AMD CPU speculation... and expert conjecture

Page 423 - Seeking answers? Join the Tom's Hardware community: where nearly two million members share solutions and discuss the latest tech.
Toggle sidebar Toggle sidebar
Status
Not open for further replies.
D

de5_Roy

Splendid
Oct 14, 2011
7,159
0
26,960
juggernautxtr

juggernautxtr

Honorable
Dec 21, 2013
101
0
10,680

glofo failing again is that a surprise?
maybe glofo should just start selling chunks to Tsmc.


 
C

Cazalan

Distinguished
Sep 4, 2011
2,672
0
20,810


Well AMD nearly doubled the transistor count and only lost 10% speed at nearly the same TDP. That's actually quite good.

AMD is betting the farm on HSA so for their sake I hope the software quickly follows. Otherwise they could have done a simple die shrink and made some much smaller/cheaper parts. They're getting some incredible density with that node.

Ivy 1.4B transistors, 160mm^2, 8.75M transistors per mm^2
Kaveri 2.41B transistors, 245mm^2, 9.84M transistors per mm^2

So GF's 28nm process is actually more dense than Intel's 22nm process.
 
etayorius

etayorius

Honorable
Jan 17, 2013
331
1
10,780
http://a.disquscdn.com/uploads/mediaembed/images/781/216/original.jpg

Something really Fkd up with the Pudget "Benchmarks", considering all the improvements AMD did to the to the Fetch, the Integrer Scheduler the L1 Cache and the addition of a Second Decoder to the modules inside Kaveri it seems really really strange Richland beating Kaveri in most Benchmarks.

I never heard of Pudget Before, but i read they were tied to very close to Intel in the past but more recently they seemed to be selling AMD Systems, perhaps Intel did Bribe them? it would not be the first time...
 
gamerk316

gamerk316

Glorious
Jul 8, 2008
13,255
149
44,840


Pudgets been one of those second tier PC makers for a long time now. Typically don't think of them, but I know the name.

As for the improvements to the frontend: What you fail to realize is, for a given benchmark, if those components that have been improved are not driving performance, then the improvements to those components will lead to next to no improvements.

EG: If you upgrade the component that is responsible for 5% of the performance by 30%, you get much less improvement then if you upgraded the other 95% by 2%.

For anything not L1 dependent, is not branch heavy, and does not use a lot of threads, you will not see as large performance gains. BTW, games are a good example of such an application. I'd imagine most of the improvement seen in the benchmarks were GPU, not CPU, side. Also don't forget the 300MHz, or 10% clock speed reduction.
 
gamerk316

gamerk316

Glorious
Jul 8, 2008
13,255
149
44,840
Not *without question*; I noted some of the results could be understated due to FPS limits. OS used (7 or 8) could be mucking results. And I've been stressing this PARTICULAR result set. And I've been trying to explain how, even with the stated changes in design, that result set could occur. Now we wait for more benchmarks and see if the pattern holds.
 
D

de5_Roy

Splendid
Oct 14, 2011
7,159
0
26,960

that sounds good... now i'm worried about yields. is the transistor count so high because the gpus can pack in parts in higher density than the cpu cores?

i'll still knock glofo :p, took them 2 years to get kaveri out. if their 28nm shp can squeeze that many transistors per unit area, then amd should follow up kaveri with a cpu release, even as experiment. or an 8 core jag/puma running a 1024 core igpu!!!
 
D

de5_Roy

Splendid
Oct 14, 2011
7,159
0
26,960

that's the plan.
but it'll require software support.
iirc trinity was supposed to do that for video editing, use the igpu as well as discreet gpu.
 
J

jdwii

Splendid
Apr 13, 2010
3,790
18
22,965
Not to much longer and we will finaly see who was the right one after all for now i was just as wrong and i over estimated the performance unless all the benchmarks are false, and i never put much in Intel or Amd benchmarks when they do them their self's. I want to see reviews from tomshardware, anandtech, and many more.
Lets remember juanrga performance figures which where 30% average for GPU benchmarks and 20% average for CPU benchmarks

I stated 25% on GPU benchmarks and 15% on CPU benchmarks so i was probably just as wrong. I'm really impressed by the CPU benchmarks and if juanrga is right about Amd not producing 8 cores anymore then we really are going to be having I3 level CPU performance from Amd for now on. My friend from Amd will not really tell me much.
One thing we should add on the forum BS post is talk about when a company is real quiet about a product never really means anything good definitely coming from Amd, 1 time i remember Intel being really quiet and that is when they finally beat the Athlon, but usually if Amd has a killer product they try and sell it really fast for example the 7970.

