AMD Bulldozer Finally!

Scougar · 28 Sep 2011 at 14:11

So what's the latest speculation for BD performance (comparable to which Intel Product perhaps?) - Not trolling, genuinely interested as trying to decide whether to get SB now, or wait and see with BD.

Martini1991 · 28 Sep 2011 at 14:36

Scougar said:
So what's the latest speculation for BD performance (comparable to which Intel Product perhaps?) - Not trolling, genuinely interested as trying to decide whether to get SB now, or wait and see with BD.

In heavily multithreaded apps, for the 8150 inbetween the 2500k and 2600k.
Likewise it should be price wise.

SeraphX · 28 Sep 2011 at 14:38

not really sure, a load of screens surfaced from Donimhaber last week? stating that apparently BD crushes i5 and in some cases challenges i7 and SB but no processor specifics were given about the amd chip in question and nothing on the screens lends credit about them originating from AMD, i'd say they were fanboy made tbh

Martini1991 · 28 Sep 2011 at 14:39

Seraph said:
not really sure, a load of screens surfaced from Donimhaber last week? stating that apparently BD crushes i5 and in some cases challenges i7 and SB but no processor specifics were given about the amd chip in question and nothing on the screens lends credit about them originating from AMD, i'd say they were fanboy made tbh

They were from the AMD PR event as the rumour goes.
Of course it should crush the i5 in the shown programs, it's executing twice the threads. If it was fanboy made it'd be crushing the 2600k.
Though I agree, they look like they've been made using microsoft publisher.

drunkenmaster · 28 Sep 2011 at 18:16

djkav said:
No not really. The only performance hit is if a single thread requires the full 256bits in the FPU. It can only execute 1 thread at that moment.

A module has 2 integer schedulers and the FPU can perform 2x 128bit operations at the same time. That's enough for the module to perform 2 threads in parallel.

If most operations are only 128bit, which they probably are. Then BD should have strong processing throughput.

Most FPU instructions are 32bit, many are 64bit, very few are 128bit, almost none are 256bit.

AVX is what Intel has and AMD is getting, it allows you to wrap several smaller "bit" instructions into one big instruction. On a non AVX(old style) FPU unit, if you only had a 32bit fpu instruction to run that clock, it would take up the entire 128bit execution unit, you could not put anything else in. With AVX, you can wrap, in a 256bit AVX instruction, 2x128bit, 4x64 or 8x32bit FPU instructions.

So worst case vs best case is 2x128bit old fpu units could process just 2x32bit instructions on one clock, while a single 256bit avx instruction can process 8 entire 32bit instructions in one clock, you've got 4x the throughput with the most common fpu instructions. 64bit, you've still got twice the throughput of the old fpu units, and 128/256bit without AVX are basically unused, with AVX, it should become common but only in processing more small instructions FAR quicker.

Basically in 99.999999% of instructions the new single shared FPU unit will be significantly faster than the FPU available on two old "full" cores anyway.

Of course, until AVX is more widely used and updated for that advantage(for Intel and AMD) isn't there. This is an architecture moving to Fusion where FPU off die will eventually end up dealing with a lot of the FPu workload...... unfortunately we've seen in computing that moving to multithread, and moving to true gpu acceleration has been a crazy slow process, AVX will be used quickly, no reason not to, its pretty easy and has huge performance advantages, fusion/fpu sharing will be a while coming

pmc25 said:
I doubt if precision ever exceeds 64bit in games. Presumably it's 32bit most of the time.

I suppose some of the physics might exceed that, but I doubt if it'd be needed.

Pretty much, physics is unlikely to aswell, as much as Nvidia want you to believe, the eye can't tell if that guy you shot off the side of a tall building is falling at 9.812983927392m/s2 or 9.8, or even 10ms. Estimation and lack of precision for physics makes things simple, Physx is completely needlessly complex which just wastes effort.

Martini1991 said:
The design before Bulldozer for a dual core is Callisto, how is that being negative? If it performs worse than 4 Callisto CPU's, then how is that a step forward for performance?

Quite simply, because you can't buy a 4 callisto CPU on one die right now. Theres really nothing more complex to it than that. IF(and it won't be) it was only 4x as fast to the dot as a dual core callisto in some benchmark that uses every extra thread effectively, thats the point, its 4 times faster, because it HAS 8 cores, and you can't buy an 8 core AMD cpu. If its exactly 33% faster than a hexcore Phenom 2, purely down to two extra cores, then again, its 33% faster.

Its kind of the equivilent of saying the 5870 isn't a step forwards, because each SP is just as fast as each SP in a 4870, theres only twice as many, how is that moving performance forward. Given the wrong super old game thats entirely cpu limited, its no faster than some old graphics core, given most newer software, its twice as fast as a 4870.

IF AMD were releasing 6/4/2 core Bulldozers that were exactly no faster than Phenom 2 6/4/2 cores, you'd have a point, but they are releasing 8/6/4 core Bulldozers, and across the range at any given price range, you're likely to be buying a chip that has at LEAST 33% more performance for the price.

