The AMD Execution Thread [2007 - 2017]
- Thread starter overclocked_enthusiasm
- Start date
-
- Tags
- amd
- Status
itsmydamnation
Veteran
Are you serious?
Bulldozer came out oct 2011, richland is now shipping to OEM's! piledriver offers a solid 10% IPC over bulldozer + 5% clock(same TDP). Now richland is bringing somewhere around 10% perf increase ( as per http://www.amd.com/us/press-releases/Pages/amd_unveils_new_apus.aspx). So in 15 months Performance at a given TDP is up around 26%.
Now lets look at what AMD has and will be delivering (assuming on time for future products)
Ontario jan 2011
llano june 2011
bulldozer oct 2011
trinity may 2012
vishera oct 2012
hondo oct 2012
richland Dec 2012
Kaveri H2 2013
Kabini Q2-3 2013
Temash Q2-3 2013
PS4 ???? 2013-2014?
Xbox ???? 2013-2014
bulldozer sucked but has there delivery of Chips really been that bad?
toms hardware have an interesting comparison of cpu's per core per clock
http://www.tomshardware.co.uk/charts/x86-core-performance-comparison/benchmarks,128.html
http://www.tomshardware.co.uk/charts/x86-core-performance-comparison/benchmarks,128.html
Richland isn't here. Can I buy it ? Even still 26% increase doesn't bring it to Ivy bridge levels let alone haswell which will most likely be out at the same time . Nor does it bring it within the same TDP.
Its a useless product . The only smart upgrade path for me as a bulldozer user is to go intel when haswell hits.
Will Richland even ship without the graphics part ?
Then we move to mobile. The trinity chips are okay there but again cpu power is extremely weak compared to intel and even the gaming will get beaten by a cheap dedicated gpu for a laptop. I'd take an i3 or i5 with a NVidia or amd gpu for a small premium over the amd chips .
Then tablets and convertables. Where are they ? The only thing I've seen from a company is vizio with the c-50 but we already got that 2 years ago in the iconia w500 . I'm sure graphics will be better than clovertrail but battery life will be barely half of the atom chips. Comparing the c-50 to an ivy bridge gpu battery life will be similar but the c-50 will be outclassed in every aspect.
AMD is simply late and will continue to be late. All amd's chips are going to go against haswel or the new atoms and be out classed once again and they still will have been 8 to 12 months late to the windows 8 launch party .
Its been 2 years since the c-50/e-350 hit the market. Where is the 28nm verisons ? Proper verisons of those would have wiped the floor with the current atoms and amd could have made money and market share , but now they will be late and what could have been attractive products will be also rans . But this is the story of amd on the cpu side and to some extent the gpu side for the last 3 or so years
You're pulling these comparisons out of thin air. You shouldn't crucify a product before it's tested.
The c-50 has been out for 2 years . I don't need to pull anything out of thing air. I've actually owned the w500 which has a c-50 in it
IT gets less than 4 hours of battery life watching videos and about 2 gaming. The tablet has a 36wh battery.
The acer iconia w510 has the atom z2760 dual core 1.8ghz atom and a 27wh battery. Anand gets almost 9 hours of battery life watching videos. More than twice the battery life .
Looking at anand's benchmarks of the w510 its faster in the majority of of tests than the e-350 which is a dual core bobcat at 1.6ghz cpu vs the 1ghz c-50.
The only place the c-50 will have an advantage is in graphics.
Now when you look at the ivy bridge the cpu portion will step all over the dual core 1ghz bobcat in the c-50 . The hd4000 will perform similar to better than the 6250 80 shader apu .
Mabye I could get excited if they had something from 2011 or even 2012 but the c-50 was announced in nov of 2011 . That is how poorly amd has executed in the tablet space.
itsmydamnation
Veteran
I just typed a huge post i lost, so i will cut it short, go look at the hotchip jaguar presentation on power consumption, then look at the published performance specs of the 5watt and 17watt jaguar chips. Its small, cheap, completely integrated soc and has very high performance. Given that its neck and neck with 35watt Intel ivy bridge chips in PC mark Vantage and has 20% over 35watt trinity AMD isn't going to have anywhere near the hard time you have assumed.
Intel's the one who's going to have to try and protect its high margin ULV chips.
Last edited by a moderator:
Eh, how about we wait on that one for a bit? Whilst a commonly recurring theme (this new AMD thing is REALLY THE ONE THAT'LL SET THE WORLD ALIGHT! NO, REALLY, THIS IS THE ONE!!!!), it has yet to practically pan out.
