AMD CPU speculation... and expert conjecture

[noob2222]

HSA is a great technonlgy, and having an APU or APUsoc is great. Even with a high end dGPU like the R9 290x, having the iGPU to offload calculations to is helpfull. Not to mention the faster bus speeds. IF they can make an apu with hyper threadding so we can have our 8 threads, for cpu intensive tasks, the HSA idea could be the Athlon 64 of today, the platform that beats intel.

HSA= more performance. Stop hating

IF ITS SUPPORTED. AMD is banking right now in software to save their ass. If its not adopted, where does that leave things? with a crappy cpu and an idle IGP.

The thing is the way everyone is talking, AMD is making weak ass cores and software that will make up the difference. Software that will NOT be in the masses on release day. Software that may take years to get into the hands of developers on a regular basis.

Software that Intel will spend every last dollar AMD has to stop. AMD is a hardware company that is now wanting to be a software company.

If the HSA adoption was more hardware based, it would be great. The problem right now is it looks to rely heavily on software. Software that will only be supported on AMD hardware. How many developers are going to ignore Intel's bribes?

 

[juanrga]

HTX slots and Hypertransport are 'dead'. The HPC community is moving away from Hypertransport. One of reasons why Cray abandoned AMD by Intel is because the Opteron don't support PCIe 3. The new Berlin APUs for servers don't support Hypertransport but support PCIe 3, including PCI endpoint mode. Also the ARM server SoC Seattle uses PCIe, not Hypertransport.

HTX slots for HEDT is still more unreliable than the idea of a SR 6/8 core SR FX from special die.

HTX provides more bandwidth, and lower latency than PCIe 3.0. The reason Cray is going away from it, is because the company who bought a majority stake in them (Intel) does not use it...so HTX is worthless if you're building IntVidia super computers.

(This is also likely the reason Cray switched from AMD processors to begin with, regardless of marketing spin)>

I don't know latencies but PCIe provides more bandwidth. HPC people reports 96GB/s PCIe 3 interconnects, whereas hypertransport maximum is 51.2 GB/s.

HTX slot is slow. Even the new HTX3 is.

Cray required AMD to provide PCIe3 support because hyper-transport was slow and doesn't scale well. Cray required this for years. AMD didn't answer (Opteron only provides PCIe 2 support) and finally Cray switched AMD by Intel.

Our move to PCI Gen 3 was critical to having higher bandwidth in and out of compute nodes and currently Intel has the best PCI Gen 3 support, so it matches our network and the performance is good.

When we built the XT, XC, XK line, the best way to talk to a processor was through AMD's proprietary HyperTransport mechanism. That's what we based our network on because there wasn't any better way to talk to a processor, so it started with the network and the Opteron happens to be a great processor. [In the XC30] it starts with the PCI Gen 3 because you need to have a more scalable network.

It is also related to the use of GPGPUs being bottlenecked in a PCIe 2 slot.

We will have PCI Express Gen 3 as well. The switch by Cray to PCI Express Gen 3 is expected on that side of the business, because PCI Gen 3 is where all the discrete GPU accelerators will be connected to, so we have Gen 3 in the works as well. There is nothing preventing putting an AMD chip into a Gen 3 based system.

What is crazy is that AMD FirePro GPGPUs are PCIe 3 whereas the Opteron only support PCIe2!!!

The new Berlin APU and Seattle SoC abandon hypertransport and provide PCIe3 support. The new AMD seems to be doing things right, is not?

 

[juanrga]

Wow!

(i)
First surprise: the 1050 GFLOP are gone. Kaveri is 856 GFLOP.

@Sigmanick Thanks by compliments!! Yes. I used the public 1050 GFLOP given by Lisa Su to obtain the freq. of the GPU. If I had known the new secret 856 GFLOP number I had given more accurate GPU freqs. because I used the same formula than AMD labs

I got the above formula from analysis of architecture. Steamroller is a 8 FLOP per core arch. I recall one 'expert' here who said I got the FLOPs incorrectly. We discuss this during weeks. He was wrong and gone from the thread. Others also said me that Steamroller would use 256-bit FMAC units. Also wrong. The above formula correspond to 128-bit FMAC units.

