AMD CPU speculation... and expert conjecture

[Ags1]

Spoiler

I wonder where this leaves SOI. AMD was the primary customer for SOI and it is almost like AMD is abandoning it because GloFo can't deliver.

it may be the other way around. glofo may be reluctant to serve the sole low-profitting, high margin, high-performance-demanding s.o.i customer instead of the likes of mediatek and rockchip.

I wonder if there is a big list of customers GloFo serves.

You do make a point, AMD had wafer supply agreements and they had that huge issue with tons of Llano stock not getting sold. That could have been a red flag for GloFo that they won't be making enough money off of AMD going SOI and relying on AMD to be their primary SOI customer.

That's extremely discouraging. I don't have high hopes at all for AMD to release parts that overclock like their parts previously did or to release a large high end chip if they are stuck on bulk.

I'm kind of holding out because I would think AMD might bring those products back if they get to SOI later and they haven't denied that, but if AMD goes bulk forever then I'm not even going to sit here and try and justify AMD continuing with big cores and not letting ARM take over. To me, availability of SOI on smaller nodes is the single largest factor of AMD bringing in FX class chips, not ARM or anything else.

Not so sure - if AMD can hit clocks of 3.8GHz on bulk, that is good enough for an FX chip too. Overclocking would suffer obviously, but I wonder what proportion of FX owners are overclockers?

 

[-Fran-]

http://www.headline-benchmark.com/results/cef25b47-463f-4aa8-a70b-f4a9c516f1c7/9a411141-526d-4452-8e78-1c9891c2efa0

There's with some cheating in the JVM, heh.

Cheers! 😛

 

[Ags1]

Well done! Can you PM me with what you did?

Looking closer, I'm not sure you did anything actually... the only difference is in the results for intermediate threading (max 4 threads) - which I know are pretty unstable - I think you are just surfing the crest of the statistical variability. Here is your two runs compared to each other - everything is the same except the CPU 2,3,4 thread results:

http://www.headline-benchmark.com/results/cef25b47-463f-4aa8-a70b-f4a9c516f1c7/083b462f-9ef2-46f6-b1bb-a3c83c1f9cce

I should probably scrap the intermediate threading category and recalculate all the results on the site :-(

But on that note, look how the FX chips scale over extra threads more smoothly than hyperthreading i7s (at least for integer-heavy threads):

http://www.headline-benchmark.com/results/cef25b47-463f-4aa8-a70b-f4a9c516f1c7/403bbefa-261b-4637-90de-a43fe7c8fc00

I think keeping the intermediate threading category (at least for aggregated CPU results) might show up a genuine architectural advantage of the AMD chip...

 

[tracker45]

This isn't looking very good for AMD.

Is this honestly the best CPU they can make ?

 

[tracker45]

referring to the 7850k

 

[UnrelatedTopic]

http://www.extremetech.com/gaming/172015-leaked-amd-roadmap-shows-excavator-arriving-in-2015-and-possibly-the-end-of-amds-big-core-x86-business


Well its like the end of the line for dCPUs. Damn I just bought my fx6300 and now its the last of its kind. Looks like Vishera is the swansong for traditional CPUs :/

 

[noob2222]

^^ not a bad find, kinda points to what has been circling the rumor mill. GF may be the reason . 1- 20nm node designed only for PPW and thats it? excavator maxing out at 65W, that doesn't look good for high end.

but it’s unlikely to stretch far enough to accommodate the full range of AMD’s product line. That’s probably part of why we see TDPs yanked down to 65W for Excavator.

Are we destined to see another clock speed drop with Carrizo ... This is whats killing the DT sales. lack of any reason to upgrade because its all a downgrade. Intel hasn't improved much over SB and it looks like AMD is going backwards.

 

[palladin9479]

Spoiler

It helps when you close firefox and all my other background stuff I have running.

Ok reposed new results and surprisingly I did better. Was running coretemp and Nvidia's GPU widget and actually paid attention during the benchmark. Integer and FPU are spot on. Graphics performance is trash, Java OpenGL isn't the best language for doing these in. Geometry is accurate but texture and pixel are both severely single thread CPU bound. My one 780 was at 20~25% tops during those benchmarks while a single core was at 100%. I suggest a rewrite of those metrics as there is entirely too much CPU overhead involved with each repetition. Java OpenGL has overhead but not that much. Also found out why the memory transfer was weird, it's a test of your cache subsystem which we can all agree Intel dominates in. I don't think java has the ability to write code that would bypass the CPU's caching to actually test the IMC.

