AMD CPU speculation... and expert conjecture

[JAYDEEJOHN]

That may end up true, but even Apple is having to readjust their volume orders.
And its just not the economy, competition is a good thing

 

[cgner]

Apple is getting hit lately, their stocks depreciate and products dont sell as well as they had hoped. Like jay said tho, competition is good. Really good. Others are getting bigger shares of the market.

 

[Cazalan]

Apple is getting hit lately, their stocks depreciate and products dont sell as well as they had hoped. Like jay said tho, competition is good. Really good. Others are getting bigger shares of the market.

The stocks have taken a hit (like any bubbles needing to burst) but they're also selling more product quarter over quarter. They still sold 1/2 a billion iOS devices last year. Staggering!

 

[Chad Boga]

After seeing the article JAYDEEJOHN posted thinking not much more maybe 10% at most from the clock speed increase(since its using the same design as radeon 6000 series looks like Amd doesn't even care this time around) take into fact that most laptop manufacturer don't buy high speed ram its going to be a troll release yes chad i used that word again.

5% boost in CPU performance on average and 10% boost in GPU performance. Wow just wow.

So below the performance of a GeForce 640?

 

[de5_Roy]

intel always benchmarkets performance improvements. the difference with amd is that intel can and has gotten away with it. multiple times. not that benchmarketing of either company hasn't been widely criticised. iirc the gpu that haswell gt3 was competing with was a gt650M, that's not even a dt gpu. there were so many details hidden away in that demo that i suspect it's validity. at least trinity's performance isn't suspect.
i think that the claimed performance improvement will come from better hybrid cfx support, new ram support, better yields on a mature process, more mature drivers and other supporting softwares. most of the richland p.r. seemed to be laptop and software focused.
i might very well be wrong about this - i noticed in a chinese(which i can't read at all) article, richland being mentioned with radeon 77xx cards. i thought that amd will add hybrid cfx support for those cards. woulda been nice if it was true.

 

[sarinaide]

We were bored and started to go through the "Core" "No core" arguments again running various test;

FX8350, bios recognizes CPU Cores#1-8 all can be disabled and isolated running independantly of the other core in the module as well as the other modulized cores. In the OS, a single core where #1 through to 8 is recognized as a single core, single thread. It suggest that AMD doesn't approach SMT using virtualization in the manner that HT does.

3770K, recognizes cores #1-4, cores 5-8 are non existent and cannot opperate a HT virtualized core only physical cores. But this is old news.

It is difficult to test HT vs CMT, HT will recognize 1 core/2 threads while the only way to make 2 threads on a AMD part is to enable 2 cores which in theory is still correct. In such senario most results show AMD's CMT to be around the true performance of native dual core, while Intels results are somewhat different, less physcal cores active affects HT perfomance, 2 cores/4threads is faster and so forth but nothing remotely close to a true native dual cores performance.

AMD have come out on numerous occasions to explain the nature of its module arch, while not native x86 cores, they still operate as independant cores just with shared resources. As to the performance hit, yes that is evident but probably down to to much shared space rather than dedicated resources. I think AMD's thinking was to lower resources aggressively to lower power usage, but had the converse effect.

 

[sarinaide]

intel always benchmarkets performance improvements. the difference with amd is that intel can and has gotten away with it. multiple times. not that benchmarketing of either company hasn't been widely criticised. iirc the gpu that haswell gt3 was competing with was a gt650M, that's not even a dt gpu. there were so many details hidden away in that demo that i suspect it's validity. at least trinity's performance isn't suspect.
i think that the claimed performance improvement will come from better hybrid cfx support, new ram support, better yields on a mature process, more mature drivers and other supporting softwares. most of the richland p.r. seemed to be laptop and software focused.
i might very well be wrong about this - i noticed in a chinese(which i can't read at all) article, richland being mentioned with radeon 77xx cards. i thought that amd will add hybrid cfx support for those cards. woulda been nice if it was true.

While the focus is Richland mobility that is expected, AMD don't yet have a top contender in the Mobile market and A-Series is suited for Mobility as well as the fact that Richland will release on Mobile before it releases on desktop. That said it is not that AMD don't have a vested interest in Richland DT doing very well. AMD are very proud of the A-Series parts, that is despite some early calous reviews on Trinity, it remains the best low cost platform and said before even discrete level graphics achieve 70FPS in BF3 64man map with lots of eye candy, that for a cheap processor is just fantastic value for money, add the Hudson platform which is full of fun and value this is potentially AMD's future bread and butter.