My prediction of incorrect GPU freq. is irrelevant because I didn't do any graphics/gaming predictions. Very little information was known about the Kaveri GPU to make a reliable prediction. However, I was more close to the final GPU freq. than that Italian site that said that GPU was 1.0-1.2 GHz part.

(ii)
I assumed 3.8GHz as minimum CPU freq. the final 3.7GHz is rather close. I did performance estimations for 4GHz but did subtract 5% for safety reasons. This means that performance estimations were about 3% higher than correspond to a 3.7GHz CPU.

However, I assumed SR is only a 20% faster than PD and in the light of recent leaks SR seems to be faster. It is highly probable that will perform better than I predicted.

All the predictions by experts such as SeronX about Kaveri CPU being a 2.6GHz part are gone. The Italian site claiming 2.9GHz is gone. People expecting more than 4.5GHz SR are gone.

(iii)
The title of one of the slides is "UP TO 12 INTEGRATED HETEROGENEOUS COMPUTE UNITS". They count the total of compute units (GCN+x86).

People believing in 6 or 8 SR core are gone. wcftech claiming that Kaveri was a 13 GCN cores APU is gone.

It is 4SR+8GCN as predicted.

(iv)
The GPU performance is fantastic. They used BF4 (ha ha). Anxiously waiting to see more benchmarks and MANTLE.

 

[de5_Roy]

HP’s Moonshot now supports AMD’s Opteron X-Series APUs
http://semiaccurate.com/2013/11/12/hps-moonshot-now-supports-amd-opteron-x-series-apus/
with pics!

kaveri is one step closer to a 'big' soc with integration of audio dsp. iirc kaveri should also have arm's trustzone 'core' for security. i wonder if amd added the audio block after redesign or was hiding it the whole time.

IF ITS SUPPORTED. AMD is banking right now in software to save their ass. If its not adopted, where does that leave things? with a crappy cpu and an idle IGP.

as much as like to complain about amd's opposite-of-stellar software support and contributions, they seem to be still working to get the er... sdk(..?) and other middlewares(..?) to isv. i finally read something new on codexl, after months. timing and execution will be crucial for amd.
i highly doubt that the cpu will be crappy and igpu would be idle. even without hsa et al, the apu by itself seems quite potent. now i am worried about amd's silence on those apus' real arena - mobile deployment.

The thing is the way everyone is talking, AMD is making weak ass cores and software that will make up the difference. Software that will NOT be in the masses on release day. Software that may take years to get into the hands of developers on a regular basis.

i agree (although i can't tell if you're being sarcastic 😛). if it's about gpu acceleration, how many mainstream softwares have adopted it since llano came out and how many stayed as they were? i am not talking about professional or content creation, but mainstream consumer software that perform task like de/compression, consumer video creation editing/conversion and so on. some of the softwares that do have support, show great potential, but the amount of softwares are low.

Software that Intel will spend every last dollar AMD has to stop.

hell yeah. i suspect that's what might get in the way of amd's server/workstation segment. although... if amd's arm products have hsa, they won't have much trouble spreading since arm is a founder member.

Software that will only be supported on AMD hardware. How many developers are going to ignore Intel's bribes?

amd isn't the only one in hsa foundation.

 

[gamerk316]

HSA is a great technonlgy, and having an APU or APUsoc is great. Even with a high end dGPU like the R9 290x, having the iGPU to offload calculations to is helpfull. Not to mention the faster bus speeds. IF they can make an apu with hyper threadding so we can have our 8 threads, for cpu intensive tasks, the HSA idea could be the Athlon 64 of today, the platform that beats intel.

HSA= more performance. Stop hating

IF ITS SUPPORTED. AMD is banking right now in software to save their ass. If its not adopted, where does that leave things? with a crappy cpu and an idle IGP.

The thing is the way everyone is talking, AMD is making weak ass cores and software that will make up the difference. Software that will NOT be in the masses on release day. Software that may take years to get into the hands of developers on a regular basis.

Software that Intel will spend every last dollar AMD has to stop. AMD is a hardware company that is now wanting to be a software company.

If the HSA adoption was more hardware based, it would be great. The problem right now is it looks to rely heavily on software. Software that will only be supported on AMD hardware. How many developers are going to ignore Intel's bribes?

