AMD CPU speculation... and expert conjecture

[noob2222]

JF actually worked for AMD. He wad misled by the marketing team trying to say bd > 980x. The same marketing team that got fired by RR for hyping BD soo much that the only thing left was to be dissatisfied. I was hoping for more than we got with BD, but it wasnt a DUD like some others were saying.

But ya, AMD is promoting software over hardware right now, and I don't see that as a good thing from a hardware company. At least FX had the hardware first, kaveri is going hardware last.

 

[FALC0N]

Gamerk316, the big problem with bulldozer was AMD's bad execution. If they had done what they kept claiming they would do, "hold the line" on single thread performance, hit 4ghz at launch, and not gobble power like crazy, it would have been pretty close to the reasonable expectation of the time.

 

[juanrga]

I didn't participate in the BD thread and cannot comment on it. But the amount of insane BS, lies, and misinformation that is being posted at this thread is incredible. I recall perfectly as all predictions made by gamerk about i3s, BF4, MANTLE (well he says "mantel") were systematically disproved,* and how de5_Roy who cannot even read a slide and obtain the FPS from it, however decided to comment on the performance that the slide was representing. LOL.

* One remain: His prediction that MANTLE brings a 5% advantage over DX. But this will be also shown to be rubbish soon.

 

[FALC0N]

Noob2222, bulldozer was a dud. Maybe not in the long term, but the initial launch was awful. A die-shrunk Phenom would have outperformed it. Yes, they were quite serviceable, but these were FX branded parts, not OEM fodder.

 

[juanrga]

I don't know which is the more laughable prediction made in this thread up to now. That from the poster who said that Kaveri couldn't be clocked above 2.9Ghz, because was using bulk or noob222 prediction that Steamroller will be a 37% slower (IPC) than Bulldozer.

 

[noob2222]

@juan

Learn to read you tool

I said your stupid cosmology ES mystery chip is 37% slower. Your the one that keeps saying that its kaveri.

@falcon

I did say overhype is what killed BD.

Performance wise it was on par with phenom at launch. At times 10% slower (lame mp3 and itunes), and in others 20-30% faster. Most of the time it was 0-5% both ways.

Some people were expecting it to be 20-30% faster all the time with the overhype.

 

[FALC0N]

Even without the hype, it was disappointing. But point well taken. More realistic marketing would have helped a bunch.

 

[szatkus]

Szatkus, what games? Those that appear in the well-known "optimized for Intel" list that Intel publicities in its own site? Those that always appear benchmarked in sites known by their bias?

None of those games were optimized for Intel.

Maybe you mean those on purpose tested at low 720p at low settings?

Or are you trying to say me that at 1080p all games will run 50% better in a 2500k than in a A10-6800k? Because I know average difference can be lower than 10%.

This is how you measure CPU's gaming performance. It's the best way to reduce GPU bottleneck and helpful to pick more futureproof CPU.

Maybe you didn't notice, or maybe you jare just ignoring it, but it was already explained here which will be the expected gaming performance of Kaveri under (i) old games,

Yup, those quite old games.

(ii) modern games,

Will be similar difference for those two. Maybe FX-8 and i7 will perform relatively better (because of 8 threads), but we're talking about A10 and i5.

and (iii) MANTLE enabled games.

Mantle will without doubt reduce CPU load on A10. But also on i5+Radeon.

 

[Ags1]

Spoiler

Palladin, Yuka, Gamerk, 8350rocks and others on the thread who tried out my benchmark app, I have built a new version with an EXE launcher for windows. It lets me control the command line and ensure the app runs at high priority by default. Also, some of your CPUs were not identified correctly - this version (on windows of course) takes a second shot at the CPU name by reading the registry.

http://www.headline-benchmark.com/download.jsp

I'll give it a try some time tonight.

Wonder how my work PC would do, but I'd rather not feel the wrath of the IT/Security department...

