[Xbitlabs]AMD Excavator Core May Dramatic Performance Increases.

[Abwx]

Power efficiency is going to be quite a bit in TSMC and GlobalFoundries ball court, yes? Of course AMD can greatly help by de-Pentium4ing their design and targeting more reasonable clocks. Remember that die shot that looked quite nice in terms of improving upon Bulldozer/Piledriver? Even more likely that was a peek at Excavator, it seemed to have dual or quad 256 bit FPUs.

Thoses implementations seems to be trivial to implement in BD
wich seems to have a quite versatile pipeline but infortunately
for such a design to materialize they need a smaller node than
28nm , i guess they are badly slowed by GF unability or eventualy
unwillingness to speed up their nodes transitions.

 

[inf64]

It has only 2 FP units that are dual threaded, just like an i3 with SMT . And in 4 threaded workloads it performs on par or better than i3 while costing less and having much faster iGPU(in case of A8/A10) . Plus in ST workloads due to close to 50% clock difference ( Athlon Ks/A8/A10s are unlocked) the ST performance deficit is basically close to nothing.

 

[Abwx]

It has only 2 FP units that are dual threaded, just like an i3 with SMT . And in 4 threaded workloads it performs on par or better than i3 while costing less and having much faster iGPU(in case of A8/A10) . Plus in ST workloads due to close to 50% clock difference ( Athlon Ks/A8/A10s are unlocked) the ST performance deficit is basically close to nothing.

It takes two cores to efficiently use this single FPU
otherwise if a single core could make full use of it
the ST CB score would be half of the 4 threads score.

Off course this would change nothing on a MT score
but in a given task if ever a core needs only 20%
of the FPU ressources over a given number of cycles
then the second core will be unable to use the full
remaining 80% efficiently , wich may be penalizing
in some given pieces of code.

Also , it makes no doubt that this FPU is extremely
powerfull , so much that AMD did share it for two cores.

 

[PPB]

I am more worried that there are no SR leaks on information, i thought it was a Q4 release (ie benches) and Q1 rollout.

I think SR will have a better IPC incerase than XC as it will be incorporating things than SB has such as the uops cache that i think does wonders. Hopefully they have the L3 cache latency issue finally fixed in this too.

I sincerely hope they dump the L3 cache altogether this round. The perf loss of going L3 cache-less on the APUs shows that all that die space could be put for a better use (or just make a smaller chip and improve die/wafer ratio).

 

[nenforcer]

I really hope AMD releases this on AM3+. But I guess it's more likely they will release on a new FMx socket

It looks like Steamroller is only going to be for APU chips on the Socket FM2+ platform. There will be no FX series AM3+ Steamroller chips for at least the time being. Maybe in one year (Q4 2014) something will show up for AM3+.

 

[Vesku]

I am more worried that there are no SR leaks on information, i thought it was a Q4 release (ie benches) and Q1 rollout.

I think SR will have a better IPC incerase than XC as it will be incorporating things than SB has such as the uops cache that i think does wonders. Hopefully they have the L3 cache latency issue finally fixed in this too.

AMD will probably wait until Steamroller products are near retail, it would be so easy for Intel to modestly tweak their lineup to pummel any gains AMD has made outside of iGP performance.

 

[wlee15]

IIRC Steamroller is to be launched in Q114, and we are yet to hear anything of that processor. And instead of bringing attention to the soon-to-be launched product, AMD is trying to bring attention to... Excavator, a product which is at least 6 quarters away.

Gentlemen, get ready for another deception once Steamroller hit the review sites.

Exactly how is AMD trying to bring attention to Excavator with a patch to a open-source programing tool?

The Pentium Pro wasn't the successor to the Pentium. The P2 was. The Pentium Pro was kind of like the Athlon FX compared to the Athlon 64.

The Athlon 600 versus the K6:
http://www.xbitlabs.com/articles/cpu/display/amd-athlon.html

Doesn't look like 50% IPC improvement to me.



Prescott/Cedar Mill to Conroe was an 8% improvement, really? How about Conroe to SB? I'm pretty sure that the reason why IVB and Haswell weren't revolutionary in performance terms was because they didn't have to be, and Intel knows that desktop CPU development isn't where their R&D should be poured into.

You're wrong about the Pentium Pro which introduced the P6 microarchitecture.

 

[mrmt]

I'm a little skeptical about this. To the best of my knowledge it was one person, JF-AMD, claiming it (and not in an especially official capacity like via presentation material). I know he used the excuse that the engineers promised him this and failed to deliver but I wonder if he didn't misunderstand something. His blog and forum posts were rife with basic misunderstandings of technical data, including how his own company's processors worked.

Indeed. AMD own marketing material mentioned a "not significant" loss on serial single-threaded workloads. The higher IPC was JFAMD creation. OTOH AMD seems to not have corrected his executive or at least ask him for shut his trap, so they should be guilty as charged too.