 

[jdwii]

So below the performance of a GeForce 640?

it would seem so based on the design and speeds reported. Probably around 6670 level. A 630M will probably have around the same performance on the laptop area. 6670 will probably be just as good as the desktop APU. Probably why Amd doesn't make anything under the 7750.

 

[sarinaide]

What all the fud on HD-D graphics suggest is same core counts just improved GCN architecture and faster Memory performance, could be quite a nice little bump for Trinity without needing a socket change. I would hazzard to guess any numbers but going on what Lower powered GCN parts did to bigger VLIW4 parts was almost double performance from the entry level point, at the same price point the HD7970 is more than double the performance of the HD6970. Going down to integrated and mobile level I am pretty sure GCN efficiency will render around 10-20% improvement on less power than that of the VLIW parts, I would guess with the Memory performance bump include could be around 15% which is considerable could be around 10-12FPS improvement at stock.

 

[cgner]

intel always benchmarkets performance improvements. the difference with amd is that intel can and has gotten away with it. multiple times. not that benchmarketing of either company hasn't been widely criticised. iirc the gpu that haswell gt3 was competing with was a gt650M, that's not even a dt gpu. there were so many details hidden away in that demo that i suspect it's validity. at least trinity's performance isn't suspect.
i think that the claimed performance improvement will come from better hybrid cfx support, new ram support, better yields on a mature process, more mature drivers and other supporting softwares. most of the richland p.r. seemed to be laptop and software focused.
i might very well be wrong about this - i noticed in a chinese(which i can't read at all) article, richland being mentioned with radeon 77xx cards. i thought that amd will add hybrid cfx support for those cards. woulda been nice if it was true.

It feels like the "review" websites are the ones doing all the benchmarketing while AMDs old marketing team did its own.

 

[truegenius]

We were bored and started to go through the "Core" "No core" arguments again running various test;

FX8350, bios recognizes CPU Cores#1-8 all can be disabled and isolated running independantly of the other core in the module as well as the other modulized cores. In the OS, a single core where #1 through to 8 is recognized as a single core, single thread. It suggest that AMD doesn't approach SMT using virtualization in the manner that HT does.

3770K, recognizes cores #1-4, cores 5-8 are non existent and cannot opperate a HT virtualized core only physical cores. But this is old news.

It is difficult to test HT vs CMT, HT will recognize 1 core/2 threads while the only way to make 2 threads on a AMD part is to enable 2 cores which in theory is still correct. In such senario most results show AMD's CMT to be around the true performance of native dual core, while Intels results are somewhat different, less physcal cores active affects HT perfomance, 2 cores/4threads is faster and so forth but nothing remotely close to a true native dual cores performance.

AMD have come out on numerous occasions to explain the nature of its module arch, while not native x86 cores, they still operate as independant cores just with shared resources. As to the performance hit, yes that is evident but probably down to to much shared space rather than dedicated resources. I think AMD's thinking was to lower resources aggressively to lower power usage, but had the converse effect.

they should have tried a die shrink of k10 to save die size :/ (other than defective apu)
k10 already showed good overclocks in form of phenom2 and better multithreading performance

 

[sarinaide]

they should have tried a die shrink of k10 to save die size :/ (other than defective apu)
k10 already showed good overclocks in form of phenom2 and better multithreading performance

K10 only looks appealing because Zambezi was by and large the same but moreso because there is this crusade for single threaded x86 performance which Zambezi didn't do well at other than that in multithreaded performance the Zambezi's about 98% of the time outright crucified the outgoing Thubans sometimes considerably. Fast forward to Vishera and not only does Vishera put Zambezi 6ft under, it by and large buries Thuban and K10 for good, A game like Skyrim which absolutely taxed the 1100T and 8150 sees the 8350 gain around 12FPS over the prior parts, in content creation K10 cannot hold a candle to the new architecture.

It would have been very easy for AMD to flog the k arch gradually improve x86 performance, but we knew in 2006 that AMD were no longer interested in flogging x86 into bedrock. The new approach while anything new goes through teething HSA is considerably faster than x86 alone and that is the bread and butter that may make AMD the champion chip in the next few years.

http://techreport.com/review/23750/amd-fx-8350-processor-reviewed

In about 99.9% of runs K10 is being buried under the heavy hand of irrelevence, it is probably the reason why AMD are where they found themselves persisting in a dead end or cul de sac. If you choose to be revolutionary its a gamble but if k10 had to continued AMD would have sunk by now. I think its time to move on.