So for that benchmark, CPU performance is a pretty solid test. Graphics is mostly a test of your CPU's single thread ability and memory is a test of your cacheing subsystem. Pretty good as long as people know what each thing does. Also I was using latest 64-bit java 1.7_45.

http://www.headline-benchmark.com/results/403bbefa-261b-4637-90de-a43fe7c8fc00/403bbefa-261b-4637-90de-a43fe7c8fc00

That's what I got. I want to know what gamer did to his setup to get that seeing as its not even being detected right. The software detected my CPU right, memory wrong though(16GB) and I think it's only using one of my 780 Hydro's. Does having SLI enabled degrade it's performance?

My CPU/RAM wasn't detected at all, but it looks like it performed right as far as the benchmarking went. (2600k & 8GB 1333 DDR3) Kinda surprised Yuka's 2700k gets detected and my 2600k didn't though. More confusing, other people's 2600k got detected right... *shrugs* And I know its not OC'd right now (it was at one point, but had a few cooling issues and haven't bothered to re-OC yet), so that probably isn't it either.

I'll toy around tonight and see if I can track down what went on there.

Doesn't make sense, I'm running an fx8350 @4.8Ghz and 16Gb DDR3-1600 yet your getting 2x memory performance. And the weight of the benchmark seems to place inordinate amount of weight on memory performance for pretty much everything. Also the 770 is showing higher then a 780 hydro, hydro's are factory overclocked higher then regular 780's as they expect you to be using a decent WC setup. I'm gonna dig around, I'm thinking something is up with the memory transfer test. I know AMD's IMC is weaker then Intel's but it's not that weak, especially seeing as mine is OC'd to 2.2Ghz.

Blah, forgot my 770 is factory OC'd. Still, I'd imagine a factory OC'd 780 would still win, unless temps/throttling are REALLY coming into play.

Ok reposed new results and surprisingly I did better. Was running coretemp and Nvidia's GPU widget and actually paid attention during the benchmark. Integer and FPU are spot on. Graphics performance is trash, Java OpenGL isn't the best language for doing these in. Geometry is accurate but texture and pixel are both severely single thread CPU bound. My one 780 was at 20~25% tops during those benchmarks while a single core was at 100%. I suggest a rewrite of those metrics as there is entirely too much CPU overhead involved with each repetition. Java OpenGL has overhead but not that much. Also found out why the memory transfer was weird, it's a test of your cache subsystem which we can all agree Intel dominates in. I don't think java has the ability to write code that would bypass the CPU's caching to actually test the IMC.

So for that benchmark, CPU performance is a pretty solid test. Graphics is mostly a test of your CPU's single thread ability and memory is a test of your cacheing subsystem. Pretty good as long as people know what each thing does.

That, would do it. I'm just going to say though, for the GPU tests, a single core on my 2600k never got above ~20% or so, if memory serves.

Still, if anything, this kinda shows how hard it is to write an unbiased benchmark at the OS level. What is looks like is happening is the CPU side of the house (driver side) is actually causing my OC'd 770 to perform faster then Palladins OC'd 780, even in GPU based testing. As for the memory test, well, as Palladin stated, AMD's cache issues are going to KILL them, and there really isn't much you can do to get around this as far as testing goes.

Thanks for the comments. Palladin, I can't explain your findings on the GPU tests. For example, in the shader test the only thing that changes is the number of iterations of a loop in a shader - the geometry and draw calls are the same, so there is no reason why CPU would climb to 100%. In fact as load on the GPU goes up, the CPU should have less load as it has to do its small part less often. CPU should go up to 100% briefly when the shader is compiled, and it is possible the compilation might take longer in later tests (if the compiler 'unrolls' the loop).

Going by the number CUDA cores in your 780 and my 430, I calculate you should be seeing roughly 2-3 times more performance in the shader test.

