First Steamroller processor core exposure

[chernobog]

I wouldn't call that kind of prediction "safe estimation". Steamroller SKUs would need to have close to a 40% combined improvement (IPC + Clockspeed + Reduction in CMT penalty) over current Piledriver SKUs to meet those expectations. Perhaps 25-30% if you mean solely in multi-threaded scenarios.

AMD mentioned that the CMT penalty will be reduced by 20% as I said before, 20% better multi-threaded performance while we know that IPC will be 15 to 20% higher than Piledriver atleast in Kaveri SKU.

They did not mentioned boosting clock speed.

Go to the first page of the thread "Steamroller could crush Haswell" and click on xbitlabs link.

 

[Vesku]

Reread the Xbitlabs write up, they list individual front end improvements as provided by 2012 Hot Chips AMD presentation. They never totaled it up to estimate % MT improvement, at least not over Piledriver. Their 30% more ops per cycle estimate was directed at original Bulldozer.

 

[ShintaiDK]

AMD mentioned that the CMT penalty will be reduced by 20% as I said before, 20% better multi-threaded performance while we know that IPC will be 15 to 20% higher than Piledriver atleast in Kaveri SKU.

They did not mentioned boosting clock speed.

Go to the first page of the thread "Steamroller could crush Haswell" and click on xbitlabs link.

The CMT penalty is around 20%. 20% reduction is 5% faster compared to the orignal 80% speed offset.

And going from SOI to Bulk is not exactly talking for frequency increase, rather the opposite.

 

[guskline]

The only SteamRoller that is out and is cool is by the King, AMD is too late!
http://www.youtube.com/watch?v=9VXr6U-BwMs

 

[chernobog]

The CMT penalty is around 20%. 20% reduction is 5% faster compared to the orignal 80% speed offset.

Recalculating... Ok.

Steamroller FX 8550 should then score 8,612

Still faster than i7 4770k by 5-8%.

FX 6300 at 4Ghz scores 5.0-5.1

FX 6550 4ghz should score 6.2-6.3

Faster than i5 2500k and on the same level as i5 4670k.

FX 8350 scores 1.1 points in single threaded, FX 8550 should score 1.3 ...

 

[guskline]

Recalculating... Ok.

Steamroller FX 8550 should then score 8,612

Still faster than i7 4770k by 5-8%.

FX 6300 at 4Ghz scores 5.0-5.1

FX 6550 4ghz should score 6.2-6.3

Faster than i5 2500k and on the same level as i5 4670k.

FX 8350 scores 1.1 points in single threaded, FX 8550 should score 1.3 ...

chernobog: all of your assumptions are based upon claims made about a core "steamroller" that no one has seen except perhaps testers. I sure hope this scenario doesn't play out like the Bulldozer debacle.

Where is the actual silicon? Where are the chips being tested?

One can blast Haswell all they want and make predictions about how a future chip could "crush" it ( your words not mine) but at least you can buy a Haswell 4770k chip right now and test it. The best AMD chip we can buy is the FX 8350. As for the release date for SteamRoller, has AMD officially designated a relase date or are these dates gleaned from the powerpoint slides?

 

[inf64]

Hopefully we will get some performance data from Far East as Kaveri is in the hands of the OEMs for some time now (according to Lisa Su).

 

[guskline]

This link http://www.extremetech.com/gaming/1...erclocks-on-tap-set-the-stage-for-steamroller shows that SteamRoller core chips with 4 core 2 module structure on a socket FM2+ could be released as early as fourth quarter 2013 but the same slide from AMD shows Vishera 8 core 4 module chips through 2013. Since a lot of the content of this thread is guess work I'll throw in my 2 cents worth. IF AMD releases an 8 core 4 module SteamRoller for the FX chipset it will probably require a BIOS upgrade and possibly a new chipset.

 

[ShintaiDK]

Recalculating... Ok.