 

[jdwii]

What all the fud on HD-D graphics suggest is same core counts just improved GCN architecture and faster Memory performance, could be quite a nice little bump for Trinity without needing a socket change. I would hazzard to guess any numbers but going on what Lower powered GCN parts did to bigger VLIW4 parts was almost double performance from the entry level point, at the same price point the HD7970 is more than double the performance of the HD6970. Going down to integrated and mobile level I am pretty sure GCN efficiency will render around 10-20% improvement on less power than that of the VLIW parts, I would guess with the Memory performance bump include could be around 15% which is considerable could be around 10-12FPS improvement at stock.

Did you read that its based on the radeon 6000HD series not the new GCN?

 

[sarinaide]

Did you read that its based on the radeon 6000HD series not the new GCN?

Read the Techreport one now, makes sense as Kaviri was designated to make the step up. That said it will be interesting to see the step up, after all Vishera was the same as Zambezi yet the performance deltas where on the oposite ends of the room, small changes can affect big results.

 

[truegenius]

phenom 2 have room to improve imc, (~19 vs ~14)

1100t lacks new instructions but still managed to score good points

x4 980 is only 5 fps behind 8350 regardless of its lower clock no turbo and half the amount of cores
i3-3225 is still competitive ! looks like mulit tasking test is not that much heavy :/

The Panorama Factory photo stitching

not related to this but still mine old 800mhz scorpion based android can process 2mp image in under 10seconds
actually, my phone have only 0.3mp camera sensor so using sweep panaroma i can take larger picture (can easily create 2mp photo by swiping)

4.5GHz, 1.5375V - OK, ~56C

phenom2 too can hit ~4.3ghz stable (amd overclocking club )


conclusion
8350 is faster than phenom 2 x4 980 and 1100t
but why ?
because
2 extra cores
better memory controller
significantly higher clock and better turbo
new instructions
smaller node (comes handy to reduce power consumption)

so if we compare apple to apple (6300 vs 1100t on same node and clock) then k10 is still competitive with new buldozzer and even piledriver

i acctually don't like cmt

and fm1 k10 are handicaped apu which can not hit 4ghz+ which is piece of cake for phenom 2, and fm1 settup costs more in comparision to am3(+) (in my place)
so that is why i hate fm1 k10
and want to have real k10 at 32nm on am3(+) platform

my guesstimations
65nm----->>>>>45---->>>>32
125w---->>>>>125--->>>>125
phenom-->>phenom2-->>phenom3
100points---->>~140---->>~175 (expected relative per core performance for top models)

 

[gamerk316]

We were bored and started to go through the "Core" "No core" arguments again running various test;

FX8350, bios recognizes CPU Cores#1-8 all can be disabled and isolated running independantly of the other core in the module as well as the other modulized cores. In the OS, a single core where #1 through to 8 is recognized as a single core, single thread. It suggest that AMD doesn't approach SMT using virtualization in the manner that HT does.

3770K, recognizes cores #1-4, cores 5-8 are non existent and cannot opperate a HT virtualized core only physical cores. But this is old news.

It is difficult to test HT vs CMT, HT will recognize 1 core/2 threads while the only way to make 2 threads on a AMD part is to enable 2 cores which in theory is still correct. In such senario most results show AMD's CMT to be around the true performance of native dual core, while Intels results are somewhat different, less physcal cores active affects HT perfomance, 2 cores/4threads is faster and so forth but nothing remotely close to a true native dual cores performance.

AMD have come out on numerous occasions to explain the nature of its module arch, while not native x86 cores, they still operate as independant cores just with shared resources. As to the performance hit, yes that is evident but probably down to to much shared space rather than dedicated resources. I think AMD's thinking was to lower resources aggressively to lower power usage, but had the converse effect.

Its important to understand HOW HTT and CMT are different.

HTT is basically the bare bones approach to SMT. You get an extra register context, but not much else. Very small performance gains, but very cheap to implement. Also, because the front end is not duplicated, it is impossible to run some tasks on a HTT core.

CMT is basically the reverse: Most of the front end resources are duplicated (scheduler and a few other shared pieces aside). As a result, a CMT core, by itself is the same as a full core. The issue becomes when BOTH cores of a BD module are used, the shared front end effects a performance hit (~20% or so).

 

[sarinaide]

Its important to understand HOW HTT and CMT are different.

