Will AMD Steamroller be their Sandy Bridge?

[Vesku]

FX 4300 vs i5 750
http://www.anandtech.com/bench/Product/700?vs=109

FX 6300 vs i7 860
http://www.anandtech.com/bench/Product/699?vs=108

FX 8350 vs i7 950
http://www.anandtech.com/bench/Product/697?vs=100

You can see here that with Vishera FX series they are roughly at the Nehalem level when comparing SKUs. A FX-4300 matches up pretty closely to an i5-750 and so on up the SKU chart. The problem is that Bulldozer was supposed to have at least that level of performance. AMD has been perpetually been behind schedule since the original Phenom.

It is actually slower in a few things in that link but really a lot of that is due to what being 20% of the market vs 80% of the market means for developer support than anything that can be designed for. After all AMD can't legally copy Intel's chips method of instruction support transistor for transistor, they can't even legally support every instruction set. AMD also can't control how things are coded and compiled, they don't have the market share and corresponding $ war chest.

If Steamroller is a 15-20% improvement and GlobalFoundries 28nm process performs better than their 32nm did at launch then wouldn't that make Steamroller FX SKUs similar to Sandy Bridge?

 

[Edrick]

If Steamroller is a 15-20% improvement

Fairly big assumption as it has not been the case for AMD in some time.

and GlobalFoundries 28nm process performs better than their 32nm did at launch then wouldn't that make Steamroller FX SKUs similar to Sandy Bridge?

Another fairly big assumption as history has shown us otherwise.

 

[SPBHM]

SB had a good IPC gain AND significantly higher clocks with lower power usage, I don't expect the same jump from AMD this time.

 

[SiliconWars]

Fairly big assumption as it has not been the case for AMD in some time.

Short memory? http://www.anandtech.com/show/6396/the-vishera-review-amd-fx8350-fx8320-fx6300-and-fx4300-tested/9

Thanks to architectural and frequency improvements, AMD delivers up to 20% better performance than last year's FX-8150 for a lower launch price, while remaining within the same thermal envelope.

Another fairly big assumption as history has shown us otherwise.

There is no history of GlobalFoundries doing a half-node.

 

[Vesku]

Fairly big assumption as it has not been the case for AMD in some time.

http://www.anandtech.com/bench/Product/434?vs=697

Bulldozer to Piledriver, no node shrink. GlobalFoundries having a good 28nm introduction is probably a big assumption, though.

 

[Edrick]

LOL

BD was actually a step backward so fixing it and putting out a better CPU in the realm of 20% improvement is nothing to write home about. Look at the past 5 years before that.

GF always seems to have issues with new processes. Im not saying they can't pull it off, but I would be surprised.

Im just saying both are big assumptions. And for this to be the case, both need to come true, not just one.

 

[guskline]

Vesku: AMD ought to hire you for their cpu marketing department (just kidding!). With AMD's recent history, I don't see that large of an improvement unless there is a radical change in core design and socket. I believe CEO Cory when he stated he was getting out of the high end desktop market.

 

[Vesku]

Vesku: AMD ought to hire you for their cpu marketing department (just kidding!). With AMD's recent history, I don't see that large of an improvement unless there is a radical change in core design and socket.

I'm not sure they'd like the characterization as perpetually late, even if it's true.

Edrick has a point that they have to hit 2 targets they have had trouble with performance and power consumption.

 

[JimmiG]

Intel is a stationary target, just like AMD was when Intel launched Conroe.
However this only applies to the performance desktop market, which is about to become totally irrelevant anyway.

 

[Vesku]

Intel is a stationary target, just like AMD was when Intel launched Conroe.
However this only applies to the performance desktop market, which is about to become totally irrelevant anyway.

Well it's not shrinking from $Billions to $0 in just a few years, so I'd imagine AMD would be OK with gaining more and more of the desktop market share.

 

[sushiwarrior]

With AMD's recent history, I don't see that large of an improvement unless there is a radical change in core design and socket.

I fail to see how "their recent history" is helping your point, when their recent history showed a 20% jump without a radical change in core design or socket.

It's very possible for AMD to pull ahead given that Bulldozer/Piledriver were shown to have several weak points (cache, pipeline, FP duality, memory controller). Fix those weakpoints and you have some easy performance gains. Although from the sounds of it, AMD will never be able to release as "perfectionist" designs as Intel due to the lack of manpower.

 

[Edrick]

Well, I am just waiting for Windows to run on this and I will be happy

http://en.wikipedia.org/wiki/IBM_zEC12_(microprocessor)

 

[mrmt]

If Steamroller is a 15-20% improvement and GlobalFoundries 28nm process performs better than their 32nm did at launch then wouldn't that make Steamroller FX SKUs similar to Sandy Bridge?

Unless you assume that AMD engineers are outright idiots, they wouldn't leave such a low hanging fruit there after having to rework Bulldozer for almost two years after the 2010 debacle. They probably opted for the smaller decoder because the heat and die space it would yield wouldn't justify the performance increase they would get.