On the other hand, I think that if AMD executes on what they have shown at CES (not necessarily a huge if, bar Kaveri coming out in 2013), their position might not be as grim as I (and perhaps others) have assumed, and they should be capable of living in the neighbourhood of black, as opposed to bleeding awfully. I'm very much looking forward to Kaveri, although my good experience with a Trinity powered laptop that also has a GCN based GPU might've added a bit of a rose tint to my glasses.
The way I see it, AMD's position relative to Intel is very likely to improve, if only because of the law of diminishing returns. Intel's architecture is so good at this point that they have to fight really hard for every performance percentage point they can get, while it's relatively easier for AMD. Still very hard, I'm sure, just not quite as much.
Even if Richland only brings about 5% more CPU performance, and Kaveri about 15% (IPC and clocks combined) that should still be a bit more than what Haswell is expected to bring over Ivy Bridge.
Kabini and Temash look quite good, so overall I'm cautiously optimistic about AMD's future. However the first few months of 2013 are probably going to be tough, but maybe GCN2 will help.
I'm not certain that Kaveri / Steamroller will be able to bump clocks up - or even keep them at the level they currently are with Trinity / Piledriver. On one hand, from what they've shown they're widening SR, which is likely to have a net negative impact. On the other, they're moving from GF 32nm (which was quite likely actually the last GF process to be tuned for AMD needs, since it was almost entirely inherited) with notable custom work being put in, to a generic SHP 28nm thing and standard libraries wherever possible. It remains to be seen how these two effects interact with other architectural choices.
With regards to the latter argument (AMD is improving more than Intel, core-to-core), I'm not sure that is accurate either, and even if it is, considering the basis against which those percentage point increases are being computed (AMD's current cores are most weak), once we move to absolute values the picture may not be so rosy. However, at this point in time it would be somewhat naïve to assume that AMD can compete pound-for-pound with Intel, but I think they can still carve a nice spot for a leaner them. And no, ARM nonsense is not the path to that, but decent x86s and strong CPU+GPU offerings might be.
I'm not certain either, I was mostly mentioning clocks for Richland, since it's apparently no different from Trinity micro-architecturally.
That said, and as skeptical as I may be about AMD's ability to bump clocks for a 100W desktop Kaveri vs. the equivalent Richland, 17~25W notebook APUs may be a different story, since these are much more power-limited. After all, if AMD is moving to 28nm, we have to assume that overall, it's worth it.
A while ago I looked at Hardware.fr's numbers and noticed that Piledriver had brought roughly a 15% to 18% performance improvement over Bulldozer (respectively in applications and games) while Ivy Bridge had brought about 7% to 9% over Sandy Bridge. Of course, Ivy Bridge is but a tick and Haswell ought to do more, but I do think we can expect to see this sort of trend for a couple of generations, precisely because AMD's basis is so low.
And as you said they don't necessarily need to catch up (they probably can't, anyway) they just need to not be so far behind.
I think GloFo 32nm Gate First process is fairly decent for low clock SKU's but fails miserably at delivering good power characteristics above 3.5GHz for Bulldozer/Piledriver. They have huge variability among wafer area and even within one CPU die. There are many reports from overclockers finding out individual modules can be easily 200MHz weaker/stronger than their neighbours at a given voltage and in some extreme cases much wider discrepancies can occur.
I'm expecting AMD to gain quite a bit of power efficiency and density when moving to 28nm HKMG Gate Last manufacturing. I do not expect Steamroller to ship at 4GHz, but maintaining current clocks of A10-5800K and possibly enabling better Turbo should be viable.
Put this another way, if Intel 32nm process was good enough for Sandy's short-pipeline design to hit 3.5-3.9GHz I do expect Glo-Fo 28nm one to be at least within 10% of Intel's standards. Therefore putting longer pipelined design on that process should land us around 3.6-4.2GHz mark fairly easily.
PS. I came across this quick Kabini die shot taken from www.chip-architect.com
At only around 114mm2 it should have much larger GPU part compared to Brazos as it was established by our forum members that CPU part should be fairly compact and not much bigger than 2 Bobcats @40nm. So we can be looking at around 50-70mm2 for iGPU part!
I wonder how this small beast will cope having only one 64bit memory channel ...
I'm expecting AMD to gain quite a bit of power efficiency and density when moving to 28nm HKMG Gate Last manufacturing. I do not expect Steamroller to ship at 4GHz, but maintaining current clocks of A10-5800K and possibly enabling better Turbo should be viable.
Put this another way, if Intel 32nm process was good enough for Sandy's short-pipeline design to hit 3.5-3.9GHz I do expect Glo-Fo 28nm one to be at least within 10% of Intel's standards. Therefore putting longer pipelined design on that process should land us around 3.6-4.2GHz mark fairly easily.