Frankly, OpenCL is more attractive then HSA, since OpenCL would work for APU's, AMD GPU's, NVIDIA GPU's, Intel iGPU's, etc. I simply don't see masses of S/W being there, ever.

Heck, OpenCL is having a tough enough time getting adopted, even with AMD pushing it for about all its existence. HSA has almost no chance in my mind.

 

[8350rocks]

HSA is a great technonlgy, and having an APU or APUsoc is great. Even with a high end dGPU like the R9 290x, having the iGPU to offload calculations to is helpfull. Not to mention the faster bus speeds. IF they can make an apu with hyper threadding so we can have our 8 threads, for cpu intensive tasks, the HSA idea could be the Athlon 64 of today, the platform that beats intel.

HSA= more performance. Stop hating

IF ITS SUPPORTED. AMD is banking right now in software to save their ass. If its not adopted, where does that leave things? with a crappy cpu and an idle IGP.

The thing is the way everyone is talking, AMD is making weak ass cores and software that will make up the difference. Software that will NOT be in the masses on release day. Software that may take years to get into the hands of developers on a regular basis.

Software that Intel will spend every last dollar AMD has to stop. AMD is a hardware company that is now wanting to be a software company.

If the HSA adoption was more hardware based, it would be great. The problem right now is it looks to rely heavily on software. Software that will only be supported on AMD hardware. How many developers are going to ignore Intel's bribes?

Frankly, OpenCL is more attractive then HSA, since OpenCL would work for APU's, AMD GPU's, NVIDIA GPU's, Intel iGPU's, etc. I simply don't see masses of S/W being there, ever.

Heck, OpenCL is having a tough enough time getting adopted, even with AMD pushing it for about all its existence. HSA has almost no chance in my mind.

That depends on how many big software companies back HSA, Oracle is already pushing Java to be HSA native with the Sumatra project. If a company like Oracle is going so far as to do that, I think Intel would ultimately have to cave.

On a side note...

I am anxious to see what other announcements come out before the conference ends. I am not sure we will get a dCPU announcement out of this conference, though I expect it will be coming.

 

[gamerk316]

Java is a dying programming language. C/C++, Objective C, and C# are all gaining traction, due to MSFT and Apple pushing them, and the fact they are simply faster languages. All the tech schools are switching back to C based courses, transitioning back from Java. In 10 years, Java will be used about as much as Ada (which ironically, I'm stuck supporting on work for the next decade. The DoD is silly sometimes...)

 

[8350rocks]

Java is a dying programming language. C/C++, Objective C, and C# are all gaining traction, due to MSFT and Apple pushing them, and the fact they are simply faster languages. All the tech schools are switching back to C based courses, transitioning back from Java. In 10 years, Java will be used about as much as Ada (which ironically, I'm stuck supporting on work for the next decade. The DoD is silly sometimes...)

Java is still used heavily in C/S architectures, specifically for the client side of things...

EDIT: Though I would honestly prefer something like python anyway...

 

[de5_Roy]

^^ kaveri is a consumer product... but berlin is for servers and possibly for workstations. add 64-bit arm that are gunning for high-revenue markets and you have a motive for hsa push.
however, both amd(and arm) and hsa are trying to penetrate a market which already has suitable tools. they'd have to offer more.. like less cost of..uh.. stuff. the biggest barrier may be educating developers on hsa and providing customer support (and cost of both).

i still don't get why mediatek's guy is at apu 13. it can't be just because of mediatek's membership. are we gonna see cheap, gcn-powered mediatek socs? radeon set to take over china and whitebox tablets? yes, it is an over the top speculation.

 

[tracker45]

Is the
A10-7850K using a 7850 graphics card in the APU ?? That GPU seems like it is too powerful to fit into an APU.

 

[de5_Roy]

Is the
A10-7850K using a 7850 graphics card in the APU ?? That GPU seems like it is too powerful to fit into an APU.

apparently, it is being called an R7 series igpu. previous rumor was R5 M200 igpu, which i now suspect might be an igpu of a laptop apu. dt kaveri might have R7 D200-600 igpu or something.

no it's not using a radeon hd 7850 igpu. 7850 has over a thousand shaders(edit: 1280) while kaveri has only 512.

the igpu is too powerful though (seems from the specs) but the raw power is not enough, it has to have proper outlet i.e. bandwidth so that there's no bottleneck. unfortunately, the apus igpus have memory bottleneck since llano.