Perhaps the issue is related to the dual-GPU configuration, or possibly to the fact that the code uses OpenGL 2.0 (for portability) and your GPU driver is doing some CPU-intensive backwards compatibility stuff. It may also be an effect that only appears for very powerful GPUs which I can't see on my more modest systems.

You are right the memory test is primarily a test of the caches - I try to overwhelm the caches by swapping blocks up to 32MB in size, but cache performance still seems to predominate.

That, would do it. I'm just going to say though, for the GPU tests, a single core on my 2600k never got above ~20% or so, if memory serves.

Well, that is what I expect to happen gamerk - this suggests to we are seeing an effect specific to running on Palladin's machine.

Honestly I believe it might be a limitation with Java OpenGL, it's never been particularly good at high performance. One of our major products we provide to customers is a modification of NASA's Whirlwind and it relies heavily on Java OpenGL, we ALWAYS have performance issues with it even on Quadro equipped systems. What version of JRE are you running on the development system? We've seen that effect performance profiles before. As for the cacheing, the only way around it is to transfer very large amounts of memory randomly. 32~128M sizes though anything larger then the L2 cache size will create a noticeably different result.

On a side note, during the texture transfer benchmark many of the textures are constantly blinking and something looks very wrong. I have a feeling that software is doing some of the rendering due to some incompatibility.

 

[blackkstar]

You guys might want to take this with a grain of salt, but I just spoke with someone who works at IBM R&D.

I tried to get what I could out of him but he said in regards to 22nm PD-SOI and IBM and Glofo, "don't ask, it's bad".

He said they are working on things for 20nm FD-SOI but it's a major pain.

I also asked him that as an FX 8350 fan and someone who is enthusiastic about high end AMD CPUs, is the future good or bad. He told me in a year or so I will have bad news for all current gen FX and he mentioned something about AMD's new chip at that time stomping an E5.

He sounded upbeat about things and said he was a fan of parallelism and a fan of AMD.

We keep hearing things from unofficial sources that sound reassuring to high end dCPU AMD but we don't see anything official out of AMD.

I want to believe but I feel it gets more and more difficult as time goes on.

 

[tracker45]

So both the next generation of intel chips and AMD chips are worse than current generations.

 

[blackkstar]

I have a question. If AMD is gonna stop making dCPUs because the desktop market is dying, does that mean they're gonna stop making dGPUs as well?

 

[UnrelatedTopic]

My question is the new APU better than a 6300 CPU wise to merit a platform upgrade? I know you guys battled it out in the last couple of pages but it was painful to read so I just skimmed parts of it. Is the A10 7850k better than my fx 6300 that I should jump ship to FM2+?

Also I doubt it that AMD will quit dGPU for the reason that you still can't match top end performance with current solutions. (APU's)

 

[jdwii]

You guys might want to take this with a grain of salt, but I just spoke with someone who works at IBM R&D.

I tried to get what I could out of him but he said in regards to 22nm PD-SOI and IBM and Glofo, "don't ask, it's bad".

He said they are working on things for 20nm FD-SOI but it's a major pain.

I also asked him that as an FX 8350 fan and someone who is enthusiastic about high end AMD CPUs, is the future good or bad. He told me in a year or so I will have bad news for all current gen FX and he mentioned something about AMD's new chip at that time stomping an E5.

He sounded upbeat about things and said he was a fan of parallelism and a fan of AMD.

We keep hearing things from unofficial sources that sound reassuring to high end dCPU AMD but we don't see anything official out of AMD.

I want to believe but I feel it gets more and more difficult as time goes on.

Thanks for sharing, i can't get any word on FX from my friend but he told me Arm is a side project and a couple of other small things.
To me it just doesn't make much sense from a marketing standpoint. Amd always talked about having more cores vs stronger fewer cores and how much that matters but now it seems like they are not listing and going back to bigger fewer cores at least according to some here and their roadmaps.

At this point their is no way i could recommend anything from Amd at this point from the AM3+ side and quite honestly i'm happy i did not buy the 8350fx and instead got some new headphones.

Will probably come to a day when both Intel and Amd are even in performance. At least Amd is trying to push the performance more Intel is not doing anything same performance lower TDP. Amd is actually trying to push the standards.

 

[noob2222]

The failure to release a product is the only thing killing DT.

@unrelated topic

The 6300 will be better in most if not all cases.