Also keep in mind that for almost every launch the fabrication process is heavily limiting Amd, this started with BD and now we have it with this processor. Odd that Amd won't just leave GF for TSMC i doubt they ever will. But then again i doubt GF will ever be competitive to even TSMC. Their has to be some kind of internal friendship still going on.
 
etayorius

etayorius

Honorable
Jan 17, 2013
331
1
10,780


The extra Decoder inside each Module by itself should be enough to gain at least a 10-15% increase, i would go as far as to to say the extra Decoder is by far the largest performance improvement from all of the changes in Steamroller Arch, i can understand the GPU Benchmarks since AMD has probably not released nor provided any Optimized Drivers for Kaveri, but the CPU benchmarks are just horrible if they are found to be true (i seriously doubt it), Steamroller main focus in CPU was to greatly improve IPC or Single Core performance, but now you`re suggesting that all the improvements AMD did to Steamroller in order to increase IPC are basically useless because as you claim "they are not driving performance", with this argument it would mean AMD are just clueless on what improves performance on their own Arch.

We know for a Fact that Piledriver is AT least 10-12% faster than Bulldozer in Single Core Exection and up to 20 Faster mostly in Multithreaded, AMD promised 20% more IPC by Ehnancing the Arch on Steamroller and another 10% out of pure higher Speed (30% target), take the 300Mhz less Speed compared to Richland and you still get the 20% more IPC with Kaveri, now how on earth is Richland overall faster than Kaveri in the Pudget Benchmarks?

Lets take your Best Case scenario of up to 10-15% more performance like you previously predicted several pages back, your predictions fell too generous compared to Pudget Benchmarks it would mean Kaveri is about 10% Slower than Richland, there for AMD is lying with their 20% more IPC compared to Richland using bad faith and false advertising for a product they may know it will fail, then inflated the price of 175USD compared to Richland 140USD, then to make all this short it means AMD wants to rip all of us with an inferior and more expensive Product.

But now seriously, we have no reason to trust Pudget or the hand behind them, seeing as how small they are, they would have no issues getting rid of AMD products and having a nice Pay Check from Intel, they have nothing to lose from backstabbing AMD in the back.

Even if it is 10% best case scenario compared to Richland it will still be faster than my PhenomII 980 at Default Clock, i get a nice GPU to the levels of the HD5770 that will boost like crazy using MANTLE and HSA in the future and i will also get a Key for Battleshit4 on Origin bundled with my APU, which i can sell easily for 30-40 Bucks... not too bad at the end of the day.

On a different note, my PhenomII 980BE C3 is one of those CPUs that perfroms about 2% slower than average, i noticed that my stock clock is 3691 Mhz compared to others exact same models, i saw one 980 default clock at 3712 and seen others at exact 3700, also my PhenomII is not OC Friendly... i can get it to 4Ghz but it will freeze once every two weeks, i been checking the temps and i got a CoolerMaster and never seen it hit more than 58C, also the default Voltage is 1.40 and i find it kinda high at default 3.7Ghz, i just got one of the bad ones i guess... so i believe when AMD claims up to 20% they really mean it should be about at the 20% mark with a 1-2% margin of error due to Fabrication of different batches of CPU like in my PhenomII, at least i can manually raise that 2% i am losing at stock, by raisin a tad bit on the Bios, i have found that my PhenomII sits better at 3.9Ghz without issues... but there is that feeling that pushes me to set it at 4 even though i know it may fail once or twice a month.
 
J

juanrga

Distinguished
BANNED
Mar 19, 2013
5,278
0
17,790
de5_Roy your 'experiment' with a "1024 core igpu" doesn't make sense. 512 cores is about the maximum allowed by DDR3.

Embra No. The first reason is AMD change of strategy in servers and the abandon of Opteron line due to architectural and market reasons (explained lots of times in this this thread). When the server roadmap was released at early 2013, some of us said that implied no FX-Steamroller. We explained why. We were ignored, and people here said we would wait for the desktop roadmap.

The 2013 desktop roadmap was released and showed no FX-Steamroller, but people here said we would wait for the 2014 desktop roadmap.

The 2014 desktop roadmap was released and showed no FX-Steamroller, but people here said we would wait for the 2015 desktop roadmap.