 

[jdwii]

Kaveri will ship in Q4 2013 and offer 856 GFLOPS computing power.

That really fell from their "first teraflop APU" talk.

To be fair was that a official statement from Amd or a hyped up statement from a fanboy?

 

[jdwii]

HSA is a great technonlgy, and having an APU or APUsoc is great. Even with a high end dGPU like the R9 290x, having the iGPU to offload calculations to is helpfull. Not to mention the faster bus speeds. IF they can make an apu with hyper threadding so we can have our 8 threads, for cpu intensive tasks, the HSA idea could be the Athlon 64 of today, the platform that beats intel.

HSA= more performance. Stop hating

IF ITS SUPPORTED. AMD is banking right now in software to save their ass. If its not adopted, where does that leave things? with a crappy cpu and an idle IGP.

The thing is the way everyone is talking, AMD is making weak ass cores and software that will make up the difference. Software that will NOT be in the masses on release day. Software that may take years to get into the hands of developers on a regular basis.

Software that Intel will spend every last dollar AMD has to stop. AMD is a hardware company that is now wanting to be a software company.

If the HSA adoption was more hardware based, it would be great. The problem right now is it looks to rely heavily on software. Software that will only be supported on AMD hardware. How many developers are going to ignore Intel's bribes?

Noob cool it alright. The Amd piledriver isn't even that bad steamroller is reported to be better. A 4 core Piledriver competes well with a I3 and a 20-30% boost in Single threading performance will make it closer to a I5 Hell a 6 Core APU might trade blows with a I7 if Amd could make one. HSA is more like a bonus for now and Mantle as well. Just think a 150$ APU competing with a 220$ CPU and a 120$ GPU if Mantle is used that would have some good value to it. Then HSA could make that 150$ APU compete with a 1000$ CPU.

 

[8350rocks]

HSA is a great technonlgy, and having an APU or APUsoc is great. Even with a high end dGPU like the R9 290x, having the iGPU to offload calculations to is helpfull. Not to mention the faster bus speeds. IF they can make an apu with hyper threadding so we can have our 8 threads, for cpu intensive tasks, the HSA idea could be the Athlon 64 of today, the platform that beats intel.

HSA= more performance. Stop hating

IF ITS SUPPORTED. AMD is banking right now in software to save their ass. If its not adopted, where does that leave things? with a crappy cpu and an idle IGP.

The thing is the way everyone is talking, AMD is making weak ass cores and software that will make up the difference. Software that will NOT be in the masses on release day. Software that may take years to get into the hands of developers on a regular basis.

Software that Intel will spend every last dollar AMD has to stop. AMD is a hardware company that is now wanting to be a software company.

If the HSA adoption was more hardware based, it would be great. The problem right now is it looks to rely heavily on software. Software that will only be supported on AMD hardware. How many developers are going to ignore Intel's bribes?

Noob cool it alright. The Amd piledriver isn't even that bad steamroller is reported to be better. A 4 core Piledriver competes well with a I3 and a 20-30% boost in Single threading performance will make it closer to a I5 Hell a 6 Core APU might trade blows with a I7 if Amd could make one. HSA is more like a bonus for now and Mantle as well. Just think a 150$ APU competing with a 220$ CPU and a 120$ GPU if Mantle is used that would have some good value to it. Then HSA could make that 150$ APU compete with a 1000$ CPU in GPU accelerated workloads.

Fixed!

 

[juanrga]

IF ITS SUPPORTED. AMD is banking right now in software to save their ass. If its not adopted, where does that leave things? with a crappy cpu and an idle IGP.

The thing is the way everyone is talking, AMD is making weak ass cores and software that will make up the difference. Software that will NOT be in the masses on release day. Software that may take years to get into the hands of developers on a regular basis.

Software that Intel will spend every last dollar AMD has to stop. AMD is a hardware company that is now wanting to be a software company.

If the HSA adoption was more hardware based, it would be great. The problem right now is it looks to rely heavily on software. Software that will only be supported on AMD hardware. How many developers are going to ignore Intel's bribes?

So much AMD hate lately...

