AMD CPU speculation... and expert conjecture

[juanrga]

No. The laws of physics says otherwise. That is the reason that both Nvidia and AMD engineers are designing their future supercomputers around APUs, and not around dCPU+dGPU anymore. Nvidia 40TFLOP design is scheduled for ~2018.

I really want to see what you say the physics are saying, because afaik: More transistors = More heat = less performance | Bigger diesize = fewer yields = higher cost

Also the problem I see with seperating the chip into 2 dies, is again afaik: space

Though there are most certainly more problems/examples than that I just used what I know so.. 😀

Several years ago, studies were made on how to design a supercomputer that would be 1000 times more powerful than actual ones. The research revealed that current engineering paradigms don't work at the exascale level. The findings were summarized in a DARPA report. One of the challenges identified in the report was that at the exascale level the principle of locality holds. A traditional design dCPU+dGPU violates the principle of locality. That is the reason why the engineers of Nvidia and of AMD have selected APUs for their respective exascale supercompute designs. A dCPU+dGPU doesn't work at that scale of performance.

I believe it is time that some people at this forum stop repeating wrong mantras such as "APUs are for cheap gaming".

 

[etayorius]

Let me make it clear, i am not saying Kaveri is crap... it is not, the GPU inside Kaveri and other features like HUMA, HSA, MANTLE and OpenCL will be massively boosted by Kaveri, i am talking about the laughable numbers Kaveri is getting compared to Richland and Trinity in regards of IPC.

As i already stated regarding Kaveri´s laughable Skyrim performance improvement against Trinity in the Spanish Forum and the Pudget review, it makes me believe Kaveri has IPC between Trinity and Richland... is this bad news? not necessarily... Kaveri is a damn great upgrade to anyone who is on a budget or who is not a hardcore gamer, like one person asked if Kaveri should be worthy upgrade compared to Llano, HECK IT IS you bet!

It is just not worth the upgrade for me (PhenomII 980 at 3.9Ghz) or anyone who has an FX or i5 CPU paired with a dGPU or greater, doing so will only be a downgrade.

Until MANTLE and HSA are well established, Kaveri will be aimed at the low-mid range budget gaming.

It`s just how it is.

But heck, we are almost hitting release Date... and if i am wrong, ill be happy to take my words back and hurry to the store and buy one, but i will not accept anything Less than 15% improvements against RICHLAND as a possible Invest in Kaveri, if Kaveri does not even manages a 15% improvement on average against Richland, there is no way i will even consider Kaveri as upgrade... hell NO.

On the contrary, the GPU inside Kaveri is really good in my opinion and MANTLE will push the performance even further, that´s heck of a Win for AMD one way or the other.

 

[palladin9479]

No. The laws of physics says otherwise. That is the reason that both Nvidia and AMD engineers are designing their future supercomputers around APUs, and not around dCPU+dGPU anymore. Nvidia 40TFLOP design is scheduled for ~2018.

I really want to see what you say the physics are saying, because afaik: More transistors = More heat = less performance | Bigger diesize = fewer yields = higher cost

Also the problem I see with seperating the chip into 2 dies, is again afaik: space

Though there are most certainly more problems/examples than that I just used what I know so.. 😀

It's mostly about heat vs thermal solutions required to cool them. A single CPU with a 125W TDP combined with a separate GPU with 250W TDP is always, 10 times out of 10, going to be higher performance then a single 125W CPU + GPU combo chip. And that's before going into the die space issue vs cost's. Juan has been blowing smoke for a long time now for no matter what they put in an APU they can always put more/ faster of the same in a dCPU + dGPU. The laws of thermal dynamics are what limits Intel / AMD here.

EX:
Richland / Kavari vs
8350 / i7 + R9 290

In every conceivable performance profile the second combo wins as whatever vector work the iGPU can do, the dGPU can do more of and faster. HSA doesn't change any of that, it just allows the iGPU to be used for something while a dGPU does something else so your not wasting 50% of your CPU's die space.

 

[esrever]

I don't know why you guys are basing everything off benchmarks of cinebench 15 which looks worse than ever as a benchmark. The reviews aren't even out yet. If you look at 11.5, the performance is up 11% at stock performance and IPC is up at least 12% in cinebench.