Steamroller FX 8550 should then score 8,612

Still faster than i7 4770k by 5-8%.

FX 6300 at 4Ghz scores 5.0-5.1

FX 6550 4ghz should score 6.2-6.3

Faster than i5 2500k and on the same level as i5 4670k.

FX 8350 scores 1.1 points in single threaded, FX 8550 should score 1.3 ...

AMD would not release 220W TDP parts and call them 9000 series FX if something better was coming. Also, there is nothing at all pointing to that AMD will make more than 2M on the desktop. However there is pointing towards AMD ditching the AM3+ socket all together in favour of FM2+.

And your numbers are still wildly exaggerated.

 

[sniffin]

AMD would not release 220W TDP parts and call them 9000 series FX if something better was coming. Also, there is nothing at all pointing to that AMD will make more than 2M on the desktop. However there is pointing towards AMD ditching the AM3+ socket all together in favour of FM2+.

And your numbers are still wildly exaggerated.

~7% better IPC from Bulldozer->Piledriver with minimal changes to the die is possible but a node shrink and redesign of the module yielding a 20% IPC increase is exaggerated? lol

 

[ShintaiDK]

~7% better IPC from Bulldozer->Piledriver with minimal changes to the die is possible but a node shrink and redesign of the module yielding a 20% IPC increase is exaggerated? lol

A nodeshrink doesnt yield IPC increase. Also the focus is on the CMT part of the module. Meaning the penalty when using both "cores" in a module. Something they need fixed when they only supply 2M CPUs for the desktop with SR.

 

[Sheep221]

chernobog: all of your assumptions are based upon claims made about a core "steamroller" that no one has seen except perhaps testers. I sure hope this scenario doesn't play out like the Bulldozer debacle.

Where is the actual silicon? Where are the chips being tested?

One can blast Haswell all they want and make predictions about how a future chip could "crush" it ( your words not mine) but at least you can buy a Haswell 4770k chip right now and test it. The best AMD chip we can buy is the FX 8350. As for the release date for SteamRoller, has AMD officially designated a relase date or are these dates gleaned from the powerpoint slides?

In fact some testers and OEMs have the engineering samples for a long time, technically even few years ahead before we even start to hear some rumoring about them.

If you work oftenly with various CPUs you probably noticed that year imprinted on the CPU itself is way earlier than the actual CPUs were sold to general public.

For example all CPUs released in 2006 and 2007 by Intel carried the 2005 model year.
The Athlon 64 single core, Sempron 64, Athlon 64 X2 and Athlon 64-FX for 754 and 939 socket carried the 2001 label and AM2 ones carried the 2005 one.
The only exception is Core i7-920 which was released and had label in the same year, which was 2008.
Same goes for newer CPUs such as sandy bridge and ivy bridge which have 2009 and 2010 on them.
This means that actual prototypes of first dual core and quad core CPUs were constructed as early as 2003 or 2004 and single core 64-bit CPUs as early as 2000.

We don't know what exists and what works are going on inside semiconductor companies. No one who works there is allowed to say anything and no one would like to say that stuff anyway info leaks however can happen.
Once we hear about something is going to be released it's either officially announced on some developer forum or press conference or expo, or the ESes were used so much already that information about upcoming hardware was easily leaked.

So to answer your question, yes, steamroller is undergoing evaluation for a long time already.

 

[inf64]

AMD emphasized ST performance advances when it comes to SR core on last year's HC. It's not an "if" whether it will happen (independently from elimination of decoder bottleneck), the question is how much of an ST IPC increase are we going to see with the new core.

 

[Ajay]

Kaveri 4C @ 2.6 GHz -> I put it at ~5.3(5.1 to 5.4) pts for R11.5.
8C Steamroller @ 2.0 GHz -> I would put it at ~9(8.6 to 9.3) pts for R11.5.