 

[NostaSeronx]

AMD own marketing material mentioned a "not significant" loss on serial single-threaded workloads.

Are you sure the slides actually said that or are you being delusional?

Arbitrary Pts != Time to Finish

 

[mrmt]

Are you sure the slides actually said that or are you being delusional?

Arbitrary Pts != Time to Finish

http://www.anandtech.com/Gallery/Album/754#6

 

[sniffin]

Power efficiency is going to be quite a bit in TSMC and GlobalFoundries ball court, yes? Of course AMD can greatly help by de-Pentium4ing their design and targeting more reasonable clocks. Remember that die shot that looked quite nice in terms of improving upon Bulldozer/Piledriver? Even more likely that was a peek at Excavator, it seemed to have dual or quad 256 bit FPUs.

This news is interesting in that it means AMD probably isn't completely retreating into the cat family and 64 bit ARM.

Edit- I believe I was thinking of this thread, http://forums.anandtech.com/showthread.php?t=2322010&highlight=steamroller , when mentioning the potential excavator die shot.

Power consumption for 32nm Piledriver isn't really that out of line anyway. Mobile Sandy i7s were operating in the region of 2-3GHz with 45W TDPs slapped onto them. Their 3GHz+ Sandy i7s had TDPs of 55+ iirc. On a similar process AMD has managed to squeeze out 3GHz+ on mobile Richland with a TDP of 35W. Granted there is more to a Sandy core than there is a Piledriver core. They made huge strides from Bulldozer -> Piledriver -> Richland Piledriver in the area of power efficiency.

Steamroller and Excavator will have more meat to them but that will be offset by process advancement. I don't think AMD will ever have a problem matching or beating Intel clocks at similar TDPs (speaking only to big cores), it's just a question of how much AMD can get done each clock and at the moment it isn't very much by comparison.

Are you sure the slides actually said that or are you being delusional?

Arbitrary Pts != Time to Finish

People tend to remember JF-AMD saying they would hold the line on IPC, but there were slides (as above)that indicated IPC would actually be lower. Nobody paid much attention though

 

[Vesku]

Power consumption for 32nm Piledriver isn't really that out of line anyway. Mobile Sandy i7s were operating in the region of 2-3GHz with 45W TDPs slapped onto them. Their 3GHz+ Sandy i7s had TDPs of 55+ iirc. On a similar process AMD has managed to squeeze out 3GHz+ on mobile Richland with a TDP of 35W. Granted there is more to a Sandy core than there is a Piledriver core. They made huge strides from Bulldozer -> Piledriver -> Richland Piledriver in the area of power efficiency.

Steamroller and Excavator will have more meat to them but that will be offset by process advancement. I don't think AMD will ever have a problem matching or beating Intel clocks at similar TDPs (speaking only to big cores), it's just a question of how much AMD can get done each clock and at the moment it isn't very much by comparison.



People tend to remember JF-AMD saying they would hold the line on IPC, but there were slides (as above)that indicated IPC would actually be lower. Nobody paid much attention though

I did, I was doing rough calculations based on retaining Phenom II IPC. JF-AMD's statements mainly encouraged me to think that they would at least hold the line. Shame they didn't even meet that level of performance. The FX 8320 is ONLY a mildly tempting drop in replacement for my 1090T (which I had ordered as soon as actual BD numbers hit the web).

 

[Shivansps]

They dont need to come up with fast performing CPU to beat Intel, they need power efficient and competitive CPUs, specially at the lower end and mainstream, I7 its not mainstream, neither are I5K.

AMD is getting beaten badly where it was the strongest, they are only holding the line with mainstream notebook cpu and FX6300/6800K at this point. They lost the low end market almost completely.

 

[nemesismk2]

They might stop when the cooling system uses/costs more than the rest of the machine

that would do it

 

[sniffin]

They dont need to come up with fast performing CPU to beat Intel, they need power efficient and competitive CPUs, specially at the lower end and mainstream, I7 its not mainstream, neither are I5K.

AMD is getting beaten badly where it was the strongest, they are only holding the line with mainstream notebook cpu and FX6300/6800K at this point. They lost the low end market almost completely.

They have the lowend as well imo. Where I live the cheapest Haswell i3 retails for $145. FX-6300s on the other hand retail for $135. An overclocked FX-6300 beats that hands down. Or I could spend a little more and grab an 8320.

Intel have no game in the low end where I am. They could destroy AMD in that segment if they really wanted but they are pricing themselves out of it. I don't know what US pricing is like though.

edit: The cheapest Ivy is $142. i3s aren't an option

 

[MisterMac]

They have the lowend as well imo. Where I live the cheapest Haswell i3 retails for $145. FX-6300s on the other hand retail for $135. An overclocked FX-6300 beats that hands down. Or I could spend a little more and grab an 8320.