In a different benchmark, its up 16% at stock and IPC increase is almost 18%

 

[-Fran-]

^ Give up your 3850 ? Only if kaveri makes a big improvement on the cpu side. Not seeing it yet, did you notice the spanish site used a closed loop for testing ? I am starting to think all the variations in benchmarks can be attributed to thermal throttling

Not really. For an HTPC you don't really need that much CPU horsepower. In particular, Just for software decoding duties, you would need lots of CPU grunt work, which I'm sure Kaveri will do just fine where Llano can't (new instructions sets + HSA promise). What you really need is decoding beauty (nVidia has the upper hand with CUDA, regrettably) and a great MoBo (sound card in particular).

Games is just a plus on the side IMO and I'm sure in that department Kaveri will be a better ride than Llano.

Cheers!

 

[jdwii]

Spoiler

APU's will never compete with dGPU / dCPU's due to cost and space / heat limits, physics is a hard master.

No. The laws of physics says otherwise. That is the reason that both Nvidia and AMD engineers are designing their future supercomputers around APUs, and not around dCPU+dGPU anymore. Nvidia 40TFLOP design is scheduled for ~2018.

A hand-made review of Kaveri

http://foro.noticias3d.com/vbulletin/showthread.php?t=421821

work in progress...

Hmm horrible CPU performance, ¡t looks like in the end Kaveri is not worth the upgrade for me anymore... oh well, ill just stick to my current CPU (980BE at 3.9Ghz) and get a R7 270X instead and a new HDD and more Ram instead of a new FM2+ system with Kaveri.

There are no Single Core benchmarks in there, but the Muti Threaded is horrible, ill be the first to declare i was wrong, i was expecting more, and kaveri just does not cut it for me... from what i understand in those benchmarks Kaveri barely beats Trinity, were not even talking Richland here.

We must be reading a different review, because the one that I see shows that the Steamroller CPU in Kaveri can be up to 30% faster than Piledriver, which agrees very much with previously leaked benchmarks of the ES

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

I actually expected Kaveri to be almost on par with the i5 2500k, but i thought it would be like 3-5% behind it, but it will still be good enough.

But you said it, up to 30% faster... but usually only 6% faster, that`s what i am seeing and avg of 6% better MultiThreaded CPU benchmarks, i assume who ever is doing the benchmarks is avoiding completely Single Core performance.

Cinebench scores can be avoided, not only they are optimized for Intel, but the new R15 version cuts down Kaveri and puts it at same level than Trinity, which is rather odd. We know that ICC forces AMD chips to run the slowest possible code to fake benchmarks, but here R15 seems to introduce extra cut-down on Kaveri. Just look at Kaveri performing much better on R11.5

Now about CPU improvements shown in the game benchmarks.

Fritz: 19% IPC
Arion: 13% IPC
Grid 2: 10% IPC
Resident Evil 5: 34% IPC
Metro LL: 12% IPC
Hitman Absolution: 32% IPC
Batman AC: 21% IPC

I don't know from where you obtain that Steamroller offers only a 6% IPC gain over Piledriver. The average is more than 20% and in a pair of cases it goes above 30%.

And it looks much better at compute. I can see Kaveri APU is 40% faster than Trinity APU, in Luxmark, 158% faster in Raytrace, and 177% faster in fluid simulation.

Ok what is IPC in your eyes and does if differ from performance per clock?

 

[etayorius]

Juanrga, the "djohny" guy, who posted the Kaveri Benchmark avoided completely Single Core benchmarks... i wonder why.

 

[esrever]

Juanrga, the "djohny" guy, who posted the Kaveri Benchmark avoided completely Single Core benchmarks... i wonder why.

probably because single core has been dead for years now.

 

[griptwister]

So... look at what I stumbled upon... ASRock came out with a MoBo that supports 64Gb of RAM for AM3+...

http://www.asrock.com/mb/AMD/Fatal1ty%20990FX%20Killer/

Looks like a "Killer" deal. Bahahaha! Looks like I found a MoBo for my primary Rig.