Ivy is going to be 10% faster than the Sandybridges they replace, does that make them worthless?

pspcpog · 28 Sep 2011 at 20:16

Haven't logged in to the forum for some time now and the
situation is same again. Guess I'll check back in a month
time then

sunama · 28 Sep 2011 at 21:02

Can I ask what all the excitment regarding BD is?
I ask this because for all intent and purposes, it would appear that the top line BD cpus should fall somewhere between 2500k and 2600k. Once the Sandybridge CPUs are overclocked, they stomp all over the BD cpus.

So, IF the above is true, why are so many people waiting for BD? Surely it is better to buy a faster (SB) cpu now, than wait 3 months for the slower, BD?

Martini1991 · 28 Sep 2011 at 21:04

sunama said:
Can I ask what all the excitment regarding BD is?
I ask this because for all intent and purposes, it would appear that the top line BD cpus should fall somewhere between 2500k and 2600k. Once the Sandybridge CPUs are overclocked, they stomp all over the BD cpus.

So, IF the above is true, why are so many people waiting for BD? Surely it is better to buy a faster (SB) cpu now, than wait 3 months for the slower, BD?

Don't ask questions.

opethdisciple · 28 Sep 2011 at 21:12

sunama said:
Can I ask what all the excitment regarding BD is?
I ask this because for all intent and purposes, it would appear that the top line BD cpus should fall somewhere between 2500k and 2600k. Once the Sandybridge CPUs are overclocked, they stomp all over the BD cpus.

So, IF the above is true, why are so many people waiting for BD? Surely it is better to buy a faster (SB) cpu now, than wait 3 months for the slower, BD?

Its all just speculation at the moment. Nobody knows where these cpus fall interms of performance.

Waiting just gives you more options at better prices.

sunama · 28 Sep 2011 at 21:25

Is there a concrete release date for Buldozer yet?

opethdisciple · 28 Sep 2011 at 21:26

I thought there was something happening on Oct 12th?

drunkenmaster · 28 Sep 2011 at 22:41

sunama said:
Can I ask what all the excitment regarding BD is?
I ask this because for all intent and purposes, it would appear that the top line BD cpus should fall somewhere between 2500k and 2600k. Once the Sandybridge CPUs are overclocked, they stomp all over the BD cpus.

So, IF the above is true, why are so many people waiting for BD? Surely it is better to buy a faster (SB) cpu now, than wait 3 months for the slower, BD?

Because Bulldozer can overclock? Because theres been no hard proof of performance, because its the first in a long line of chips on a new architecture, which is (almost) always when an architecture is least efficient.

Sandybridge, call a 2600k 100% performance, where is Phenom 2 right now, and for most people who bought Phenom 2's, a couple years ago with quad cores, you're moving from a quad core, say at 60% of 2600k speed, to something with 95% of the speed.

If there is genuine competition then Intel might drop prices, they might not, Bulldozer mk1 isn't the most interesting, Piledriver is better but not hugely better, the third version due 2013 is being hinted at being epic(so might Haswell).

Generally speaking, its not so much about the performance, people like new tech, which is good.

For AMD users, a single bulldozer will be MUCH cheaper than a Sandybridge + mobo, just drop in a CPU to an existing setup.

Theres still quite a lot of potential for Bulldozer to beat Sandybridge quite a lot in certain benchmarks, which is something Phenom 2, even hexcores, really couldn't do.

Zarf · 28 Sep 2011 at 23:41

Scougar said:
So what's the latest speculation for BD performance (comparable to which Intel Product perhaps?) - Not trolling, genuinely interested as trying to decide whether to get SB now, or wait and see with BD.

Gaming wise SandyB is probably going to bench better on todays games, but considering my antique Q6600 is still coping fine either of them will handle todays games with ease. Things will probably be closer on future games which are more heavily threaded.
Overclocking potential looks to be equal, and price/performance will be similar, probably with a slight edge to AMD once motherboard costs are counted.
Where things will get interesting is the FX8130, this should be similar in price to a 2500K, but since it has all 8 cores if it overclocks like it's big brother it'll bury the 2600k at almost £100 cheaper.

Overall It's going to be close enough that whichever way you go you won't be disappointed. I suspect that you'll get the upgrade itch long before either is seriously taxed.

manveruppd · 29 Sep 2011 at 00:22

Calling it SandyB makes it sound like a character from Grease or something...

SheepBeast · 29 Sep 2011 at 01:26

post #5145 should be required reading before posting in this thread. so many people here don't seem to understand anything about how cpu architecture actually works, yet still draw baseless conclusions about BD performance. thankyou dm.

Final8y · 29 Sep 2011 at 06:50

AMD Confirms 32nm Yield Issues at Global Foundries

AMD just announced revised revenue projections for Q3. Revenue is up compared to Q2 by 4 - 6%, but AMD had originally expected an increase of 10%. The reason for the revised projections? Llano supply is limited by apparently poor yields on Global Foundries' 32nm process. We had heard rumors to this effect for a while, but now they're officially confirmed by AMD.