Yikes! There were rumors that 28nm bulk was going to have low clocks, I guess they were true. No wonder AMD is trying to bulk up the IPC significantly. Geez, now they'd need a huge boost in IPC to match Intel's ST performance.

 

[krumme]

Yikes! There were rumors that 28nm bulk was going to have low clocks, I guess they were true. No wonder AMD is trying to bulk up the IPC significantly. Geez, now they'd need a huge boost in IPC to match Intel's ST performance.

I dont beliewe it. BD is a high freq design. Doing this is a castration of one of the fundamental principle of the design. Then nothing is left and cmt is washed completely out. I guess with mubadalas ATIC investment in GF anything is possible but this looks borderline reasonable.

 

[chernobog]

Yikes! There were rumors that 28nm bulk was going to have low clocks, I guess they were true. No wonder AMD is trying to bulk up the IPC significantly. Geez, now they'd need a huge boost in IPC to match Intel's ST performance.

So if that is the scenario that AMD is in it right now then I would not be shocked, but look at positive side now. Excavator will be a beast.

Excavator - TSMC 20nm - 2014/2015 - 6 core APU anyone?

 

[Abwx]

High frequency is still actual , even if it s not higher than PDriver.

 

[SiliconWars]

Rockchip's 28nm tablet SoC is capable of up to 1.8 GHz - http://globalfoundries.com/newsroom/2013/20130617.aspx

This is on GF's 28nm SLP.

 

[chernobog]

Rockchip's 28nm tablet SoC is capable of up to 1.8 GHz - http://globalfoundries.com/newsroom/2013/20130617.aspx

This is on GF's 28nm SLP.

You kinda fail, why?
- ARM chip / ARM9(mobile)
- This is still early 28nm from GF if I stand correct, 28nm is ready for mass volume in 2014
- Does this even have a mere connection to x86-64 chip? Nope.

So why are you mentioning then?

No personal attacks
-ViRGE

 

[Arachnotronic]

High frequency is still actual , even if it s not higher than PDriver.

Hey, that looks like 2x 128 bit FMACs!

 

[SiliconWars]

You kinda fail, why?
- ARM chip / ARM9(mobile)
- This is still early 28nm from GF if I stand correct, 28nm is ready for mass volume in 2014
- Does this even have a mere connection to x86-64 chip? Nope.

So why are you mentioning then?

Because I have a brain and can compare it to current TSMC 28nm A9's.

 

[chernobog]

Because I have a brain and can compare it to current TSMC 28nm A9's.

You have the experience but I don't see connection between ARM and x86-64. How would this reflect on Steamroller?

 

[Durp]

I bet there will be a lot of Intel fanboys shuffling to AMD when they see AMD being superior, fanboys are in the end back stabbing douches and I would not be suprised if you start yelling AMD is the best when it finally punches Intel in the face and take its crown, again.

That is called being a performance "fanboy" which is a non biased and normal position to take when buying a processor, this is normally a good thing. That is exactly what most of these so called Intel fanboys are. Performance fanboys who happen to own Intel processors because they happen to destroy anything AMD has on the market.

[snip]

Don't start
-ViRGE

 

[SiliconWars]

You have the experience but I don't see connection between ARM and x86-64. How would this reflect on Steamroller?

If we're talking about the process being at fault for Steamrollers supposedly low clocks, there doesn't seem to be much wrong with GF's 28nm based on their first A9's SLP clock speed. It's not great but it seems to be average for what the smaller node would bring.

Steamroller might come in at <3GHz clocks but I don't believe it will be the fault of GF's 28nm bulk process.

 

[Abwx]

Steamroller might come in at <3GHz clocks but I don't believe it will be the fault of GF's 28nm bulk process.

There s no way 2M 4C Kaveri will be at 3ghz given that there s 100W TDP versions , this would imply lower perf/watt than Richland.

http://www.extremetech.com/computing/159635-amds-future-apu-and-gpu-code-names-revealed-analyzed