 

[etayorius]

Juanrga, the "djohny" guy, who posted the Kaveri Benchmark avoided completely Single Core benchmarks... i wonder why.

probably because single core has been dead for years now.

We are not quite there yet, was not the whole reason why bulldozer was such a fail and Piledriver was introduced?

 

[Ranth]

No. The laws of physics says otherwise. That is the reason that both Nvidia and AMD engineers are designing their future supercomputers around APUs, and not around dCPU+dGPU anymore. Nvidia 40TFLOP design is scheduled for ~2018.

I really want to see what you say the physics are saying, because afaik: More transistors = More heat = less performance | Bigger diesize = fewer yields = higher cost

Also the problem I see with seperating the chip into 2 dies, is again afaik: space

Though there are most certainly more problems/examples than that I just used what I know so.. 😀

Several years ago, studies were made on how to design a supercomputer that would be 1000 times more powerful than actual ones. The research revealed that current engineering paradigms don't work at the exascale level. The findings were summarized in a DARPA report. One of the challenges identified in the report was that at the exascale level the principle of locality holds. A traditional design dCPU+dGPU violates the principle of locality. That is the reason why the engineers of Nvidia and of AMD have selected APUs for their respective exascale supercompute designs. A dCPU+dGPU doesn't work at that scale of performance.

I believe it is time that some people at this forum stop repeating wrong mantras such as "APUs are for cheap gaming".

But wouldn't we run into problems with diesize and thermal limits/density negating the performance increase of the principle of locality?

 

[Cazalan]

I believe it is time that some people at this forum stop repeating wrong mantras such as "APUs are for cheap gaming".

They are for cheap gaming until someone actually ships a more powerful APU to break the paradigm. What looks good on paper doesn't mean they can actually FAB it. Until then its vaporware.

 

[juanrga]

Spoiler

Understanding Juanrga is a lot more easier if you realize that he's basically the first real ARM fanboy to defend ARM so heavily.

The footnotes in the Nvidia slides where they're making all these claims are laughable at best. Lines like "estimates" and "20 percent margin of error" and "PS3 only using one core and no SPUs".

Nvidia always does this with Tegra. They hype it up like none-other. In fact, they even boasted Tegra 2 was on 360 levels and you'd have the power of a console in your pocket. And look how that turned out.

The truth is Tegra K1 final silicon is a long ways away, at least 10 months. We will be talking about Excavator at that point, and the other ARM companies wilil probably be with ARM64 then too.

Nvidia always starts hyping Tegra way before it's actually due. They're very good at marketing, and if Qualcomm or Samsung were out here hyping their ARM products that would ber coming out around the time of Tegra, people would be going crazy too.

You have used my comments about Denver core made in another part to launch here an anti-Nvidia rant... and you call "fanboy" to others! LOL

But at least this time you are spelling my nickname correctly. Thanks!

Haha, I'm learning 😉

But Juan, the thing is that I don't think it's necessarily bad to be a fanboy to a certain extent. I'm arguably a fanboy for HEDT platforms with serious power, and you're more for low power ARM. I don't know, maybe fanboy is not the right word for what I'm describing, since we both are more aligned with platforms as opposed to brands.

But Nvidia has a massive history of lying about Tegra and making it sound far better than it really is. I would give it a few weeks or so until Nvidia's lies start to show up. We're already seeing them with the 2w figure being used when playing regular games, like Angry Birds and Bejeweled clones, which are all 2d games that use sprites and probably aren't even GPU accelerated.

I'm always very pessimistic when Nvidia talks about ARM. If this was Qualcomm or Samsung, or even Apple (which I despise), I'd be a little more optimistic about what they're claiming. I just feel like this is basically the 5th time Nvidia has gotten on stage and talked about how amazing Tegra was, how it would beat the x86 CPU of the day, etc.