 

[kviksand81]

Fun fact about AMD Desktop roadmap: I can find the server roadmap for 2014 on http://www.amd.com/us/press-releases/Pages/amd-unveils-2013june18.aspx (Yeah... That IS AMD's OFFICIAL homepage... But the Desktop Roadmap for 2014... Nope... Not there... And mark you the detail level about when and what on the server roadmap... The "current" desktop roadmap doesn't have that kind of details... I'm just sayin'...

EDIT:
It is in fact very questionable that the "published" desktop roadmap has anything to do with AMD...? Lots of their slides are available through slideshare.net and no AMD Desktop Roadmap 2014 is to be found here either:

http://www.slideshare.net/AMD/presentations

Nor under documents:

http://www.slideshare.net/AMD/documents

So... Think I'll just wait a little longer and keep running my Phenom II 1100T before I jump to conclusions on whether or not AMD will release an Steamroller based FX-CPU and buy a FX-8350... Next couple of months or so could be interesting...

 

[tracker45]

Steamroller doesn't look very powerful at all.

Very disappointing.

 

[Cazalan]

I have a question. If AMD is gonna stop making dCPUs because the desktop market is dying, does that mean they're gonna stop making dGPUs as well?

No. They want you to buy Kaveri and a dGPU as well.

 

[anxiousinfusion]

I have a question. If AMD is gonna stop making dCPUs because the desktop market is dying, does that mean they're gonna stop making dGPUs as well?

Probably not for a long time. But just as with sound cards, it will eventually come to a point where only an extremely niche minority will actually need dedicated graphics cards.

 

[Cazalan]

Steamroller doesn't look very powerful at all.

Very disappointing.

Powerful compared to what? It's a $150 product.

For the mainstream consumer it will be great. It's not an enthusiast part.

 

[jdwii]

My question is the new APU better than a 6300 CPU wise to merit a platform upgrade? I know you guys battled it out in the last couple of pages but it was painful to read so I just skimmed parts of it. Is the A10 7850k better than my fx 6300 that I should jump ship to FM2+?

Also I doubt it that AMD will quit dGPU for the reason that you still can't match top end performance with current solutions. (APU's)

Your basic question is this will a A10 7850K(is that the name) beat a 6300 in CPU tasks.

Based on my estimates i will continue to say the A10 7850K will have a 15% boost in single core performance compared to the A10 6800K, which is already 10% faster than a 6300 in single core performance but a bit slower in multithreading performance probably around 35% slower. My guess is the A10 7850K will be just about as good as your CPU(6300 3.5Ghz one) in multithreading but around 25-30% faster in single core performance. I can't say if its worth it for you to upgrade from your 6300fx and board that is up to you.

 

[jdwii]

One of Steamrollers improvements was with scaling with PD/BD a module was a 1.6 core with Steamroller should be closer to a 1.8 core. So basicly a 4 core steamroller would have 3.6 scaling and a 4 core PD would have 3.2 core. Take into fact that steamroller has 25% better per clock performance you could add to that number(3.6) and get 4.5. Now take a 6 core PD and do the number and you get 4.6. So at the same clock speed a Steamroller Quad core should give around the same performance as a 6 core PD in multithreading and 25% boost in single core performance.
Also this number would also relate to the A10 7850K being just as good as the I5 2500K in multithreading even the 6300 is. With PD A quad core did not give the same multithreading performance it was closer to a I3 with HT then a full I5.

 

[tracker45]

Is the A10 7850K the best out of the Steamroller range ?

 

[palladin9479]

At this point their is no way i could recommend anything from Amd at this point from the AM3+ side and quite honestly i'm happy i did not buy the 8350fx and instead got some new headphones.

The 63xx series is still amazing for the money, the 8350 is also pretty good provided you actually have a task that wouldn't be done with the 63xx.

The problem I still see people using is they look at absolutes and not realities. Intel's highest performing CPU is better then AMD's higher performing CPU, nobody doubts that. Money wise it's incredibly hard to not recommend the $119 USD FX6300 for lots of builds. The FX8350 itself $170 USD with the FX8320 being $160. The only Intel CPU in that range is the i3 with the occasional low power i5 dipping into the 83xx range. So many people want to do a "well the fx8 can do 8 threads and so can the i7, lets compare them", or "the fx4 is the low end so lets compare it to the i3" and end up with slanted results. When comparing CPU's of the same price range the conclusion changes dramatically.