HTT is basically the bare bones approach to SMT. You get an extra register context, but not much else. Very small performance gains, but very cheap to implement. Also, because the front end is not duplicated, it is impossible to run some tasks on a HTT core.

CMT is basically the reverse: Most of the front end resources are duplicated (scheduler and a few other shared pieces aside). As a result, a CMT core, by itself is the same as a full core. The issue becomes when BOTH cores of a BD module are used, the shared front end effects a performance hit (~20% or so).

Thats what I was implying by AMD attempting to use to restricted a front end, to little shared resources and narrow instruction exectution leads to the hit, Vishera limited it to a degree lets hope that the SR front end reduces the performance penalty further.

 

[sarinaide]

http://techreport.com/r.x/amd-fx-8350/stream.gif
phenom 2 have room to improve imc, (~19 vs ~14)

http://techreport.com/r.x/amd-fx-8350/aida-mandel.gif
http://techreport.com/r.x/amd-fx-8350/aida-julia.gif
1100t lacks new instructions but still managed to score good points

http://techreport.com/r.x/amd-fx-8350/multi-fps.gif
x4 980 is only 5 fps behind 8350 regardless of its lower clock no turbo and half the amount of cores
i3-3225 is still competitive ! looks like mulit tasking test is not that much heavy :/

The Panorama Factory photo stitching

not related to this but still mine old 800mhz scorpion based android can process 2mp image in under 10seconds
actually, my phone have only 0.3mp camera sensor so using sweep panaroma i can take larger picture (can easily create 2mp photo by swiping)

4.5GHz, 1.5375V - OK, ~56C

phenom2 too can hit ~4.3ghz stable (amd overclocking club )


conclusion
8350 is faster than phenom 2 x4 980 and 1100t
but why ?
because
2 extra cores
better memory controller
significantly higher clock and better turbo
new instructions
smaller node (comes handy to reduce power consumption)

so if we compare apple to apple (6300 vs 1100t on same node and clock) then k10 is still competitive with new buldozzer and even piledriver

i acctually don't like cmt

and fm1 k10 are handicaped apu which can not hit 4ghz+ which is piece of cake for phenom 2, and fm1 settup costs more in comparision to am3(+) (in my place)
so that is why i hate fm1 k10
and want to have real k10 at 32nm on am3(+) platform

my guesstimations
65nm----->>>>>45---->>>>32
125w---->>>>>125--->>>>125
phenom-->>phenom2-->>phenom3
100points---->>~140---->>~175 (expected relative per core performance for top models)

Floating point opperands as we already know with AMD's first module arch suffer the performance hit, K10 is the end of a generation of arch evolution, Zambezi and Vishera represent the first phase with Stream roller third phase and Excavator 4th phase from AMD's own interpretation, in each arch evolution changes to the front side are made so if by SR the differential between the Thuban and SR parts in FPU operations increases further then it can be said that already modulation is more efficient than native cores.

Another thing which does happen is the FX4300 and A10/A8 parts beat the 1100T and 980BE so its not purely down to core count. PhenomII had its day but it was time for AMD to get proactive lingering on K6/K7 design was eventually going to end up in AMD falling behind and they did.

 

[mayankleoboy1]

conclusion
8350 is faster than phenom 2 x4 980 and 1100t
but why ?
because
2 extra cores
better memory controller
significantly higher clock and better turbo
new instructions

AVX, FMA, BMI, are all gimmicky (at this point). I doubt any real software uses any of these instructions.
Maybe AVX is only 90% show, 10% real world usage, because it works on floating-point.


AVX2 of Haswell should speed up a lot of software, because it works on Integers.

 

[esrever]

Steamroller and kaveri is still at Glofo

 

[gamerk316]

AVX, FMA, BMI, are all gimmicky (at this point). I doubt any real software uses any of these instructions.
Maybe AVX is only 90% show, 10% real world usage, because it works on floating-point.


AVX2 of Haswell should speed up a lot of software, because it works on Integers.

Most SSE/AVX, or any other CPU opcodes are used only when the compiler inserts them during optimization. There are VERY few circumstances where your data is formatted in the right way to use these instructions, AND the coder knows to use them.

