[Digitimes] Kaveri delayed

[nehalem256]

Because this "Intel fan" (and AMD shareholder) understands that AMD is sinking untold millions into R&D for big chips that don't sell. That money can be saved since the Jaguar chips will probably sell like gangbusters and actually be profitable.

While everybody else cheers that AMD's chips are super duper cheap, I cringe knowing that the company is selling 315mm^2 dies into <$200 chips.

Except sadly due to the WSA with GF AMD wouldn't save money selling smaller cores.

 

[pablo87]

If true, its not surprising. AMD is juggling a lot of balls with relatively limited resources.

Not good news for s/h as the WSA is the financial elephant in the room - Kaveri = volume * bigger die. Will Richland tide them over until then? Yes but at firesale prices...

The thing is, this could have been easily avoided by spinning a version of Kabini with a kaveri like gpu and dual channel memory...

 

[Arachnotronic]

Except sadly due to the WSA with GF AMD wouldn't save money selling smaller cores.

Move the smaller cores and GPUs to GloFo. Done.

 

[Erenhardt]

Move the smaller cores and GPUs to GloFo. Done.

Then GPU department will also loose to nvidia. What will be left from AMD then? Memory sticks? (http://www.amd.com/pl/products/desktop/radeon-memory/Pages/system-memory.aspx)

 

[Ventanni]

I feel like we could get a significantly better product out of AMD if they weren't contractually obligated to partner with GF. How can AMD compete if their foundry can't either?

 

[Arachnotronic]

Then GPU department will also loose to nvidia. What will be left from AMD then? Memory sticks? (http://www.amd.com/pl/products/desktop/radeon-memory/Pages/system-memory.aspx)

Wasn't GloFo giving out powerpoints about how it's going to have 14nm in 2014? lol

 

[ShintaiDK]

Wasn't GloFo giving out powerpoints about how it's going to have 14nm in 2014? lol

They sl^Hhipped 28nm since 2010 :awe:
And should have shipped 22nm since Q3 2012 and 20nm since Q4 2012 :biggrin:

 

[Idontcare]

I feel like we could get a significantly better product out of AMD if they weren't contractually obligated to partner with GF. How can AMD compete if their foundry can't either?

AMD is currently structured to basically have none of the benefits of being fabless (negotiating power, choice of foundry, flexibility in commitments to production capacity, etc) while still having all the downsides of being an IDM (contractual obligations to subsidize R&D, contractual obligations to pay for a fab whether they use it or not, etc).

In short their current business structure is exactly the kind of structure you would expect to be a recipe for economic disaster in an industry where its competitors are free to enjoy the benefits of being fabless or to take full advantage of being an IDM.

AMD gets to do neither, but does still have to deal with the worst of both worlds.

 

[Vesku]

They sl^Hhipped 28nm since 2010 :awe:
And should have shipped 22nm since Q3 2012 and 20nm since Q4 2012 :biggrin:

Yes, it's pretty difficult to do an internet search to see the current news on GF 28nm because the results are filled with almost 3 years of GF announcements that "28nm is in production".

I'll be pleasantly surprised if Kaveri is available in retail before next year.

 

[Fox5]

AMD is currently structured to basically have none of the benefits of being fabless (negotiating power, choice of foundry, flexibility in commitments to production capacity, etc) while still having all the downsides of being an IDM (contractual obligations to subsidize R&D, contractual obligations to pay for a fab whether they use it or not, etc).

In short their current business structure is exactly the kind of structure you would expect to be a recipe for economic disaster in an industry where its competitors are free to enjoy the benefits of being fabless or to take full advantage of being an IDM.

AMD gets to do neither, but does still have to deal with the worst of both worlds.

If I was an AMD employee seeing management pull games like that, I'd have jumped ship to pretty much anywhere else. I hear ARM picked up a lot of AMD employees.

In fact, I was at a conference in Texas recently where someone said that AMD used to be the lifeblood of the local microprocessor industry, and now it's pretty much all switched over to being ARM focused. There were lots of researchers in attendance from ARM, AMD and Intel. AMD research seems solid, but undermanned. Intel seemed to be a bit of a disarray about the management upheavals, but the ARM seemed to be completely in control.

 

[beginner99]

But they dont build less, they build more.