Frankly, OpenCL is more attractive then HSA, since OpenCL would work for APU's, AMD GPU's, NVIDIA GPU's, Intel iGPU's, etc. I simply don't see masses of S/W being there, ever.

Heck, OpenCL is having a tough enough time getting adopted, even with AMD pushing it for about all its existence. HSA has almost no chance in my mind.

OpenCL is part of HSA. Second, HSA allows better control and performance than OpenCL. Third, you also said us that nobody was going to use MANTLE...

 

[de5_Roy]

s/a remember's the first tflops apu amd announced:
http://semiaccurate.com/2013/11/12/amd-misses-expectations-kaveri/
has the older promo slide showing first 1050 Gflops.
the first time i read about kaveri delay, which i didn't want to believe:
http://semiaccurate.com/2012/11/06/amds-kaveri-apu-slips-again-2014-now/
before multiple 'follow ups' from amd telling us that kaveri was still 'on track'.

"AMD is collaborating with its technology partners and the open source community to provide developers with tools that enable them to build server applications that utilize both CPU and GPU compute capabilities available in its revolutionary HSA based server APUs."
http://www.xbitlabs.com/news/other/display/20131111230028_AMD_Teams_Up_with_Open_Source_Community_to_Develop_Heterogeneous_Server_Software.html

Report: GlobalFoundries to manufacture SoCs for Apple devices (common platform alliance in action)
http://techreport.com/news/25638/report-globalfoundries-to-manufacture-socs-for-apple-devices
not directly amd-related.. but glofo seems to be making it's priorities clear.

 

[Ranth]

Juan? AMD hate? I don't think Noob is hating, he's just pointing out possible pitfalls in AMD's strategy. Just because you're not agreeing one everything X does, is not equal to hating X...

Also would people stop with the attacks/insults/taunts? Anger/arrogance are not useful in a discussion. So calm down, and discuss properly.

 

[gamerk316]

Java is a dying programming language. C/C++, Objective C, and C# are all gaining traction, due to MSFT and Apple pushing them, and the fact they are simply faster languages. All the tech schools are switching back to C based courses, transitioning back from Java. In 10 years, Java will be used about as much as Ada (which ironically, I'm stuck supporting on work for the next decade. The DoD is silly sometimes...)

Java is still used heavily in C/S architectures, specifically for the client side of things...

EDIT: Though I would honestly prefer something like python anyway...

Python was quite powerful in its time, but the world have moved on to more powerful scripting languages. I prefer Ruby personally, but that's probably the Smalltalk in me speaking out right now.

C/S may still be teaching Java, but a lot of schools are moving back toward C again. Apple is pushing this a LOT (since Objective C isn't that far from the base C syntax).

 

[gamerk316]

http://semiaccurate.com/2013/11/12/amd-misses-expectations-kaveri/

The major issue here:

The targets for Kaveri that were announced by AMD were 4 Ghz for the CPU side and 900 Mhz for the GPU. What the company is actually delivering is 3.7 Ghz on the CPU and 720 Mhz on the GPU which is an eight percent miss on the CPU and a massive 25 percent miss on the GPU side.

The GPU miss is very, very worrying. The question needs to be asked, why a 25% miss?

I think the CPU holds the answers: Heat. I have a suspicion the GPU portion is now so powerful, AMD is finding they can't keep the package cool with a normal HSF. If so, them you can basically guarantee that these things will NOT OC (being close to the maximum possible clocks) particularly well. This also implies AMD won't be able to increase GPU power generation-over-generation as easily, now relying strictly on architecture improvements, rather then clock ramps and adding more compute units.

If this is due to heat, then Kaveri is nearing the end of its current architectural life, since AMD is going to struggle to find ways to make it better. I'm going to be VERY interested in power numbers when this comes out.

 

[noob2222]

IF ITS SUPPORTED. AMD is banking right now in software to save their ass. If its not adopted, where does that leave things? with a crappy cpu and an idle IGP.

The thing is the way everyone is talking, AMD is making weak ass cores and software that will make up the difference. Software that will NOT be in the masses on release day. Software that may take years to get into the hands of developers on a regular basis.

Software that Intel will spend every last dollar AMD has to stop. AMD is a hardware company that is now wanting to be a software company.

