Why AMD dumped the K10/Stars/10h Architecture ?

[emad_ramlawi]

* ALERT THIS IS A LONG RANT FROM A PERSON WHO LOVES AMD AND WISHES THEM TO GET BACK IN THE RIGHT TRACK,
WITH MANY LINKS TO PROVIDE EVIDENCE AND FACTS *

Sorry i just don't see any reason they stopped evolving that great architecture, particularly the Athlons family that had real cores
(bare in mind there are AMD athlons in production but they have a pile diver core instead like the AMD Athlon X4 750K)

I am taking about :

AMD Athlon II X4 640 Propus 3.0GHz Socket AM3 95W
http://www.newegg.com/Product/Product.aspx?Item=N82E16819103871

sells currently for 80 USD, cheap to produce and simple and is all around good performer in single thread and multi-threaded, tasks
thanks to its real cores.

Lets go back in time and check those :

http://www.amd.com/us/products/desktop/processors/athlon-ii-x2/Pages/AMD-athlon-ii-x2-processor-model-numbers-feature-comparison.aspx

notice AMD at one point of its time, gave a true meaning to its green color in there label, with there :

Energy Efficient AMD Athlon™ II X4 Quad-Core Processor :

620e 2.6 GHz 45nm SOI 2MB socket AM3 45W
615e 2.5 GHz 45nm SOI 2MB socket AM3 45W
605e 2.3 GHz 45nm SOI 2MB socket AM3 45W

release date September 2009 (Stepping C2) till 2011

And please look at this review :

http://www.anandtech.com/show/2861/9

Notice the 605e at idle : 110W
and at load : 137.7 W

Back at the time when AMD TDP mean something and was 100% Genuine, before Intel wrecked the TDP meaning all together, and
AMD sinking to intel new TDP definition and losing ...

Now since we went back in time, lets go forward with the latest iteration of the athlons, which was in the FM1 socket motherboards
(Liano Series), like the A8-3850.

Manufactures in 32nm, bare in mind as well that all FM1 chips, had 1/3 there size allocated for the internal GPU.

Please take a look at this pic for the die shot :

http://www.realworldtech.com/includes/images/articles/llano-1.jpg?b22ba0

now lets take the lowest quad core liano desktop chip to make an example the :

AMD A6-3650 Llano 2.6GHz Socket FM1 100W

Lets look at some power consumption reviews:

http://www.guru3d.com/articles_pages/amd_a6_3650_apu_review,9.html
http://www.legitreviews.com/article/1655/16/
http://hothardware.com/Reviews/AMD-A63650-Llano-APU-Performance-Review/?page=11

You Will come up with the conclusion that the A6-3650 idles about : 40~50 TDP
and at full loads ~140W

Which is comparable with an i5 core sandy bridge system give or take 10 TDP

That said as seen in the Die shop nearly half of the chip is dedicated to the IGP, and the AMD CPU cache was limited due to make
side for the internal Radeon GPU, which is much better than Intel® HD Graphics 2000 found in the i5 sandy bridge,
so i am pretty sure the IGP is sucking 30~40 Watts.

so the cores alone operate at 40W-100W which is very comparable to intel most power efficient core i3 series, which intel labels 65W or 55W,
as we will see now:

Intel Core i3-2100 Sandy Bridge 3.1GHz LGA 1155 65W consumes according to this review:

http://www.anandtech.com/show/4083/the-sandy-bridge-review-intel-core-i7-2600k-i5-2500k-core-i3-2100-tested/21

IDLE : 73 W
LOAD : 106 W

the Intel Core i3-3220 Ivy Bridge 3.3GHz LGA 1155 55W consumes according to this review:

http://www.bit-tech.net/hardware/2012/11/26/intel-core-i3-3220-review/7

IDLE : 46 W
LOAD : 78 W

And the above CPU was build using 22nm and not 32nm

Note that i didn't include the Phenoms in the above power efficient rant cause they are abit high on power usage, compared to intel, and i believe the athlons, if AMD worked on
them they would've produced great results, how the hell intel came up with Sandy Bridge do you think ??? only from years with working with the previous
Core family ...

Regarding performance, please check AMD Phenom II X4 965 which is build using the Stars/K10 arch,
and this CPU with its 3.4 GHZ stand toe to toe with Intel Core2 Quad Q9650 @ 3.00GHz

http://www.cpubenchmark.net/cpu.php?cpu=AMD+Phenom+II+X4+965&id=370
http://www.cpubenchmark.net/cpu.php?cpu=Intel+Core2+Quad+Q9650+%40+3.00GHz&id=1050

meaning that the GHZ difference was only 0.4 ~ 0.6 which i bet you noticed looking at game system requirements which is was like this usually :

Quad core intel running at 2.4 Ghz or Quad Core AMD 3.0 ghz.

Nowadays thanks to the Bulldozer/Pilediver/GayJustinBieber Arch amd needs to be running at 4.0 Ghz
to compete with intel 3.0 Ghz which is a whopping 1.0 Ghz difference, and the new AMD 5.0 Ghz processors will prove me right(er).

