[Xbitlabs]AMD Excavator Core May Dramatic Performance Increases.

[eyeofcore]

As Advanced Micro Devices is preparing to launch its next-generation microprocessors with Steamroller high-performance x86 cores, enthusiasts are revealing secrets about its fourth-generation Bulldozer core code-named Excavator. As it appears, that processing engine will support 256-bit AVX2 floating point instructions, which may mean that it will feature rather revolutionary changes from existing Bulldozer cores.

Source

AMD to Tape Out First 20nm, 14nm FinFET Chips Within Next Two Quarters.

Advanced Micro Devices said Thursday that it would tape out the first products to be manufactured using 14nm FinFET and 20nm planar process technologies in the coming quarters. The company did not elaborate on actual products, but the names of the process technologies do indicate that the chip designer will work with both GlobalFoundries and Taiwan Semiconductor Manufacturing Co.

Source


My thoughts/opinion;
I know that AMD is behind Intel and so are TSMC and GloFO though if AMD and their partners manage to get viable 20nm and even 14nm FinFET yields and production capacity would not that be dangerously close to Intel also we know that Intel has some problems with 14nm because of nature of physics and those chips will be red hot because of density and increased requirements of proper cooling.


I am sure that 14nm FinFET's will be for ARM and maybe MIPS chips from AMD while 20nm will be used for Excavator, so... What kind of performance can we expect from Excavator? On par Haswell or exceed those kind of performances?

 

[ShintaiDK]

On par Haswell or exceed those kind of performances?

Try the 3rd option.

 

[inf64]

I don't know about it getting on par with Haswell but it was known for a long time that either SR or XC(my bet is the XC) will have "full-width" 256bit internal operations meaning it will have native 256bit FP pipelines and won't break the 256bit AVX/FMA ops in 2x128bit ops. The info is here and is dated May 2013

 

[Ventanni]

AMD has a long, long ways to catch up to Intel in the IPC department, and even if they do catch up, they have to do so at a lower power usage if they really want to compete where the money is at (servers). Secondly, with the exception of the Jaguar cores, they have to fab their chips with GloFO, which, I'm sorry, is about the last company AMD wants to be tied to.

Steamroller is supposed to bring about pretty dramatic performance improvements through flow path optimizations of the cores, but there has been rumors of some of the Excavator tech sneaking into Steamroller as well like a memory controller redesign (much needed for AMD APUs!). This is all rumor and speculation though, but I'm excited to what we're about to see.

Realistic Expectations:
- Up to 15-20% IPC improvement over Piledriver (+30% over Bulldozer)
-Slightly lower clocks due to process change (32nm SOI to 28nm bulk)
-GCN 1.0 GPU

It'll be modest, but don't expect miracles. AMD is in a bit of a bind right now because in order to actually make real money off of these big cores, they need to be competitive in the server markets. And, in order to be competitive in the server markets, they need Excavator completely finished like yesterday.

 

[inf64]

AMD has a long, long ways to catch up to Intel in the IPC department, and even if they do catch up, they have to do so at a lower power usage if they really want to compete where the money is at (servers). Secondly, with the exception of the Jaguar cores, they have to fab their chips with GloFO, which, I'm sorry, is about the last company AMD wants to be tied to.

Steamroller is supposed to bring about pretty dramatic performance improvements through flow path optimizations of the cores, but there has been rumors of some of the Excavator tech sneaking into Steamroller as well like a memory controller redesign (much needed for AMD APUs!). This is all rumor and speculation though, but I'm excited to what we're about to see.

Realistic Expectations:
- Up to 15-20% IPC improvement over Piledriver (+30% over Bulldozer)
-Slightly lower clocks due to process change (32nm SOI to 28nm bulk)
-GCN 1.0 GPU

It'll be modest, but don't expect miracles. AMD is in a bit of a bind right now because in order to actually make real money off of these big cores, they need to be competitive in the server markets. And, in order to be competitive in the server markets, they need Excavator completely finished like yesterday.

If Carizzo (APU based on Excavator core) is about to launch in H1 2015 they must have finished it by now. SOG I have linked above is a good indication they have done so and also the recent gcc compiler updates referring to "bdver4" support( having AVX2 in it which SR core doesn't have ).

 

[Fjodor2001]

I think the question is how close Excavator will be to Broadwell, which will be it's direct competitor at that time.

AMD is increasing CPU performance on their mainsteam CPUs at a faster rate than Intel currently. At some point the difference in performance will be close enough so AMD CPUs are considered "good enough", and then AMD can perhaps sell more again. Especially since focus lately also has been on the iGPU, which is where AMD has an advantage compared Intel.