"Tegra 2 faster than Core 2 Duo":
http://androidcommunity.com/kal-el-nvidia-chipset-benchmarked-against-core2duo-and-tegra-2-20110218/

"Tegra 3 faster than Core 2 Duo":
http://www.engadget.com/2011/11/09/nvidia-says-tegra-3-is-a-pc-class-cpu-has-screenshots-to-prov/
http://www.brightsideofnews.com/news/2011/2/21/why-nvidiae28099s-tegra-3-is-faster-than-a-core-2-duo-t7200.aspx

Tegra 4 faster than Snapdragon 800:
http://www.phonearena.com/news/Is-the-NVIDIA-Tegra-4-faster-than-the-Qualcomm-Snapdragon-800_id40741

in reality, not really:
http://tabshowdown.blogspot.com/2013/11/apple-a7-vs-nvidia-tegra-4.html

Now, Tegra K1, faster than the latest x86 cpu!

See the trend? Nvidia has done nothing but cherry pick benchmarks and exaggerate potential of their Tegra chips for nearly as long as they have existed.

You can do your own googling and take a look, I just gave this post a few minutes of my time. But Nvidia has pretty much always lied about their performance of their Tegra chips.

This is the second time that you ignore what I say just to post your anti-Nvidia rants here.

I am not "fanboy" for low power anything. I am a fan of efficient and clean ISAs/muarches. I only know one ISA that is efficient, clean and modern. If you know more let me know. If tomorrow AMD and Intel join and develop a new "x86-B" that is not an extension of an bloated and ancient ISA I will be fan of that as well.

My claimed 16FLOPs/clock figure was not said by any PR marketing guy. I got it. I can also estimate the DMIPS/MHz from the arch. I also got by myself the TDP. I got a dissipation slightly above the 6W. Marketing slides claim other number. In fact, in the other thread, people is asking from where the 6W number has come. You know it because you are participating in it, but with a different nickname sd***

 

[juanrga]

No. The laws of physics says otherwise. That is the reason that both Nvidia and AMD engineers are designing their future supercomputers around APUs, and not around dCPU+dGPU anymore. Nvidia 40TFLOP design is scheduled for ~2018.

I really want to see what you say the physics are saying, because afaik: More transistors = More heat = less performance | Bigger diesize = fewer yields = higher cost

Also the problem I see with seperating the chip into 2 dies, is again afaik: space

Though there are most certainly more problems/examples than that I just used what I know so.. 😀

It's mostly about heat vs thermal solutions required to cool them. A single CPU with a 125W TDP combined with a separate GPU with 250W TDP is always, 10 times out of 10, going to be higher performance then a single 125W CPU + GPU combo chip. And that's before going into the die space issue vs cost's. Juan has been blowing smoke for a long time now for no matter what they put in an APU they can always put more/ faster of the same in a dCPU + dGPU. The laws of thermal dynamics are what limits Intel / AMD here.

EX:
Richland / Kavari vs
8350 / i7 + R9 290

In every conceivable performance profile the second combo wins as whatever vector work the iGPU can do, the dGPU can do more of and faster. HSA doesn't change any of that, it just allows the iGPU to be used for something while a dGPU does something else so your not wasting 50% of your CPU's die space.

Thanks by this new personal attack. I could not expect anything different from you.

Now you only need to take your deep understanding of thermodynamics and meet all Nvidia, AMD, Intel, Cray... engineers (Cray is one of the seven partners of Nvidia in their exascale supercomputer project).

Please meet them and explain them your "laws of thermal dynamics". It is funny that you believe that the same people who designs the current supercomputers are not aware that "10 times out of 10" a dCPU+dGPU offers more performance than an APU. Poor engineers, they must be smoking the same than me. :lol:

end-sarcasm

In fact, the same laws of physics (that you don't understand) say that your a dCPU+dGPU would offer about 10x less performance than the APU designs that those same engineers are developing for exascale level computing.

No. The laws of physics says otherwise. That is the reason that both Nvidia and AMD engineers are designing their future supercomputers around APUs, and not around dCPU+dGPU anymore. Nvidia 40TFLOP design is scheduled for ~2018.

I really want to see what you say the physics are saying, because afaik: More transistors = More heat = less performance | Bigger diesize = fewer yields = higher cost

Also the problem I see with seperating the chip into 2 dies, is again afaik: space

Though there are most certainly more problems/examples than that I just used what I know so.. 😀

Several years ago, studies were made on how to design a supercomputer that would be 1000 times more powerful than actual ones. The research revealed that current engineering paradigms don't work at the exascale level. The findings were summarized in a DARPA report. One of the challenges identified in the report was that at the exascale level the principle of locality holds. A traditional design dCPU+dGPU violates the principle of locality. That is the reason why the engineers of Nvidia and of AMD have selected APUs for their respective exascale supercompute designs. A dCPU+dGPU doesn't work at that scale of performance.