Intel have no game in the low end where I am. They could destroy AMD in that segment if they really wanted but they are pricing themselves out of it. I don't know what US pricing is like though.

edit: The cheapest Ivy is $142. i3s aren't an option

That's because your segment of DIY Lowbudget perf\price enthusiasts is even lesser than the absolute highend crazy modding machine segment.

Most in your segment buy pre-built - and the unit price for an i3 will be much less when bought in 10k at a time.


There's the HTPC crowds, but they wouldn't buy a FX6300 in the first place, or a mid\top i3 for that matter. They'd go different routes

 

[frozentundra123456]

They have the lowend as well imo. Where I live the cheapest Haswell i3 retails for $145. FX-6300s on the other hand retail for $135. An overclocked FX-6300 beats that hands down. Or I could spend a little more and grab an 8320.

Intel have no game in the low end where I am. They could destroy AMD in that segment if they really wanted but they are pricing themselves out of it. I don't know what US pricing is like though.

edit: The cheapest Ivy is $142. i3s aren't an option

The i3 is not really low end. They have both pentium and celeron for that. Both cheap, good enough performance for everyday use, and very efficient. The fx6300 is attractive for enthusiasts who will use a discrete card, but not so much for mainstream, especially prebuilts, because it has no igp.

 

[monstercameron]

The i3 is not really low end. They have both pentium and celeron for that. Both cheap, good enough performance for everyday use, and very efficient. The fx6300 is attractive for enthusiasts who will use a discrete card, but not so much for mainstream, especially prebuilts, because it has no igp.

^^agreed and the celerons are every where and are pretty cheap!

 

[Sheep221]

I sincerely hope they dump the L3 cache altogether this round. The perf loss of going L3 cache-less on the APUs shows that all that die space could be put for a better use (or just make a smaller chip and improve die/wafer ratio).

how it will work without L3? you need that in multi core CPUs

 

[AtenRa]

Indeed. AMD own marketing material mentioned a "not significant" loss on serial single-threaded workloads. The higher IPC was JFAMD creation. OTOH AMD seems to not have corrected his executive or at least ask him for shut his trap, so they should be guilty as charged too.

That was for the CMT(Cluster Multi-Threding) design vs CMP(Chip Multi Processor), it was not meant for comparison between Phenom vs Bulldozer.

 

[Shivansps]

The i3 is not really low end. They have both pentium and celeron for that. Both cheap, good enough performance for everyday use, and very efficient. The fx6300 is attractive for enthusiasts who will use a discrete card, but not so much for mainstream, especially prebuilts, because it has no igp.

Exactly my point, back in 2009 i could never have imagined a escenario where it was a good idea to buy a Celeron instead of a AMD, since Sandy Bridge Celerons and Pentiums perform better than A4 line, and since Ivy they are not far on IGP either, and H61 mbs are very cheap too.

Attractive options from AMD on desktop are the FX6300 and 6700/6800K. On Notebook Richland is good, but Kabini and Temash are having a hard time, less back to 2009 and check if it was the same situation with E-350, who had no match.

 

[PPB]

how it will work without L3? you need that in multi core CPUs

Who told you that? So tell me how all those 2 module apus are working right now? Or the K10.5 cores in Llano?

 

[glugglug]

Has a company ever produced an x86 processor that has a 50% jump in IPC over its predecessor?

I can think of a few possibilities, although in recent times there has been stagnation.

80386 vs. 80486 (it might even be a 500% jump).
80486 vs Pentium
K5 vs K6 (again, much more than 50% jump).
Pentium vs. Pentium Pro?
Pentium 4 vs. Core (or maybe Core vs Core2)?
K7 vs K8?

 

[krumme]

But excavator certainly have the potential to do the same against bd. Lol

 

[Ajay]

But excavator certainly have the potential to do the same against bd. Lol

When is EX coming out - maybe 2015? If AMD is super lucky, GFL will have 20nm by the end of that year (but most likely not till sometime in 2016). Unless things change, by some combination of cash contributions to node R&D and the potential to produce console APUs, the standard 20nm process @ GFL will be LP - not so good for AMD; hopefully, AMD and GFL have worked out and arrangement for a high power node for AMD APUs. If there is no SHP/HP node (depending on your definition) for EX, then it won't just be beaten by Skylake, it will be an utter joke.

This is just going from sad to pathetic in real terms. What makes things even worse is the state of AMD chipsets. I've been looking into to this recently and I can't believe how far they are behind Intel in this arena as well. The satisfaction rates over at Newegg and Amazon on AMD motherboard just isn't very good compared to some Intel based offerings.

Sorry to be so pessimistic, but I didn't realize how poor the over-all ecosystem was for AMD until I was looking for a motherboard for a pair of GTX 460 to run F@H 24x7. I thought AMD might offer me an inexpensive way out, but even if I went up in price, the overall ratings just were not what I expected.