 

[noob2222]

The problem is Amd is not in their own little world with a release from Intel coming out this is one horrible upgrade and yes i'd like to know if Amd did say 20-40% boost(marketing guys should NEVER use numbers they need to go back to marketing i know this as i've read lots of marketing books) and no one cares about 1 test i care about all of them on average. Again their is know reason why they didn't put a better GPU besides a little speed bump, not to mention with the rumors about IGPU from Intel, Amd should be really trying to make a better GPU this time around or at least use the one from the radeon 7000HD series. This should be called Trinity 1.1 because that's all were getting. With Trinity I could recommend Amd again and have people get a A10 vs a I3 as a thought it was better overall now with haswell i don't know have to wait to see benchmarks. If Intel is right when saying their next GPU is on par with a gtx 650M then Amd has issues since even a 5670 is better then trinity.

The 650 is on par with a 6850

Hope Intel is pulling a Amd and exaggerating their numbers as well.

First off, your reading wayy too much into the GT3E

A highly forgiving DirectX 11-generation title, DiRT 3, was made to run on both GPUs. Visitors noted that even if not as smooth as the GTX 650, Intel's Haswell iGPU did produce playable frame-rates.
http://www.techpowerup.com/178741/Intel-quot-Haswell-quot-GT3-Graphics-Twice-as-Fast-as-quot-Ivy-Bridge-quot-.html

full of typos, according to Anandtech, its gt3E and 650M.

so first things first ... dirt 3 ...

The hd 4000 is already capable of somewhat ok performance in that game. doesn't take much more to push the 60 fps limit

so why not something else?

because if you double the performance of crap, its still crap. Also notice the 7660D is pushing close to a 640 already.

secondly this is the gt3E

With Haswell, the ULV part with the Crystallwell memory costs more to make, but not a huge amount more. Think $5 or less for the DRAM, and maybe double that tally for construction losses. For this, Intel is trying to charge OEMs $50-60 on a CPU that already breaks the bank at $225 or so. Several OEMs contacted SemiAccurate recently to complain about Intel’s pricing, one calling it, “Absurd”, and that was the nice one.
http://semiaccurate.com/2012/10/03/oems-call-intels-haswell-pricing-absurd/

bare price for gt3E? $300 range for ultrabooks. I don't think AMD is too thratened at $150.

And the last part ... where exactly is the 650M performance? unfortunately you have to look at a ... iffy website to see a direct comparison.

http://www.videocardbenchmark.net/video_lookup.php?gpu=GeForce+GT+650M&id=122


650M, 1278
650, 1779
640, 1304
7660d, 833
Intel hd4k, 472

according to that, in order to achieve 650M peformance, it has to triple the current 4000, doubling puts it right above at the hd7660 ... for an embedded memory part. who knows if the chart is linear or not as far as actual rl performance, obviously not since anandtech showed the 7660D pushing close to the gt 640 at times. Not reliable but some numbers are there, so lets see what Nvidia sais.

According to Nvidia, the gt 650 ~= 670MX so don't get that part mixed up, 650 != 650M by any means.

 

[sarinaide]

GT3 was touted to max skyrim at 1080P, I already used a GtX 550ti and HD 6790 to prove how impossible a claim that was but the Intel faithful said, no we doubled the EU (I am assuming thats Intel code for stream processors). The 550ti and 6790/6850 are entry level mainstream cards that barely manage 30FPS on maxed skyrim settings at 1080P. A person I know who has been a hardware reviewer for over 20 years, professional gamer and overclocker immediately responded to that video as complete nonsense. The assumption is that GT3 will match mid level Llano parts but be around 40%~50% slower by the time Richland is out.

If we factor in GCN parts on Kaviri along with a better arch focusing on latency and cache improvements, we could well see Intel deciding to pull the plug on integrated as its consumes die space and uses power that Intel can theoretically do without.

GT3 and Haswell will be the furthest thing from a APU slayer but from a intel perspective it will be better at least.

 

[esrever]

GT3 will be 7660G level performance in high end ultrabooks for longer battery life.

 

[jdwii]

Read the Techreport one now, makes sense as Kaviri was designated to make the step up. That said it will be interesting to see the step up, after all Vishera was the same as Zambezi yet the performance deltas where on the oposite ends of the room, small changes can affect big results.

Edit.
http://www.tomshardware.com/news/Richland-APU-iGPU-Clock-Speeds-Specs,20748.html

looks like this will have the GCN nice at least that's good news i was just angry that they would use the same design as trinity with a clock bump when Intel is releasing a nice tock this time.
"GPU on the Richland APUs will be part of the HD8000 Series based on the GCN architecture"

Now with this phenom yes its time to do something new. Was a module design smart no to many latency problems with sharing cache, don't even understand why Amd did this thought they had smart engineers unless the marketing team was behind it somehow or dumb management who don't understand anything.