Just stop and think, I have a $1600 USD graphics solution coupled with a $300 USD case and a $400~500 USD cooling solution. A $170 USD CPU is running all that without any issues and I foresee zero reasons to change it anytime soon.

 

[jdwii]

At this point their is no way i could recommend anything from Amd at this point from the AM3+ side and quite honestly i'm happy i did not buy the 8350fx and instead got some new headphones.

The 63xx series is still amazing for the money, the 8350 is also pretty good provided you actually have a task that wouldn't be done with the 63xx.

The problem I still see people using is they look at absolutes and not realities. Intel's highest performing CPU is better then AMD's higher performing CPU, nobody doubts that. Money wise it's incredibly hard to not recommend the $119 USD FX6300 for lots of builds. The FX8350 itself $170 USD with the FX8320 being $160. The only Intel CPU in that range is the i3 with the occasional low power i5 dipping into the 83xx range. So many people want to do a "well the fx8 can do 8 threads and so can the i7, lets compare them", or "the fx4 is the low end so lets compare it to the i3" and end up with slanted results. When comparing CPU's of the same price range the conclusion changes dramatically.

Just stop and think, I have a $1600 USD graphics solution coupled with a $300 USD case and a $400~500 USD cooling solution. A $170 USD CPU is running all that without any issues and I foresee zero reasons to change it anytime soon.

It gets a lot easier to not recommend those chips when their is no upgrade path.

 

[de5_Roy]

Yes, what is it equivalent to ?

kaveri's igpu is equivalent to radeon hd 7730 gfx card, with ddr3 vram. any entry level gcn card with gddr5 vram can outclass kaveri's igpu.
a10 7850's cpu should be quicker than richlands. but it's a bit harder to predict cpu performance for now. wait till it comes out and gets reviewed.

^^ not a bad find, kinda points to what has been circling the rumor mill. GF may be the reason . 1- 20nm node designed only for PPW and thats it? excavator maxing out at 65W, that doesn't look good for high end.
Are we destined to see another clock speed drop with Carrizo ... This is whats killing the DT sales. lack of any reason to upgrade because its all a downgrade. Intel hasn't improved much over SB and it looks like AMD is going backwards.

it's glass half full or half empty kinda sitch. apus are getting consistent upgrades as speculated when llano, trinity came out.
one issue i've discussed before is that fx was never designed for oems who buy in bulk quantities. it was designed for d.i.y. enthusiasts who, for the most part, ignored fx since zambezi came out (at least, that's what the sales and shipment reports suggest). if amd had launched a chipset like 980G at zambezi launch, they coulda recovered some of the loss from d.i.y. segment.

I have a question. If AMD is gonna stop making dCPUs because the desktop market is dying, does that mean they're gonna stop making dGPUs as well?

highly unlikely. discreet gpus are different from cpus. amd can't shoehorn an fx in a laptop, but they can sell a 7970M with 5750M apu or core i5/i7 cpu. discreet gpus have a far deeper reach into pc ecosystem. hell, if amd starts selling 16core arm cpu, they can just implement pcie bus in the soc, validate it for dt/laptops, and still use radeons. they'll still be able to sell dgpus even if intel snips pcie 3.0 off future intel cpus (unlikely).
moreover, amd's gpu division is highly competitive with nvidia's gpus (discreet gpus), despite amd's continued neglect and abuse (see what they did with r9 290/290x, 7970, vce, early 7850 cards, refusing to acknowledge single card stutter, microstuttering in multi card and so on).

It gets a lot easier to not recommend those chips when their is no upgrade path.

i consider "upgrade path" nearly as mythical as "futureproofing". i see people asking question containing such terms and people recommending parts claiming "upgradability" and "futureproof", but nearly all of them are prone to replace their base cpu+mobo when the "time" actually comes. accidents are exeption to this. in case of accidents, upgrades become viable.... until you face the facts that both intel amd amd do not offer significant upgrades over their respective high end lineup. all you really have are sidegrades or downgrades within respective brands.
the "performance for money" and "performance per watt for money" are still much more viable.