 

[mikk]

I think the question is how close Excavator will be to Broadwell, which will be it's direct competitor at that time.

Next year is Steamroller time (APU only) and the year after that Excavator which means Skylake is more likely as competitor at that time.

 

[eyeofcore]

Try the 3rd option.

There is no 3rd option you rabbid Intel fanboy.

Insulting other members is not allowed here
Markfw900
Anandtech Moderator

AMD has a long, long ways to catch up to Intel in the IPC department, and even if they do catch up, they have to do so at a lower power usage if they really want to compete where the money is at (servers). Secondly, with the exception of the Jaguar cores, they have to fab their chips with GloFO, which, I'm sorry, is about the last company AMD wants to be tied to.

Steamroller is supposed to bring about pretty dramatic performance improvements through flow path optimizations of the cores, but there has been rumors of some of the Excavator tech sneaking into Steamroller as well like a memory controller redesign (much needed for AMD APUs!). This is all rumor and speculation though, but I'm excited to what we're about to see.

Realistic Expectations:
- Up to 15-20% IPC improvement over Piledriver (+30% over Bulldozer)
-Slightly lower clocks due to process change (32nm SOI to 28nm bulk)
-GCN 1.0 GPU

It'll be modest, but don't expect miracles. AMD is in a bit of a bind right now because in order to actually make real money off of these big cores, they need to be competitive in the server markets. And, in order to be competitive in the server markets, they need Excavator completely finished like yesterday.

AMD has won Verizon and they made a big contract and AMD's Opteron's are being used not Xeon's so that is a good start and AMD said that IPC of Kaveri should be 25% and I can see it happening since Steamroller cores are a new iteration of Bulldozer architecture not like Piledriver that was a refresh.

Excavator should have 50% jump in IPC over Bulldozer/Piledriver, Excavator will be 20nm since it will be ready by then though 14nm FinFET's will most likely be limited to ARM CPU's and AMD is planning to use 14nm FinFET's for their ARM SKU's.

Jaguar APU's already have GCN 1.5 and Kaveri should have GCN 2.0 and Excavator APU's should have GCN 2.5/3.0 ...

 

[Arachnotronic]

Skylake will have AVX-512 by the time this chip rolls out...

 

[inf64]

Skylake will have AVX-512 by the time this chip rolls out...

You mean very late 2015 for Skylake? If so then AMD will have no problems as they will already be competing with Broadwell-K parts. Unless you think intel will release Boradwell-K and then after a few months release Skylake too, overlapping these two (which makes zero sense).

Carizzo will compete with shrunk Haswell (Broadwell) and follow-up to Carizzo(based on whatever core in 2016) will compete with Skylake. It might be later than Skylake by a few quarters but that's not a big deal.

 

[Sweepr]

There is no 3rd option you rabbid Intel fanboy.

Even with a 50% IPC jump (dont hold your breath) vs BD/PD a hypothetical ~4GHz 2M/4C Excavator would be slower than a quad-core Haswell perf-wise (especially 8-thread Core i7s) and by 2015 there will be Skylake with AVX 3.2 support, higher IPC than Haswell and quite possibly more than fours cores on the mainstream platform. So yes, there is definitely a third option.

In case some people missed:

According to Intel, the delay of Broadwell will not affect the company's next line of processors, Skylake, as the chips are based on new architecture.

www.macrumors.com/2013/10/16/intel-...l-chips-until-2014-due-to-manufacturing-issue

 

[frozentundra123456]

Kind of a non-story IMO. Who are these "enthusiasts"?

If you read the rest of the article, AMD even had no comment, so it is just more speculation. Besides, it will be software dependent, and if AVX2 becomes popular Haswell will see a huge performance boost in those applications as well.

 

[Abwx]

Kind of a non-story IMO. Who are these "enthusiasts"?

If you read the rest of the article, AMD even had no comment, so it is just more speculation. Besides, it will be software dependent, and if AVX2 becomes popular Haswell will see a huge performance boost in those applications as well.

TReport used a X264 bench that included AVX2 optimisations ,
perfs are indeed better but saying that it s huge is somewhat
completely unfounded to say the least.

 

[frozentundra123456]

TReport used a X264 bench that included AVX2 optimisations ,
perfs are indeed better but saying that it s huge is somewhat
completely unfounded to say the least.

How can anything be "somewhat completely"? In any case point is the statement is just random speculation by an un-named enthusiast. Not even from AMD.