I believe it is time that some people at this forum stop repeating wrong mantras such as "APUs are for cheap gaming".

But wouldn't we run into problems with diesize and thermal limits/density negating the performance increase of the principle of locality?

At contrary, it is the same principle (based in the laws of physics) which puts thermal and power limits to the traditional dCPU+dGPU. Scientists/engineers refer to it as the "power wall" or the "energy wall".

Or do you believe that you know more than all the engineers working at AMD, Nvidia, Intel, Cray... together?

I believe it is time that some people at this forum stop repeating wrong mantras such as "APUs are for cheap gaming".

They are for cheap gaming until someone actually ships a more powerful APU to break the paradigm. What looks good on paper doesn't mean they can actually FAB it. Until then its vaporware.

Wrong. Berlin APU is not for cheap gaming. Its successor (Toronto APU) is not for cheap gaming...

Moreover, we already know (there is consensus among researchers) that only an APU can power exascale level compute. And no it is not waporware, but actual research made under government and other grants.

In fact, the original 10TFLOP APU design made by AMD details a FSA system architecture that is the precursor of the HSA finally introduced in Kaveri.

 

[jdwii]

^ lol when he say's we ignore what he says it make me feel bad

 

[juanrga]

Let me make it clear, i am not saying Kaveri is crap... it is not, the GPU inside Kaveri and other features like HUMA, HSA, MANTLE and OpenCL will be massively boosted by Kaveri, i am talking about the laughable numbers Kaveri is getting compared to Richland and Trinity in regards of IPC.

As i already stated regarding Kaveri´s laughable Skyrim performance improvement against Trinity in the Spanish Forum and the Pudget review, it makes me believe Kaveri has IPC between Trinity and Richland... is this bad news? not necessarily... Kaveri is a damn great upgrade to anyone who is on a budget or who is not a hardcore gamer, like one person asked if Kaveri should be worthy upgrade compared to Llano, HECK IT IS you bet!

Therefore the Spanish review showing that the CPU in Kaveri offers about a 20% IPC gain (more than 30% in a pair of cases) is what you call "laughable numbers"? Wow! What were you expecting? Maybe a 2000% increase in IPC in a single generation?

The numbers that they are obtaining are very much that I was expecting. Re-read my article.

The Spanish reviewer finishes is review with

Otros juegos como Battlefield 4, Diablo 3, Naruto Ninja Storm, Xcom EU, etc han mejorado sobremanera. Pero comentaré más, mas adelante.

Por lo demás, tenemos justo los resultados anunciados y esperados.

Translated to:

Other games like Battlefield 4 Diablo 3, Naruto Ninja Storm, Xcom EU, etc have improved greatly. But I will comment more later.

Otherwise, we just have the announced and expected results.

 

[juanrga]

Bonus

http://foro.noticias3d.com/vbulletin/showthread.php?t=421821&p=5036333&viewfull=1#post5036333

Kaveri APU vs i3+HD7750

The 7750 is the GDDR5 version. The Kaveri APU uses iGPU and non-final drivers. As predicted Kaver iGPU is between the 7750-DDR3 and the 7750-GDDR5.

With final optimized drivers, the OC Kaveri could be near the 7750-GDDR5 level for 1080p gaming.

 

[noob2222]

@juan

If you want to average the benchmarks, dont just pick the best 7 benchmarks and ignore the other 9 just because they dont coincide with your predictions of kaveri being equal to an i5 2500k.

 

[jacobian]

Another set of benchmarks are out, including overclocked A10-7850K.

http://wccftech.com/amd-kaveri-a10-7850k-overclocked-45-ghz-benchmarked-a105800k/

Most benchmarks show a healthy gain on A10-5800K. However, when speaking of CineBench R11.5 multi-threaded:

"Only at overclocked frequencies does the A10-7850K matches the FX-6300 which is a 6-Core Piledriver based CPU."