The 2015 desktop roadmap was released and showed no FX-Steamroller, but people here now says that we would wait for the 2016 desktop roadmap...

griptwister It has been explained dozens of times in this thread that main goal of AMD behind APUs is not cheap gaming (as some still believe) but compute. Berlin (Kaveri version for servers) will be used for compute, the iGPU will be not used to play games in a server ;-)

During APU13, it was mentioned that MANTLE has asymmetric GPU support. It was explained that the discrete GPU could do rendering of the scene, whereas the iGPU in the APU could do the post-processing, for instance. In this case a A10-7850K would be more powerful than a i7-4770k.

I suppose that this is another example of this thread going in circles again and rediscovering stuff explained before.
 
etayorius

etayorius

Honorable
Jan 17, 2013
331
1
10,780


Why does AMD not use some sort of Sideport Memory for Kaveri? it would had increase the bandwidth.
 
J

jdwii

Splendid
Apr 13, 2010
3,790
18
22,965
Juanrga facts are starting to fail, look he has even resorted to stating things such as " that main goal of AMD behind APUs is not cheap gaming" note that this is used to overcome his information previously about steamroller being 20% more powerful on average(which he used to say 30% but will deny it) and 30% more IGPU performance.

right before Bulldozer came out JF-AMD started to state that IPC and single core performance did not matter. Maybe he did not know that Intel did not give us 1 core for 220$ but 4 which most software used anyways and according to Amd's roadmaps will only be produced from Amd.
Like i said its about detecting BS and fanboyism, The older you get and the more your on fourms the more you will notice this.
 
jimmysmitty

jimmysmitty

Champion
Oct 5, 2007
21,604
599
59,140


Of course this would require the software to utilize Mantle, which as of yet has not even been released for the GPUs it was designed for.

I can see the benefit of it but I think its too much as most professionals pay top dollar for hardware dedicated to the purpose and what will matter is much like the gaming aspects of the APU. I remember when it was found that a i3 3000 series with a HD7770 was the same price as the then top end APU from AMD but performed better in games of course.

Price and if the software can use it and as of right now I don't see software maxing out quad cores for a few more years.

And jdwii, he is not far off. The APU was not designed with just cheap gaming in mind. They want HSA (what he is talking about with the CPU and GPU doing the work together) and as well are replacing entry and mainstream level builds with it. You can't find a mITX AM3 mobo or microATX AM3 mobo with more than 2 slots of RAM. Mainly FX is to be like Intels LGA2011. Top of the line. FM2 is to be everything else and I would not be surprised if they don't kill even top end and use only APUs.

Did you read what I read from Rory Reed? When he stated desktops have more than enough power? He isn't interested in giving you a better CPU than Intel. He is focusing on whatever will make them money instead which is sad as now Intel has no reason to keep their CPUs pushing as fast as they were. When Core 2 hit it was massive. Then Nehalem jumped again. Then SB was even better. But IB and Haswell are mostly not that great except in certain situations and now Broadwell is not supposed to be out until the 2H of 2014 when every step was supposed to be Q1 of the year.

As I said, AMD is focusing on money more than competing with Intel.
 
Reynod

Reynod

Administrator
Mar 26, 2007
14,921
16
42,215


Let me tell you the same thing I told the last guy ... stop trolling or there will be consequences.

Your deliberately trying to provoke another user and that will not be tolerated.

 
J

juanrga

Distinguished
BANNED
Mar 19, 2013
5,278
0
17,790
etayorius, sideport memory allowed ATI to bypass the CPU memory controller to access system RAM and increase performance. Kaveri has hUMA. From a die show of Kaveri shown at CES, Hans (cpu architect) was able to guess that the memory controller is twice so large as in Trinity/Richland. There was some speculation at S|A forums about if it did mean a quad-memory channel, but it seems to be that final answer is hUMA. The double controller (one for CPU and another for GPU) makes something similar to sideport memory.

jdwii I don't know what are you doing or trying. If you re-read past messages, I have been saying that AMD main emphasis on APUs is compute for months. In fact I used the same arguments before: Berlin APU is not for cheap gaming it is for heterogeneous compute in servers. Several people here rejected the idea, but this is not anything new. AMD has been saying that they want to use the iGPU as a kind of giant FPU since they presented Fusion many years ago.

jimmysmitty Of course, you need software support! Note that the claim about using the iGPU for post-processing was made by an Oxide developer during an APU13 talk.
 
gamerk316

gamerk316

Glorious
Jul 8, 2008
13,255
149
44,840
The extra Decoder inside each Module by itself should be enough to gain at least a 10-15% increase, i would go as far as to to say the extra Decoder is by far the largest performance improvement from all of the changes in Steamroller Arch, i can understand the GPU Benchmarks since AMD has probably not released nor provided any Optimized Drivers for Kaveri, but the CPU benchmarks are just horrible if they are found to be true (i seriously doubt it), Steamroller main focus in CPU was to greatly improve IPC or Single Core performance, but now you`re suggesting that all the improvements AMD did to Steamroller in order to increase IPC are basically useless because as you claim "they are not driving performance", with this argument it would mean AMD are just clueless on what improves performance on their own Arch.
Click to expand...