 

[inf64]

Or we can say it's just XBitlaBS. We have exhibited la BS from that website many times now .

 

[Abwx]

How can anything be "somewhat completely"? In any case point is the statement is just random speculation by an un-named enthusiast. Not even from AMD.

Not huge at all if you prefer , actualy just an incremental
improvement , a 4770 is 10% better than a 3770 , is that huge.?.

http://techreport.com/review/24879/intel-core-i7-4770k-and-4950hq-haswell-processors-reviewed/12

 

[Abwx]

Or we can say it's just XBitlaBS. We have exhibited la BS from that website many times now .

Right that Anton Shilov is often out of track , just look
at his article messing AVX2 and FP performances.

 

[frozentundra123456]

Not huge at all if you prefer , actualy just an incremental
improvement , a 4770 is 10% better than a 3770 , is that huge.?.

http://techreport.com/review/24879/intel-core-i7-4770k-and-4950hq-haswell-processors-reviewed/12

So you expect somehow it is an incremental improvement for haswell but will be a huge improvement for excavator? I do think the potential improvements are greater than shown in that article, because I don't think it was fully optimized for AVX2. That is why I am not that impressed with it overall in has well though. Even with the dominant market share of Intel very few applications are taking advantade of it.

 

[guskline]

Any speculation on the type of socket for this chip?

 

[Abwx]

So you expect somehow it is an incremental improvement for haswell but will be a huge improvement for excavator?

Where did i say that it would bring a huge improvement
for EXcavator.???.

Would you mind pointing me where you did read that ,
i guess that you d be annoyed to sustain such curious
rethoric.

I do think the potential improvements are greater than shown in that article, because I don't think it was fully optimized for AVX2. That is why I am not that impressed with it overall in has well though. Even with the dominant market share of Intel very few applications are taking advantade of it.

Now it s the bench fault , yet TR did all they could to show HW
under a good light , going as far as cutting Vishera perfs by as
much as 15% compared to their previous same benchs.

 

[Insert_Nickname]

Even with a 50% IPC jump (dont hold your breath) vs BD/PD a hypothetical ~4GHz 2M/4C Excavator would be slower than a quad-core Haswell perf-wise (especially 8-thread Core i7s) and by 2015 there will be Skylake with AVX 3.2 support, higher IPC than Haswell and quite possibly more than fours cores on the mainstream platform. So yes, there is definitely a third option.

What do you expect from a chip with about half the execution resources of the Haswell quadcore? If you stop thinking of the 2M/4T as a quadcore, and look at them as a dualcore with HT, they start to look a lot better. At 4.5GHz Piledriver trades blows with an i3-3240. That's the fastest Ivy Bridge dualcore on the market...

AMD should stop their insistence in referring to the 2M/4T as quadcores and 4M/8T as octocores. Its really hurting their perception among enthusiasts.

If you read the rest of the article, AMD even had no comment, so it is just more speculation. Besides, it will be software dependent, and if AVX2 becomes popular Haswell will see a huge performance boost in those applications as well.

If AVX2 really takes off, it will result in a large number of toasty Haswells from what I'm hearing...

 

[AtenRa]

AVX will be completely irrelevant by that time. Where are the AVX 2.0 applications ??? People where tooting about Haswell and AVX 2.0 months before Haswell release.

Lets see SR with Kaveri first and then we talk about XC. But i dont expect them before Q3 2015.

Kaveri will compete against Haswell, Haswell refresh and a little bit against Broadwell at the end of 2014 start of 2015.
XC will compete against Broadwell and Skylake after H2 2015

 

[Phynaz]

enthusiasts are revealing secrets about its fourth-generation Bulldozer core code-named Excavator

No better source to be found!

 

[sniffin]

I'm pretty sick and tired of talk about how AMD is going to eventually bend Intel over in x years time when blahblah is released. After nearly 8 years of playing second fiddle I think its about time people gave it a rest. Not to mention predicting Excavator performance at this stage is sort of ridiculous.

What do you expect from a chip with about half the execution resources of the Haswell quadcore? If you stop thinking of the 2M/4T as a quadcore, and look at them as a dualcore with HT, they start to look a lot better.

So you're suggesting we distort reality because the performance of their 8 cores isn't what an 8 core should be? Performance has NOTHING to do with core count. Similarly, modules are NOTHING like hyperthreading. They don't even perform anything like it.

 

[ShintaiDK]

Any speculation on the type of socket for this chip?

FM2+ or FM3