The results are keyed by cpu name and no. of available cores, so I expect your results would just be entered under a separate key. Well, at least I hope so!

 

[jdwii]

...Am I the only one here getting a strong sense of deja-vu?

The [strike]Bulldozer[/strike] Kaveri rumor thread, with Gamerk2 playing himself, noob2222 playing himself, and juanrga playing the star role of JF-AMD.

Not JF-Amd he was still better more like brain i think is his name

 

[jdwii]

Maybe gamerk360, But at least there was good reason with bulldozer. A new, promising architecture, the return of the FX brand, the announced pricing, all pointed to a quality part that never materialized.

In this case, all signs point to a quality part, but there is no indicate that these parts will generate the 30% speed boost needed to catch the 2500k.

We might see a big IPC improvement, but the clocks are lower and that will offset the IPC gains somewhat.

Not to mention we are seeing the same old Amd again with no paper launches and their own specil benchmarks software. When Amd really has a winner like the 7970hd they push it old the door as fast as they can.

 

[jdwii]

Gamerk316, the big problem with bulldozer was AMD's bad execution. If they had done what they kept claiming they would do, "hold the line" on single thread performance, hit 4ghz at launch, and not gobble power like crazy, it would have been pretty close to the reasonable expectation of the time.

I didn't participate in the BD thread and cannot comment on it. But the amount of insane BS, lies, and misinformation that is being posted at this thread is incredible. I recall perfectly as all predictions made by gamerk about i3s, BF4, MANTLE (well he says "mantel") were systematically disproved,* and how de5_Roy who cannot even read a slide and obtain the FPS from it, however decided to comment on the performance that the slide was representing. LOL.

* One remain: His prediction that MANTLE brings a 5% advantage over DX. But this will be also shown to be rubbish soon.

I don't know which is the more laughable prediction made in this thread up to now. That from the poster who said that Kaveri couldn't be clocked above 2.9Ghz, because was using bulk or noob222 prediction that Steamroller will be a 37% slower (IPC) than Bulldozer.

I'm going to go with your prediction about the Amd Kaveri being equal to a I5 2500K

 

[-Fran-]

I'm going to go with your prediction about the Amd Kaveri being equal to a I5 2500K *

■ In HSA enabled software.

Cheers!

 

[jdwii]

I'm going to go with your prediction about the Amd Kaveri being equal to a I5 2500K *

■ In HSA enabled software.

Cheers!

Like i said its like saying i can get my car(Kaveri) to go to 0-60 in 4.2 seconds with heavy wind(HSA software) and the competitor car (I5) gets around 4.3 seconds(without HSA software)

 

[de5_Roy]

I didn't participate in the BD thread and cannot comment on it. But the amount of insane BS, lies, and misinformation that is being posted at this thread is incredible. I recall perfectly as all predictions made by gamerk about i3s, BF4, MANTLE (well he says "mantel") were systematically disproved,* and how de5_Roy who cannot even read a slide and obtain the FPS from it, however decided to comment on the performance that the slide was representing. LOL.

* One remain: His prediction that MANTLE brings a 5% advantage over DX. But this will be also shown to be rubbish soon.

lol. i clearly recall you avoiding posting your own calculations for weeks, instead ad hominem attacking me (i had already admitted several times that your given methods were not yielding reliable results for me), which resulted in me accusing you of lying (with accompanying explanation why). the reason i didn't report you to the mods was because i wanted to compare your data (and because the posts were right on the thread, for all to see). i have also called out your lies on several other occasions, enough proof that you cannot be remotely believed, regardless of the end result.

 

[jdwii]

"instead ad hominem attacking me"
He has nothing left then except i didn't say that lol

 

[FALC0N]

I think the HSA argument is valid one. AMD has worked hard creating the idea, the chips to support it, and cultivating support for it among the industry. If they succeed, they deserve credit for it. IF they are successful.