PS. I came across this quick Kabini die shot taken from www.chip-architect.com
At only around 114mm2 it should have much larger GPU part compared to Brazos as it was established by our forum members that CPU part should be fairly compact and not much bigger than 2 Bobcats @40nm. So we can be looking at around 50-70mm2 for iGPU part!
I wonder how this small beast will cope having only one 64bit memory channel ...
Are you claiming that AMD is moving its CPU manufacturing to TSMC? GF is gate-first until 20nm.
It's shorter than BD, but the full pipeline length isn't entirely certain. BD has a penalty ~20 cycles for conditional branch mispredicts, while Sandy Bridge has ~15 for misses within the uop cache.
It's several cycles more if it's not in the uop cache.
So SB is somewhat close to BD if it misses the uop cache, and about ~25% shorter otherwise.
upnorthsox
Veteran
So 50-70mm^2 then for steamroller? Is that 2 or 4 cores?
itsmydamnation
Veteran
SR isn't any wider, its still 2 ALU + 2AGLU , the FP actually gets narrower 2FMAC 128bit + 1 MMX not 2 and 2. The only thing that is wider is the decode which isn't any wider there just two per module not one. They will also have a loop buffer which saves power and much imporved L/S ( papermaster stated some L/S operations are 2 cycles quicker in SR then BD/PD). Fixing the I1$ cache issues should have both a performance and power improvement. So at this stage i would say there isn't anything thats an obivious increase in power draw or an obivious decrease in clocking potencial .
True but GF SOI 32nm has been a pretty bumpy road so who really knows how they compare. That said SR/EV are heading down the road of automated design so being on a process that has more then just yourself on it should have a tools benifit.
SR should be a big improvement, 30% high instruction throughput a clock, if that doesn't turn into performance improvement then maybe AMD need to get north korean rocket scientist on the engineers ass
.the thing about kabini is not that intel cant compete with it, it is if they will want to. i5 and i7 ULV chips sell for 250-350 USD, thats a long way from what kabini will be sold as, celerons wont cut it and how silvermount performs is yet ot be seen.
According to Mitch Aslup, he once designed a simple scalar in-order x86 processor that could go all the way up to 3GHz with only a 6 stage pipeline.
What that should tell you is that maximum clock speed is not merely a function of pipeline length. It depends a lot on the stuff you're doing during those pipeline stages. You can't really compare two completely different uarchs this way.
But I doubt GF's 28nm bulk is so inhibitive vs its 32nm SOI that it's lowering the practical clock speed limits by so much.
Oh s****
forgot that 'tiny' detail ...Well, that was my impression from reading many sources during last year, but now after checking Global Foundries web page it clearly states they are sticking to Gate First approach.
Also AMD confirmed during last year conference call regarding Wafer Supply Agreement that Kaveri will be fabbed at GloFo 28nm bulk, so no, I'm not claiming they are moving to TSMC. I simply made mistake/misremembered this one.
- AMDs CFO Thomas Seifert
Yes, BD/PD and SB are close to each other with BD/PD featuring somewhat longer pipeline, but where SB at 3.5GHz still has plenty of clock and thermal headroom, BD/PD is running out of it very quickly mainly down to manufacturing process problems (my opinion). If GloFo 28nm bulk fixes some of the issues they experiencing at 32nm SOI this should compensate somewhat for added circuit complexity of Steamroller (who knows AMD might even add few pipelines there).
It's hard to tell from their press release blurb how much better 28nm is compared to 32nm as they are comparing it to 40nm bulk:
If we assume 45nm SOI was about on par with 40nm bulk by looking at how bad power consumption was when moving K10.5 to Llano (Athlon II parts with fused iGPU) we can hope this new 28nm process will be much more efficient. On the other hand if we assume all the blame for high speed Llano core power woes was down to rushed K10.5 porting we might not have great jump. I hope the reality will be somewhere in-between and GloFo 28nm can offer 20%-30% lower switching power at the same clock compared to 32nm therefore 'unlocking' higher clock celling thanks to removing some of that power/thermal limit for high performance CPU's
Upcoming laptop parts based on Kaveri are completely different case as 32nm didn't suck this bad with low clock products compared to 45nm or even other manufacturers processes.
This is Jaguar low power core not a Steamroller, and Gipsel estimated 4 core CU with L2 to be around 25mm2. Even assuming he is off by 50% we are nowhere near 50mm for 4 Jaguar cores.
Last edited by a moderator:
- Status
