop these things.

 

[Embra]

http://www.dailyfinance.com/2014/01/03/can-amd-win-the-high-end/

 

[wh3resmycar]

A lot of people don't want a 360...they want to play Dota or Starcraft after checking their Email.

yes, but i'm willing to bet an arm and a leg that those people are not looking forward to p@wn people while keeping visual settings low-medium. i've been playing dota 2 for a long time now and playing it @ 720p or lower at fluctuating FPS really hurts the experience. and that 384 GCN cores will not be enough to satisfy any competitive multiplayer gamer out there. my 7750 ddr3 which is twice as fast didn't well do well enough.