And i don't want to start with the power efficiency of the Bulldozer/Pilediver/GayJustinBieber, or poor single threaded performance.

Oh and the performance of current Liano FM1 APU in the CPU side, still beats FM2 generation of processors in multitasking environments.

so why did AMD start the new arch ???? and abandoned looking up to the stars.

 

[DComander1x]

Im sure AMD had to, the K10 was a tweak of the K8 arch from around 2004, and Intel, at the time of Nehalem, had a slight lead on AMD, so they needed to come out with something more competitive to Intel, and faster on the per thread level and higher clock speed as well.

 

[8350rocks]

Why not pose this question instead:

Why has Intel not abandoned an archaic architecture that they've been running concurrently since the late 90s with a tweak each generation?

Is it lack of innovation? Is it lack of forward thinking?

The reality is, those "real cores" would not compete this day and age. In reality, things are going more and more multithreaded.

For example...Jaguar "small cores" from AMD, are in the low power segment and you don't see any of those with a higher TDP than around 45W or so. Those are based on K10/stars architecture...and they don't compete in the same areas the new PD architecture does.

While it's a great APU for what it does, it isn't something you could effectively scale well. Consider that the Phenom II architecture high end stuff had a TDP of 140W. The highest end FX series now has 125W TDP and destroys the old K10 architecture in many benchmarks if you look at Piledriver. With Piledriver, even single threaded performance is as good or better than the old K10 architecture.

They were running out of headroom and saw the writing on the wall that Intel refuses to acknowledge. That's why 5th Gen Intel architecture won't be coming next year, neither will the die shrink to 14nm. Intel is hitting the "wall" AMD foresaw.

Now, how will this turn out for them? They are turning around right now, and from a cash flow/cost perspective, they are actually healthier now than they were during K10 architecture.

 

[DComander1x]

Why not pose this question instead:

Why has Intel not abandoned an archaic architecture that they've been running concurrently since the late 90s with a tweak each generation?

Is it lack of innovation? Is it lack of forward thinking?

The reality is, those "real cores" would not compete this day and age. In reality, things are going more and more multithreaded.

For example...Jaguar "small cores" from AMD, are in the low power segment and you don't see any of those with a higher TDP than around 45W or so. Those are based on K10/stars architecture...and they don't compete in the same areas the new PD architecture does.

While it's a great APU for what it does, it isn't something you could effectively scale well. Consider that the Phenom II architecture high end stuff had a TDP of 140W. The highest end FX series now has 125W TDP and destroys the old K10 architecture in many benchmarks if you look at Piledriver. With Piledriver, even single threaded performance is as good or better than the old K10 architecture.

They were running out of headroom and saw the writing on the wall that Intel refuses to acknowledge. That's why 5th Gen Intel architecture won't be coming next year, neither will the die shrink to 14nm. Intel is hitting the "wall" AMD foresaw.

Now, how will this turn out for them? They are turning around right now, and from a cash flow/cost perspective, they are actually healthier now than they were during K10 architecture.

Exactly, though Jaguar, is essentially bobcat, which is basically a modified Athlon XP with 64 bits, hypertransport, tech tweaks and a lower voltage, though thats what I got from what AMD said.

 

[emad_ramlawi]

hmmm interesting answers, that made me re-consider...

I work at software company, we released a product that can actively use multi-threads and logical cores, i reckon we can utilize up to 8 logical cores in one process, and own application can spawn in its process 30 threads ...

BUT this software was released in 2012, and was major release for my company, and think about that for a second only in 2012 newly released windows software (Zoolz) can utilize many threads, what about all the rest of the stuff that rely on single core performance ....

my main point was that AMD could've created an 8 core K10 athlon and will consume possibly less than 100W

if this athlon k10 quad core CPU ----> 605e 2.3 GHz 45nm SOI 2MB socket AM3 45W

Moving to 28nm i reckon its feasible to create that by doubling the above CPU, that will score good in multi threaded as do pilediver CPU`s, as well as good in single threaded if AMD wants to make an 125W version of it and make it Turbo Boost to higher frequencies, and be able to hit 3.3 Ghz when only 4 cores are taxed or something.

Also when i look at the AMD bulldozer/pilediver server cores, i kinda see what you talking about:


Model Number Cores Core Speed AMD Turbo CORE L3 Cache TDP Socket Type

6380 16 2.5GHz 3.4GHz 16MB 115W G34


The above is impressive, its Pilediver server core and i choose it cause its TDP resembles a desktop abit, and in server releases, you can fully see AMD latest innovations.

16 threads, thats pretty neat and impressive, i reckon using 28nm we can easily reach 20 threads ... which if implemented properly can change
our way of compute, and start using different types of application that are coded to use those threads properly .


But it all boils down to software .

 

[Cazalan]

so why did AMD start the new arch ???? and abandoned looking up to the stars.

Internal politics and engineer pride. At some point the lead designer will want his own chip that he championed from "scratch".

 

[8350rocks]

But it all boils down to software .

BINGO! How optimized the software becomes for multithreading makes all the difference in the world.