It seems like for guys who don't need the iGPU, the FX-6300 is still a good value, being priced $120 on newegg right now.

 

[Ranth]

At contrary, it is the same principle (based in the laws of physics) which puts thermal and power limits to the traditional dCPU+dGPU. Scientists/engineers refer to it as the "power wall" or the "energy wall".
Or do you believe that you know more than all the engineers working at AMD, Nvidia, Intel, Cray... together?

Would you care to explain in detail what physics you are refering to because that makes no sense AT ALL...... What principle? The laws of physics, well that broad, and leave me with no clue of what you're trying to say..

So what I think you're saying is that cramming more onto a single dies reduces power-consumption, thermal density, increase performance? Sorry if I find that hard to believe.

 

[Reynod]

Actually Kaveri looks pretty good.

I did warn the fanbois not to get all baronmatrix over it ... did anyone listen ... no.

If the damned thing was that good don't you think the reviewers would have slipped and told one of us by now who has some credibility?

 

[de5_Roy]

i'll freak out if toms review shows that the igpu is still bottlenecked and amd traded off keeping the igpu well-fed for cpu bandwidth.
and if there isn't a good number of well-known and used softwares supporting hsa (using vce's treatment as reference).
edit: not in this thread. in the review thread. there will be a rant.

 

[Reynod]

I honestly have not had a chance to ask one of them but I imagine it will be seen positively ...

crashman may have a rant ... we encourage him to "let it all out" once in a while ... otherwise he gets depressed and starts retesting the PSU's under a dead short to see what holes the smoke comes out of.

I think they stopped his extreme overclocking efforts because he deliberately likes blowing stuff up.

crash is my fav because he is as diplomatic as a Chinese Tank driver ... when the cameras are not on.

 

[-Fran-]

Well, de5_roy... The iGPU inside the APUs has been bottlenecked by the RAM since Llano (hitting 2133 as a ceiling, IIRC). We were all hoping for a GDDR5, EDRAM or DDR4 approach with Kaveri, but timing and because reasons we're stuck with DDR3 still. Since APUs are a cheap solution while HSA catches up peoples attention, they can't force people to switch, yet again, from Trinity/Richland to a more expensive platform without everything sorted out. Keeping Kaveri in FM2 is a cost decision, not a technical one, I think.

Anyway, the Spanish review gives more hope than not. Let's wait for Toms and the rest to confirm and see how it turns out.

Cheers!

 

[gamerk316]

No. The laws of physics says otherwise. That is the reason that both Nvidia and AMD engineers are designing their future supercomputers around APUs, and not around dCPU+dGPU anymore. Nvidia 40TFLOP design is scheduled for ~2018.

I really want to see what you say the physics are saying, because afaik: More transistors = More heat = less performance | Bigger diesize = fewer yields = higher cost

Also the problem I see with seperating the chip into 2 dies, is again afaik: space

Though there are most certainly more problems/examples than that I just used what I know so.. 😀

Several years ago, studies were made on how to design a supercomputer that would be 1000 times more powerful than actual ones. The research revealed that current engineering paradigms don't work at the exascale level. The findings were summarized in a DARPA report. One of the challenges identified in the report was that at the exascale level the principle of locality holds. A traditional design dCPU+dGPU violates the principle of locality. That is the reason why the engineers of Nvidia and of AMD have selected APUs for their respective exascale supercompute designs. A dCPU+dGPU doesn't work at that scale of performance.

I believe it is time that some people at this forum stop repeating wrong mantras such as "APUs are for cheap gaming".

Ah, but you assume that space/heat/power is not an issue. In the consumer space, it is, and given the design sacrifices that you have to make to slap a CPU/GPU on the same die, maximum performance is still made via a dedicated CPU, dedicated GPU, and a very fast link between them.

Also FYI, off die memory also violates locality. That's why we have more and more levels of CPU cache; its to hide the fact memory access is SLOW compared to the speed the CPU operates at. If you had infinite money, infinite die space, infinite levels of thermal control, then you could slap on 16GB of L1 cache if you wanted to, to maximize performance.

You have maximum performance, and maximum performance that makes financial sense.