 

[noob2222]

^^ They deserve credit, but the problem may lie with the other innovations AMD came up with. Intel will wait till its adopted then exercise the cross license agreement and adopt HSA later. Let AMD front the bill getting it all started.

If that happens, nothing changes. Intel will have the faster cpu because AMD spent all their efforts on the software. Intel won't just let AMD run with the ball without doing anything to stop them, one way or another.

 

[juanrga]

Szatkus, what games? Those that appear in the well-known "optimized for Intel" list that Intel publicities in its own site? Those that always appear benchmarked in sites known by their bias?

None of those games were optimized for Intel.

They are.

Maybe you mean those on purpose tested at low 720p at low settings?

Or are you trying to say me that at 1080p all games will run 50% better in a 2500k than in a A10-6800k? Because I know average difference can be lower than 10%.

This is how you measure CPU's gaming performance. It's the best way to reduce GPU bottleneck and helpful to pick more futureproof CPU.

Nope, If I want play games at 1080p at medium settings testing CPUs at low settings and 720p is misleading. Moreover, as commented before, the tendency in games will be towards offloading the CPU generating GPU-bounded games.

Maybe you didn't notice, or maybe you jare just ignoring it, but it was already explained here which will be the expected gaming performance of Kaveri under (i) old games,

Yup, those quite old games.

I already explained what criterion I use to label games as old or new.

(ii) modern games,

Will be similar difference for those two. Maybe FX-8 and i7 will perform relatively better (because of 8 threads), but we're talking about A10 and i5.

Modern games have something more than just more threads. This was explained before as well.

and (iii) MANTLE enabled games.

Mantle will without doubt reduce CPU load on A10. But also on i5+Radeon.

This was also explained before. MANTLE not only reduces CPU load, but reduces the gap between different CPUs.

 

[FALC0N]

^^ They deserve credit, but the problem may lie with the other innovations AMD came up with. Intel will wait till its adopted then exercise the cross license agreement and adopt HSA later. Let AMD front the bill getting it all started.

If that happens, nothing changes. Intel will have the faster cpu because AMD spent all their efforts on the software. Intel won't just let AMD run with the ball without doing anything to stop them, one way or another.

Intel doesn't do it because its not in their best interest to do it. AMD has the more powerful GPU at this time. Moving the GPU into the realm of general processing benefits AMD and hurts intel. Why would intel show any interest in that?

And it may not be quite as simple as intel adopting HSA later. Its not just GPU on die. Its a tight integration between the two that allows programs and programmers to access them seamlessly. AMD has been working towards this for a long time and its not clear how quickly intel could move in this direction, if they chose to.

Of course, all this is predicated on AMD's vision for HSA succeeding.

 

[juanrga]

Juan direct quote from your bsn article. Kaveri scores for the CPU are obtained from taking Trinity/Richland scores as base and utilizing the assumed 20% gain from doubling the decoder per module minus a correction factor of the 5%.

True, but it doesn't mean that I added 20% to the scores. I assumed that Steamroller is 20% faster than Piledriver and then considered the frequencies. From the article: "Combining all this data [2], I predict that the CPU of the top Kaveri APU will be about 26% faster than top Trinity APU and about 17% faster than top Richland APU."

I can argue that doubling the decoder does not necessarily mean 2x performance as you have done in your calculations, But your 5% correction factor should be close. Especially if ipc has increased as i expect it has.

See above. Dual decoder: --> ~20% gain. I don't know from where you got 2x. Not from my article of course.

My concern is throttling, trinity/richland run hot as it is, how will kaveri cope.

I am said that Kaveri is quieter than Trinity/Richland. We will see.

Also just to elaborate on my comment about you being proven right, a88 chip set alone is yielding a 5% improvement in my own testing with fm2+ extreme 6 and that is with trinity. So i can see your 17% being very close.

Yes, you are right; there are lots of factors to consider here, specially when comparing Kaveri to Intel:

Dual decoder --> ~20%

There are other hardware improvements: from additional changes in the front-end to bigger L1 cache or improved L2 cache. Some leaked benchmarks suggest that they account for another 10% gain.