Different types of application have bottlenecks in different parts of the CPU. Taking your example, if the application's performance is not driven by the decoder module, then improvements to that area aren't going to lead to significant performance increases for that particular application.

Lets take games as an example: Do you really think the L1 cache is driving performance? No, the larger L2/L3 are more important. So improvements to the L1 aren't going to have significant effects to performance, even if the improvements to the L1 are quite large.

Combine that with a clockspeed decrease of 300MHz, and you could easily see why some applications would offer virtually the same performance, even if several parts of the CPU were improved up to 30%.

When I look at the improvements AMD made, I see improvements that will reduce some of the overhead to CMT, and that will make applications that scale across all the cores run better. For applications that don't scale as well, such as games, I really don't see much in the way of CPU performance increases. I'd suspect those apps are going to be driven by improvements to the iGPU, not CPU.

We know for a Fact that Piledriver is AT least 10-12% faster than Bulldozer in Single Core Exection and up to 20 Faster mostly in Multithreaded, AMD promised 20% more IPC by Ehnancing the Arch on Steamroller and another 10% out of pure higher Speed (30% target), take the 300Mhz less Speed compared to Richland and you still get the 20% more IPC with Kaveri, now how on earth is Richland overall faster than Kaveri in the Pudget Benchmarks?
Click to expand...

"In some tasks". AMD's words, not mine. See my above point: If a particular part of the CPU is not driving performance for a particular app, that app should not expect to see any performance increases by increasing the performance of that part of the CPU.

For example: Say AMD made the cost overhead of CMT 5%, rather then the 20% PD has. So you'd say "15% more IPC", and thus, 15% more performance. But for tasks that never use CMT [one core per module being utilized], this 15% performance increase to CMT has no performance impact on that task. Combine that with a clock decrease, and you'd actually expect performance to drop due to lower clocks and no change to IPC.

Lets take your Best Case scenario of up to 10-15% more performance like you previously predicted several pages back, your predictions fell too generous compared to Pudget Benchmarks it would mean Kaveri is about 10% Slower than Richland, there for AMD is lying with their 20% more IPC compared to Richland using bad faith and false advertising for a product they may know it will fail, then inflated the price of 175USD compared to Richland 140USD, then to make all this short it means AMD wants to rip all of us with an inferior and more expensive Product.
Click to expand...

Remember the BD launch? When the PII X4 965 BE was beating the FX-4100 head to head in the majority of benchmarks? Its happened before, just two years ago.

I don't think it will be that bad. There aren't many pure single-threaded applications, so the number of apps that lose performance is going to be vanishingly small. Also remember that games are driven more by GPU performance, and its possible driver side issues on the GPU side could be understating the results. Simply one data set, and we'll see where the others line up.

But now seriously, we have no reason to trust Pudget or the hand behind them, seeing as how small they are, they would have no issues getting rid of AMD products and having a nice Pay Check from Intel, they have nothing to lose from backstabbing AMD in the back.
Click to expand...

And...we're back to conspiracy theories. Rather then try and explain the results [poor CPU design, shoddy drivers, issues with the OS], we fall back to bias instead. And people wonder why politics in this country is so bad right now...

Even if it is 10% best case scenario compared to Richland it will still be faster than my PhenomII 980 at Default Clock, i get a nice GPU to the levels of the HD5770 that will boost like crazy using MANTLE and HSA in the future and i will also get a Key for Battleshit4 on Origin bundled with my APU, which i can sell easily for 30-40 Bucks... not too bad at the end of the day.
Click to expand...

Notice I've never been one of the people claiming that a PII CPU would beat Kaveri? That's just sad people are still claiming as such. For BD, ok, but that hasn't been true since PD came out. As for Mantle, I've made my opinions on the API very clear by now.
 
Status
Not open for further replies.

TRENDING THREADS

Latest posts

Moderators online

Share this page

COMPANY
RESOURCES
FOLLOW
Top Bottom