If the HSA adoption was more hardware based, it would be great. The problem right now is it looks to rely heavily on software. Software that will only be supported on AMD hardware. How many developers are going to ignore Intel's bribes?

So much AMD hate lately...

Frankly, OpenCL is more attractive then HSA, since OpenCL would work for APU's, AMD GPU's, NVIDIA GPU's, Intel iGPU's, etc. I simply don't see masses of S/W being there, ever.

Heck, OpenCL is having a tough enough time getting adopted, even with AMD pushing it for about all its existence. HSA has almost no chance in my mind.

OpenCL is part of HSA. Second, HSA allows better control and performance than OpenCL. Third, you also said us that nobody was going to use MANTLE...

Its not hate per say, its the fact that IF AMD spends all their time/money/effort to push software to fix their hardware problem, where does that leave the if it fails?

AMD could end up spending every last penny they have trying to push this software solution only to watch Intel pay more to stop it.

What is left when that happens? Some slow (3.7 ghz vs 4.2 ghz) l3 cacheless cores and an IGP.

People praising this software only solution without even thinking of the concequences is what I hate.

AMD needs to fix the hardware issues (l3 cache, Main memory latency, ect) not just push software patches.

It seems some people couldn't be happier with this solution, contantly blabbing about GFLOP this and GFLOP that wich requires the Highly Specialized Application software to achieve.

 

[noob2222]

http://semiaccurate.com/2013/11/12/amd-misses-expectations-kaveri/

The major issue here:

The targets for Kaveri that were announced by AMD were 4 Ghz for the CPU side and 900 Mhz for the GPU. What the company is actually delivering is 3.7 Ghz on the CPU and 720 Mhz on the GPU which is an eight percent miss on the CPU and a massive 25 percent miss on the GPU side.

The GPU miss is very, very worrying. The question needs to be asked, why a 25% miss?

I think the CPU holds the answers: Heat. I have a suspicion the GPU portion is now so powerful, AMD is finding they can't keep the package cool with a normal HSF. If so, them you can basically guarantee that these things will NOT OC (being close to the maximum possible clocks) particularly well. This also implies AMD won't be able to increase GPU power generation-over-generation as easily, now relying strictly on architecture improvements, rather then clock ramps and adding more compute units.

If this is due to heat, then Kaveri is nearing the end of its current architectural life, since AMD is going to struggle to find ways to make it better. I'm going to be VERY interested in power numbers when this comes out.

This apparently was another side effect of the bad decision to try bulk silicon.

 

[gamerk316]

As the software guy you need to convince for AMD to succeed:

I agree with noobs post above. I am not going to waste my time optimizing my applications for HSA, MANTLE, CUDA, OpenCL, XB1, PS4, AVX, SSE 4.2, SSE 4.1, SSE4a, SSE3, and SSE2. And don't even get me started on supporting both a Win8 and traditional Windows user interface if I want to be featured in the Windows store! And lots not forget Linux/OSX support either; those are big things now apparently.

And even if someone does go that far, all you'll do instead is complain about how the program crashes every two minutes, tanks performance on some really rare setups, has missing features, and so on and so forth.

Its not happening. Deal with it. Kaveri is an attractive low cost SOC, but that's about it, and all it ever will be.

 

[lilcinw]

Juan? AMD hate? I don't think Noob is hating, he's just pointing out possible pitfalls in AMD's strategy. Just because you're not agreeing one everything X does, is not equal to hating X...

Also would people stop with the attacks/insults/taunts? Anger/arrogance are not useful in a discussion. So calm down, and discuss properly.

+1

People need to learn that criticism is not necessarily an attack. Keep your cool and respond with facts instead of emotion.

 

[gamerk316]

He has the psychological need to always be correct, so he feels almost obligated to respond to any post that questions his conclusions, and views anything that disagrees with those conclusions as an attack against him.

I would know; I do it too.

 

[de5_Roy]

What is left when that happens? Some slow (3.7 ghz vs 4.2 ghz) l3 cacheless cores and an IGP.

i wonder if this necessitates ... you know.. some skus with L3 cache😗 and ... may be better silicon to tackle clockrate and heat issue. ofc glofo will have to be on board with amd.

we might see kaveri revisions like 1.1 or 1.5 if the early units turn out to be bad.