Kaveri comes with 2400MHz support. Most reviews of Trinity chips used 1600MHz (when the default is 1866MHz) and this favored Intel chips. Using the memory of 2400MHz must add another 5% to pure CPU workloads.

W8.1 increases performance of AMD APUs up to 10% compared to W8. Moreover, W8 increased performance of AMD by 5-10% compared to W7. This is due to scheduler optimized for CMT architecture. Most reviews of Trinity chips used W7 (and its bad scheduler) and this favored Intel chips. Using W8.1, an AMD APU is between 5-15% faster.

Intel ICC favors Intel chips. Benchmarks and software compiled with ICC (Sysmark, Cinebench...) use the Cripple_AMD function and force non-Intel chips to run slower code. I am said that AMD has convinced software houses to compile the next version of their software with a different compiler. We will see new versions of some popular software to run better in AMD hardware soon. We could see 5-20% gains.

 

[juanrga]

jdwii, I didn't say that Kaveri is equal to i5-2500k.

de5_Roy, you started to post insults and lies after it was shown that you are unable to perform a 10-year old calculation. I have ignored your posts lately until that you launched another ad hominem against me one page ago.

 

[juanrga]

^^ They deserve credit, but the problem may lie with the other innovations AMD came up with. Intel will wait till its adopted then exercise the cross license agreement and adopt HSA later. Let AMD front the bill getting it all started.

If that happens, nothing changes. Intel will have the faster cpu because AMD spent all their efforts on the software. Intel won't just let AMD run with the ball without doing anything to stop them, one way or another.

First, AMD, Nvidia, and Intel develop both software and hardware. You only attack AMD by developing software.

Second, your claim that "AMD spent all their efforts on the software" is plain wrong. With Kaveri AMD is pushing at the same time a new architecture and new node. Intel only makes one thing at the same time. The hardware improvements in Kaveri are too numerous to be mentioned all here, but include a mayor revision of the Piledriver architecture, from the front end to the L2 cache almost everything, including L1 cache, FPU, and IMC are improved.

Besides that, AMD introduces HSA hardware improvements, such as hUMA (unified memory controller for both CPU and GPU) or HSAIL ISA in the CPU. Add the ARM core for security purposes and also the iGPU, which has been changed entirely, with the old VLIW4 architecture replaced by a new GCN architecture as that found in the new R9 cards.

Third, Intel and Nvidia are also developing their own HSA-like approaches. Intel makes this with Xeon+Xeon-Phi. Phi is based in new AVX extensions, which Intel is promoting heavily. The main change in Haswell CPU was the update in AVX. Using ordinary software, Haswell performs the same than Ivy Bridge, using software compiled for the new AVX2 extensions Haswell run much faster. Intel has collaborated with software houses to develop AVX software. For instance, some review sites reviewed Haswell chips using new H264 software optimized for AVX2. This is a quote from the tomshardware review of Haswell:

Our AVX2-based results from Core i7-4770K almost match the preview piece’s exactly, while the AVX-accelerated Ivy and Sandy Bridge numbers are close too. We now see that AVX2 helps a four-core Haswell part outperform Sandy Bridge-E’s six cores in AVX-optimized code.

Therefore if AMD develops HSA and HSA-enabled software you make here a show, but then Intel makes the same with its own neo-heterogeneous approach, you remain silent. Funny double standard.

 

[pangolin_user]

http://wccftech.com/amd-working-hynix-development-highbandwidth-3d-stacked-memory/
amd apu is on roll.. they know what they need on how to do it right.

 

[noob2222]

^^Intel's success doesn't depend on hsa being successful. All of intel's software could fail and they would still make 4B per year.

AMD on the other hand is pushing lower end hardware in favor of HSA. Thats the problem. If AMD software fails, they have only a low end processor and a mid level gpu.