AMD "Steamroller/Excavator" -info, speculations and experience

[EniGmA1987]

Is the FM2 socket going to provide a processor that will be able to feed a mid-high end graphics card?

A Richland processor can already feed a GTX 680 perfectly fine if the game is coded to use the GPU well. There is no bottleneck of any kind like people seem to throw around these days. It is when a game is needing more CPU power that you start to see a difference, which many games do need a good chunk of CPU. We do not know how fast Kaveri will be just yet. It may be 20% faster than Richland all around, it may be 2% faster with a much better GPU. We also dont know how well future games will be coded. So it is really a complete gamble at this point in time whether Kaveri will be a really great little CPU to have, or a complete piece of crap.

[informal]

the only time FX series will get out done is single thread. Kaveri will still look a dual core in things like cinebench. I'm not talking about it reading 2 cores 4 threads either. I'm talking about the MP ratio on advance setting.

Nope. I have 3.2x mp ratio on my 750K in C15. That is exactly 20% lower than what "true" QC with no SMT get in this metric(like 3570K etc.). Kaveri will have dedicated decoders so 20% penalty there will be gone. This will make it basically scale as a true QC chip since the mp ratio will be ~3.8x+ or in the range of "true" QC designs. Then on top of that you will have *some* IPC gain over richland, it can be 1% or it can be ~15%, depending on the test. No wonder why AMD lists ~30% more ops/cycle, this is exactly what you will get when you multiply both the effect of dedicated decoders and the IPC gain due to various other improvements: 1.3x~=1.2x1.1 (20% being due to no penalty as we have 8 intsr/cycle per module instead of just 4 in multithreaded workloads; up to ~10% IPC which might be on the lowish side here but I'm being careful here).

Kaveri will look to outside software like a genuine QC product (looking at MP scaling factors, execution threads etc.). The possible fact that it won't be as fast as say 3570K( both at stock of course) doesn't mean it won't be a "true" QC. It will mean it is just a slower true QC product. The reasons behind this are various but let's say intel will still have more integer and SIMD/fp capability "per thread" in this comparison so it's no wonder this might still be the case. Excavator core will expand on SR core so in 2015 we might finally have a core that will be roughly on an equal footing with Haswell/Broadwell (core per core, frequency being design specific). Intell will still have an advantage in thread count as AMD won't be going over 3 modules/6 threads, but most software on desktop cannot still even utilize 4 cores properly so it won't matter.

[demonkevy666]

Nope. I have 3.2x mp ratio on my 750K in C15. That is exactly 20% lower than what "true" QC with no SMT get in this metric(like 3570K etc.). Kaveri will have dedicated decoders so 20% penalty there will be gone. This will make it basically scale as a true QC chip since the mp ratio will be ~3.8x+ or in the range of "true" QC designs. Then on top of that you will have *some* IPC gain over richland, it can be 1% or it can be ~15%, depending on the test. No wonder why AMD lists ~30% more ops/cycle, this is exactly what you will get when you multiply both the effect of dedicated decoders and the IPC gain due to various other improvements: 1.3x~=1.2x1.1 (20% being due to no penalty as we have 8 intsr/cycle per module instead of just 4 in multithreaded workloads; up to ~10% IPC which might be on the lowish side here but I'm being careful here).

Kaveri will look to outside software like a genuine QC product (looking at MP scaling factors, execution threads etc.). The possible fact that it won't be as fast as say 3570K( both at stock of course) doesn't mean it won't be a "true" QC. It will mean it is just a slower true QC product. The reasons behind this are various but let's say intel will still have more integer and SIMD/fp capability "per thread" in this comparison so it's no wonder this might still be the case. Excavator core will expand on SR core so in 2015 we might finally have a core that will be roughly on an equal footing with Haswell/Broadwell (core per core, frequency being design specific). Intell will still have an advantage in thread count as AMD won't be going over 3 modules/6 threads, but most software on desktop cannot still even utilize 4 cores properly so it won't matter.

adding in a decoder doesn't change the fact the L1 instruction cache is still shared and will probably still work the same way for Kaveri as it does for FX.

[tbone8ty]

Any idea when these will be available?

[xVeinx]

Feb 2014?

[FlanK3r]

I think January (its my tip)

[informal]

adding in a decoder doesn't change the fact the L1 instruction cache is still shared and will probably still work the same way for Kaveri as it does for FX.

It has received a major revamp(size, hit rate) along with many other parts of the front end (that are still also shared). This is all public info that we have known for almost a year.

[Darakian]

Not steamroller news but AMD just posted patches for excavator support to the linux kernel.
http://www.xbitlabs.com/news/cpu/dis...Increases.html
https://sourceware.org/ml/binutils/2.../msg00155.html
Xbit also has an article on AMD moving to 20 and 14nm
http://www.xbitlabs.com/news/other/d..._Quarters.html

tl;dr the patch basically says excavator is in the works (steamroller got added over a year ago so... ya)
the lithography article doesn't have any hard timelines either so it's more fluff than anything.
Still, people talking.

[tbone8ty]

heres to a 28nm processor for am3+!

[Yeroon]

Hey i know this is slightly off topic, but anyone have any new news on the binned PS4 chips for retail? Kinda wondering when/how those are going to show up and what configuration.

[tbone8ty]

probably not gonna happen

but would possible make a decent steambox??

but yah probably not, cpu is not that strong

[The Stilt]

Mullins should be available for embedded solutions (other than console) at some point.

[FlanK3r]

Hm, Kaveri has the same VID interface standard as Richland? Richland+Trinity has SVID 2.

[tbone8ty]

http://wccftech.com/amd-kaveri-steam...ition-haswell/

[xVeinx]

If Kaveri is only going to have 85% of the float performance of the previous cores, I'd like to think they wouldn't release it (i.e., back to tweaking).

[FlanK3r]

this is not so simply - FXs have L3 cache and there is performance boost (you can see it in superpi or in harder benchmarks as Cinebench too)

[tbone8ty]

Can't wait for kaveri mobile chips!

Wanna get something like the MSI gx70 but Richland was a little bit lacking to power the 8970m

[xVeinx]

this is not so simply - FXs have L3 cache and there is performance boost (you can see it in superpi or in harder benchmarks as Cinebench too)

True, but this would be after the improvements. There wasn't more than about 15% performance difference to begin with, and this doesn't make sense with the improvements we're seeing with integer...

[FlanK3r]

I heard about lower clocks with Kaveri chips. No more 4 GHz...True/false? I dont know now.

[radier]

Source

[The Stilt]

I heard about lower clocks with Kaveri chips. No more 4 GHz...True/false? I dont know now.

If you have ever been sitting in a traffic jam caused by a road construction site, you'll know that these...



... move very slowly, but also get the job done very effectively.

Hammer time.

[tbone8ty]

Source

Yah I remember there was something mentioned in The article explaining steamrollers arch that it was gonna be lower clock speed. But more efficient per clock

[informal]

Stilt I LOVE the reference but I still hope it's not sitting at stock ~3Ghz range.

PS "Hammer please don't hurt them" :P

[FlanK3r]

If with lower clock will be effective why not lower clocks after ...

[xVeinx]

dropping the clock from 4 to 3 GHz is a steep drop. Seems there is a lot changing under the hood this time around....