No idea but maybe because broadwell will be a true SOC meaning PCH is in the chip itself and hence chips are bigger and more factories are required.

For me it just makes sense to use the old fabs for desktop.

 

[nehalem256]

Move the smaller cores and GPUs to GloFo. Done.

They tried to move the small cores to GF(28nm Bobcat). Cancelled.

 

[SlickR12345]

Not a big deal, I mean yeah its always great to have these products in time, but in reality every company is late, Intel has been late so many times and outright wrong.

I mean their presentations in 2010 showed that we'd have 5x the performance from then and in reality we got about 2x the performance from then, less power consumption, but in terms of performance its been slow.

That said AMD need to get rid of Bulldozer and get back with the Phenom II line, those processors were really good and if they just kept developing them it would have been great. Instead they went for Bulldozer and now their processors suck for the most part.

They need to bring in strong desktop competition. I mean imagine a refined Phenom on 22nm, better channels, turbo mode, new expanded instructions set, more cache and faster lines, running on 3.5GHz and the ability to overclock on all models easily up to 5GHz.

That would be a produce I would buy over Intel, especially in the $250 segment and lower.

 

[NTMBK]

That said AMD need to get rid of Bulldozer and get back with the Phenom II line, those processors were really good and if they just kept developing them it would have been great. Instead they went for Bulldozer and now their processors suck for the most part.

Oh god, not this myth again!

Llano used an improved version of the Phenom II core on GloFo 32nm. Trinity then used Piledriver on GloFo 32nm. Trinity was an improvement over Llano in CPU performance.

 

[jacktesterson]

Oh god, not this myth again!

Llano used an improved version of the Phenom II core on GloFo 32nm. Trinity then used Piledriver on GloFo 32nm. Trinity was an improvement over Llano in CPU performance.

This.

 

[PPB]

I always wondered the die size it would take to make a L3 cache-less version of vishera on current 32nm node, considering the marginal loss of performance it gives comparing Trinity and vishera at same clocks.


I know it has all that L3 cache because it was just the same die for server and FX parts, but considering big CPU's strategy out of server line to let in the APU line in SR generation parts, I dont think it would be a bad idea.

 

[Fox5]

Oh god, not this myth again!

Llano used an improved version of the Phenom II core on GloFo 32nm. Trinity then used Piledriver on GloFo 32nm. Trinity was an improvement over Llano in CPU performance.

At the end of the day, all that matters is performance per watt. Trinity had resonant clock meshing which llano did not, plus an improved gpu.
At 100W, llano had 4 cores running at 3ghz and 400 radeon cores running at 600mhz.
At 100W, Trinity has 2 modules running at 3.8ghz and a gpu with 384 cores running at 800mhz.

Looking at a 2.9ghz llano versus a 3.8ghz trinity, there is a 31% advantage in clock speed.
http://www.anandtech.com/bench/Product/399?vs=675

AMD accounted for a ~10% clock speed boost from resonant clock meshing.
Cherry picking a few benchmarks:
Trinity is 29% faster in sysmark 2012 overall.
22% faster in adobe cs4.
12.7% faster in divx encode.
5% faster in x264 encode 1st pass.
20% faster in x264 encode 2nd pass.
12.5% faster in windows media encoder 9.
9.5% faster in 3dsmax cpu test.
20.5% faster in cinebench r10 single thread.
8% faster in cinebench r10 multithread.
4% faster in POV-Ray SMP benchmark.
7.6% faster in excel monte carlo.
-<1% slower in sorenson squeeze.
23% faster in winrar.
21% faster in cinebench 11.5 single thread.
-5% slower in cinebench 11.5 multi thread.
-5.5% slower in dragons age.
-2% slower in dawn of war 2.
6% faster in WOW.
22.5% faster in starcraft 2.

My conclusions:
A hypothetical resonant clock mesh llano would have traded blows with trinity, but trinity would generally have been faster.
Single threaded performance of trinity is way higher than llano.
Multithreaded performance of trinity roughly on par with llano, but shows the weakness of the module design compared to dedicated cores. This is especially so in apps that stress the fpu, more so if they're legacy apps that don't take advantage of the improved FPU.