The official statement is below:

The less-than-forecasted preliminary third quarter 2011 revenue results are primarily due to 32 nanometer (nm) yield, ramp and manufacturing issues at GLOBALFOUNDRIES in its Dresden, Germany factory that limited supply of "Llano". Additionally, 45nm supply was less than expected due to complexities related to the use of common tools across both technology nodes. AMD continues to work closely with its key partner GLOBALFOUNDRIES to improve 32nm yield performance in order to satisfy strong demand for AMD products.

The bigger concern in the near term is the impact this will have on the ramp of Bulldozer. Llano wasn't a huge chip, Bulldozer is.

Source: AMD

http://www.anandtech.com/show/4894/amd-confirms-32nm-yield-issues-at-global-foundries

gareth170 · 29 Sep 2011 at 07:46

does this mean BD will be limited in supply?

stulid · 29 Sep 2011 at 08:04

I vote for a change of thread title, its been "coming soon" for ages.

Andy__C · 29 Sep 2011 at 08:22

That has to be a major contributor in the delay....

And the revenue growth being 50% of its target - you can see maybe why heads had to roll in the AMD mgmnt.

drunkenmaster · 29 Sep 2011 at 08:30

gareth170 said:
does this mean BD will be limited in supply?

Potentially, if a company can't provide you with enough chips theres not a huge amount of options. But realistically we're still talking about millions of chips getting made a year, and this effects AMD's bottom line and the amount HP orders by the 100k's. When it comes to distributors getting 5k in and OCUK ordering 100 at a time it shouldn't be too much of a problem.

We'll have to see, supply isn't hugely down on what they expected, its a 4-6 % rise over 10% +/-2%, so it could have been as low as a 8% rise over Q2, and this is now only 6%. Part of the problem is 45nm is dropping as 32nm increases so the problem isn't all 32nm either(though it is the cause).

Bulldozer is certainly bigger, and theres been info coming out about Glofo/32nm problems for some months now. Yields are increasing afaik, they biggest problem is they aren't increasing fast enough.

Main problem for consumer is server versions of the chip sell for 10 times the cost, so if you have limited production, cash wise for AMD its FAR more important to get server chips out over desktop chips

Ages ago it became pretty clear, Bulldozer taped out almost 2 years ago, if 32nm was working as it should be, it would have launched in Q2, the respins are almost certainly down to making adjustments for the process/yield issues. Its a shame, but don't forget Intel is having a hell of a time with 22nm, and 32-28nm gate first is a lot smaller than Intel's 32nm, the 28nm gate first will be closer to 22nm from Intel. I'm also guessing thats where a lot of problems are coming from, the equipment and testing, production its geared for both 32 and 28nm, its probably the single most complex process to date and they've added complexity in letting the equipment essentially be scaleable.

Hopefully a long a crap 32nm ramp will help iron out issues and the 28nm transition will be problem free and VERY fast compared to starting a new process with new equipment.

TSMC and glofo are having issues with 28nm bulk, Intel with 22nm, everyones having pretty major issues, billions and billions were wasted on 32nm bulk that never even appeared. Processes are getting highly problematic now and honestly we're going to have these issues with almost every new process. Intel don't get effected as they put one fab to the new process, the ramp/getting it right might take 2 years instead of 6 months it used to, and time between processes might increase but they still have 3-4 other fabs in max production of the old process. AMD don't have that luxury with only one real fab right now, late 2012 that problem will be dramatically less bad. TSMC have an issue as only way to double gpu power and keep power sane is new processes, they have to have something new every 2 years or the gpu guys have major major problems.

EDIT:- having just read a story by Charlie, apparently, Llano's yields are pretty abysmal, 5-15% for fully working Llano's, and 60% for all the chips together, a8/a6/a4, so thats a lot of salvaged parts. what you'd want is, 40-50% A8, and the salvaged parts taking you to 70-80% yields, because A8's are far more expensive you're basically making 3 times as much money. Its not all bad, AMD made a "per good die" deal with Glofo a while ago which is awesome. Normally you pay $5-10k for a wafer, and you get what you get off it, if yields suck you lose money. AMD are only paying for working chips, meaning Glofo are eating the cost, AMD could be losing money this quarter rather than making less profit. The only issue is, if theres demand for 10 mil llanos a quarter, and you only get 5, you're losing out in well, you're starting to talk in fractions of billions rather than millions of lost profit.

Anyway the interesting bit, Bulldozer DOESN'T have as bad yields despite being 50% bigger than Llano, that is epic news frankly, the strange news, Trinity doesn't have anywhere near as bad yield issues as Llano? Bulldozer, no gpu, Llano, huge gpu and awful yields, with rumours its all the gpu side(but apparently false), Trinity, huge gpu and much better yields, so what the heck is wrong with Llano?