And then comes Steamroller. The only way you could get 15%-20% increase in IPC is through a drastic reworking of the architecture, like forgeting the shared FPU or adding a third ALU in AMD corem and even here they would lose a lot of clock speed. So far we heard nothing like that, just a few tweaks on top of the bigger decoder, all that in a process from a subpar foundry that is very much late.

If you are expecting a Sandy Bridge out of AMD efforts, you will have another Randy "40%" Allen or another John "IPC won't go down" Fruehe moment.

 

[Vesku]

I fail to see how "their recent history" is helping your point, when their recent history showed a 20% jump without a radical change in core design or socket.

It's very possible for AMD to pull ahead given that Bulldozer/Piledriver were shown to have several weak points (cache, pipeline, FP duality, memory controller). Fix those weakpoints and you have some easy performance gains. Although from the sounds of it, AMD will never be able to release as "perfectionist" designs as Intel due to the lack of manpower.

Just want to point out that Sandy Bridge is not ahead of Ivy Bridge nor Haswell with stock SKUs.

 

[sushiwarrior]

Unless you assume that AMD engineers are outright idiots, they wouldn't leave such a low hanging fruit there after having to rework Bulldozer for almost two years after the 2010 debacle. They probably opted for the smaller decoder because the heat and die space it would yield wouldn't justify the performance increase they would get.

And then comes Steamroller. The only way you could get 15%-20% increase in IPC is through a drastic reworking of the architecture, like forgeting the shared FPU or adding a third ALU in AMD corem and even here they would lose a lot of clock speed. So far we heard nothing like that, just a few tweaks on top of the bigger decoder, all that in a process from a subpar foundry that is very much late.

I don't see any low hanging fruit that Haswell targeted, yet it achieved ~10% IPC boosts on an already seasoned/developed architecture. Getting 10-20% IPC boosts doesn't mean that there must be glaring holes in the architecture, it just means that you can always improve on efficiency given more time to work on the architecture. AMD has even more chance to, because they have problems in the aforementioned areas (mem controller, cache, etc.) which aren't low hanging fruit, but rather difficult problems which could give massive performance increases.

Dare I say that a processor with AMD core/module arch and Intel mem controller/cache would be far faster than anything currently out there.

 

[Vesku]

If you are expecting a Sandy Bridge out of AMD efforts, you will have another Randy "40%" Allen or another John "IPC won't go down" Fruehe moment.

I'm not claiming that level of improvement I'm asking people's opinions on the possibility.

 

[mrmt]

I don't see any low hanging fruit that Haswell targeted, yet it achieved ~10% IPC boosts on an already seasoned/developed architecture.

Intel track record and R&D budget are far different than those of AMD. You can't compare the two.

 

[mrmt]

I'm not claiming that level of improvement I'm asking people's opinions on the possibility.

Then low, very low, way low.

 

[Abwx]

FX 4300 vs i5 750

If Steamroller is a 15-20% improvement and GlobalFoundries 28nm process performs better than their 32nm did at launch then wouldn't that make Steamroller FX SKUs similar to Sandy Bridge?

They stated 30% more operations per cycle re bulldozer
at the module level, that s about 20% re Piledriver.

To reach this number they will undoubtly increase
the module efficency when two threads are executed
but since they cant have 100% scaling in this matter
they ll have to increase single thread efficency by
an amount that can be barely predicted as the less
scaling is increased the more ST perf must be increased.

Based on Piledriver numbers we can expect the said 20%
at the module level and up to 15% on single thread/module.

 

[sushiwarrior]

Intel track record and R&D budget are far different than those of AMD. You can't compare the two.

A significant portion of Intel's R&D is spent on fab technology and design, so not as much of it as you think is spent on chip design.

But that is besides the point, you're saying that it can't be done without radical core design changes, I'm saying that it's certainly achievable, it's just a question of whether AMD can pull it off.

 

[mrmt]

But that is besides the point, you're saying that it can't be done without radical core design changes, I'm saying that it's certainly achievable, it's just a question of whether AMD can pull it off.

When did you see 15-20% average IPC increase from one generation to another?

 

[sushiwarrior]

When did you see 15-20% average IPC increase from one generation to another?

From IB to Haswell was 10% on a mature arch, from Bulldozer to Piledriver was 20% on a weak arch. I think 15% (middle ground) is a perfectly viable outcome.

 

[Vesku]

Core 2 to Nehalem?

Intel Core 2 Quad Q8400 - 2.66GHz - 4MB L2
vs
Intel Core i5 750 - 2.66GHz - 1MB L2 - 8MB L3

http://www.anandtech.com/bench/Product/89?vs=109

 

[Schmide]

Core 2 to Nehalem?

Intel Core 2 Quad Q8400 - 2.66GHz - 4MB L2
vs
Intel Core i5 750 - 2.66GHz - 1MB L2 - 8MB L3

http://www.anandtech.com/bench/Product/89?vs=109

It's all about the cache dude. 12mb vs 10mb

Intel Core 2 Quad Q9450 - 2.66GHz - 12MB L2

vs

Intel Core i5 750 - 2.66GHz - 1MB L2 - 8MB L3


http://www.anandtech.com/bench/Product/51?vs=109

 

[Vesku]

I still see a lot of 20+% increases going from Q9450 to i5 750.