And observations:
The Stars core was designed around needing an L3 cache to share data for multithreading. The Bulldozer/Piledriver module has a shared L2 cache between the integer cores, and a more powerful FPU to make up for only having one.
At low clock speeds, Kabini with its shared L2 cache would probably destroy Trinity in multi threaded benchmarks.
Trinity is a larger die than llano, so it is slightly more expensive.

Compared to Phenom II:
It's hard to say how much power the igpu consumes. However, I feel that a 65W Phenom II is a reasonable comparison to a 100W Llano.
That gets you the 2.6ghz Phenom II X4 910 versus a 2.9ghz (or up to 3ghz) llano.
http://www.anandtech.com/bench/Product/399?vs=85
In comparison, Llano is almost always faster, only in the occasional multithreaded benchmark does Phenom II match or exceed Llano's performance.


Now, Kabini compared to Trinity:
http://www.anandtech.com/show/6974/amd-kabini-review/5
Trinity here has a 53% clock speed advantage. The top end Kabini at 25W would match Trinity's clock speed at 25W.

PCMark 7:
Trinity only 43.5% faster

Cinebench R11.5 single thread:
Trinity is 79% faster.

Cinebench R11.5 Multi:
Trinity is only 36% faster than Kabini.

x264 pass one:
Trinity is 62% faster.

Pass 2:
Trinity is 48% faster.

Conclusion:
Kabini compares favorable to Trinity in multi threaded performance, which was already Trinity's achilles heel. But in terms of single threaded performance, the Piledriver core seems to easily have the best performance per watt of any AMD 32nm architecture.

 

[Soulkeeper]

I wouldn't trade my llano for a trinity/richland even if it cost me no $.

 

[NTMBK]

I wouldn't trade my llano for a trinity/richland even if it cost me no $.

Why? Improved CPU performance, significantly improved GPU performance, equal power use at load and lower power use at idle- what's not to like?

 

[unon]

If you look at various reviews online, llano cpu only power consumption was rougly equal to 3ghz pII / AthII. So I'm not sure good things can be said about glofo 32nm, at start at least. Btw trinity doesn't have rcm(or it's disabled), maybe richland does. The improved frequency is due to improved 32nm and bulldozer.

 

[JDG1980]

AMD is currently structured to basically have none of the benefits of being fabless (negotiating power, choice of foundry, flexibility in commitments to production capacity, etc) while still having all the downsides of being an IDM (contractual obligations to subsidize R&D, contractual obligations to pay for a fab whether they use it or not, etc).

In short their current business structure is exactly the kind of structure you would expect to be a recipe for economic disaster in an industry where its competitors are free to enjoy the benefits of being fabless or to take full advantage of being an IDM.

AMD gets to do neither, but does still have to deal with the worst of both worlds.

This raises the obvious question: why did AMD ever agree to a horrible deal like that in the first place? What was in it for them? Was Hector Ruiz really that stupid, or was there some reason to think this would be a good idea at the time that just didn't pan out?

 

[SiliconWars]

It's simple really, they didn't have enough money to pay for the fabs or the 3,000 workers. GF wanted to ensure they had business for those fabs, and I'm sure AMD was more than happy to basically continue what they had except on a customer basis. The only problem has been the collapse of the PC market meaning AMD has had to pay penalties, but to be fair to GF they are quite willing to amend the agreement every year.

 

[pelov]

This raises the obvious question: why did AMD ever agree to a horrible deal like that in the first place? What was in it for them? Was Hector Ruiz really that stupid, or was there some reason to think this would be a good idea at the time that just didn't pan out?

It panned out as expected. Shortly after selling off GloFo, Hector Ruiz became chairman of GloFo.

Hector Ruiz isn't stupid but those in charge of AMD at the time certainly were

 

[frozentundra123456]

This raises the obvious question: why did AMD ever agree to a horrible deal like that in the first place? What was in it for them? Was Hector Ruiz really that stupid, or was there some reason to think this would be a good idea at the time that just didn't pan out?

Weren't they pretty much forced into it by debt from the ATI acquisation? I was a big fan of the purchase at the time, but the benefits have been slow to materialize. At best they paid way too much and were slow to market with products that utilized the tech. If I recall correctly, Intel even was first out with an igp, although admittedly a lousy one. And we still don't have a true HSA chip until Kaveri, which looks like sometime in 2014 now.

 

[Xpage]

Hector Ruiz was really that stupid, and greedy too.