Dresdenboys' blog: AMD Bulldozer - Patent based research part 2

[Behemot]

However, where did you get the 480 Shaders from? It would make it slighty superior to the Radeon 5670 that got 400.

Fusion parts are good for either the Notebook market, and for the low end and mainstream Desktop (That is also where AMD is aiming them at), because getting more things integrated in the Processor piece is cheaper than having them separate.
The problem isn't the new Socket for Notebooks, but the fact that we will have again two Desktop Sockets coexist. And the idea of segmentating the market in both Fusion for low end and mainstream, then Bulldozer for high end desktop, doesn't sound very good if you want the features of both platforms. No one would like buying a Llano then hearing news that Bulldozer will come first to Socket AM3.

It's from some picture of the processor. The expected power of IGPU is the most interesting thing of all - it's more than enough for all low-end to mainstream applications (office, HTPC, light gaming). You can still use more powerfull dedicated VGA.

I see no problem in features - as long as we know Llano does not lack anything compared to GPU-less processors, so again no difficulties with features. Intel now has 3 active sockets (775, 1156, 1366) and plans to have 4 (1155, 1156, 1366, 1365 or something like that) - it is crazy but apparently sells well and that's what matters.

[-Boris-]

How about Bulldozer v1 in both AM3 and another socket which Llano uses.

Desktop will go from AM3 to "?" in the future, you get to choose either >2 channel memory/PCIe or pins for GPU usage.


Laptop for Llano will probably be S1g5/FS1.

That is what I believe. Would be nice with something like 1156 with 32-40 PCIe lanes.

[Oliverda]

My thoughts:
Bulldozer having DDR-2 / DDR-3 IMC: Socket AM3 (938 pins)
Bulldozer having DDR-3 IMC: Socket AM3 (941 pins)

Llano is a quad-core design with 480 shaders.
Brazos is a dual-core design.

IMHO 941 pins AM3 CPU will be called as AM3+.

Maybe AM3+ mobos will support separate power planes for each modules in BD.

[FlanK3r]

what about DDR2 IMC? No, please no...DDR2 is in future death. Yes, i have now DDR2 mobo with 965 BE, but for future not, not not. Or, do u can beat DDR3 2100+ in benhcmarks with ddr2 1066 ?

[-Boris-]

what about DDR2 IMC? No, please no...DDR2 is in future death. Yes, i have now DDR2 mobo with 965 BE, but for future not, not not. Or, do u can beat DDR3 2100+ in benhcmarks with ddr2 1066 ?

Depends, with the IMC from Phenom II there isn't much difference.

[Behemot]

Depends, with the IMC from Phenom II there isn't much difference.

Exactly, due to HyperTransport (FSB no longer is bottleneck) the memory does not really matter. The results are almost the same because of the higher the bandwith, the higher the latency. Overall performance in real world apps still the same no matter if you have DDR2-533 or DDR2-1200

Compared to GDDR, the DDR chips stagnate badly

[-Boris-]

Exactly, due to HyperTransport (FSB no longer is bottleneck) the memory does not really matter. The results are almost the same because of the higher the bandwith, the higher the latency. Overall performance in real world apps still the same no matter if you have DDR2-533 or DDR2-1200

Compared to GDDR, the DDR chips stagnate badly

Don't blame the memory chips, the problem is clearly an AMD only issue.
While Phenom II have seen very small advantages of DDR3, Intels LGA 1156 processors have much better bandwidth.

http://www.anandtech.com/show/2832/7

In many cases Athlon 64+ 6400+ shows better memory bandwidth than Phenoms, due to full speed IMC.
Overclocking the IMC cures the problem, but 3.5GHz+ IMC isn't really realistic.


A bit off topic, but just want to show how severe the IMC problem really is in the Phenoms.

http://www.madshrimps.be/?action=get...66&articID=909
66% higher IMC frequency gives 54% higher write performance. With DDR2 800MHz! DDR 400MHz wouldn't be far away in write performance!
Clearly bottlenecked by the IMC. At least read perfomance isn't crippled as much as write.

http://www.madshrimps.be/?action=get...65&articID=909
Overclocking memory gives much less perfomance increase in Nature test than overclocking the IMC.
However, read performance increases, hinting that it's the write operations that the IMC bottlenecks mostly, at least with DDR2 800, with faster memory the IMC should bottleneck read operations even more.

http://www.madshrimps.be/?action=get...67&articID=909
The Phenom II starts getting memory starved already after 1400MHz.


I would love to see these tests with a 3.6GHz Phenom II with at least 1600MHz memory.
And coupled with some real world game benches.

[blindbox]

Don't blame the memory chips, the problem is clearly an AMD only issue.
While Phenom II have seen very small advantages of DDR3, Intels LGA 1156 processors have much better bandwidth.

http://www.anandtech.com/show/2832/7

In many cases Athlon 64+ 6400+ shows better memory bandwidth than Phenoms, due to full speed IMC.
Overclocking the IMC cures the problem, but 3.5GHz+ IMC isn't really realistic.


A bit off topic, but just want to show how severe the IMC problem really is in the Phenoms.

http://www.madshrimps.be/?action=get...66&articID=909
66% higher IMC frequency gives 54% higher write performance. With DDR2 800MHz! DDR 400MHz wouldn't be far away in write performance!
Clearly bottlenecked by the IMC. At least read perfomance isn't crippled as much as write.

Phenom IIs run at 2000+ MHZ IMC at default, at least for the AM3 versions.

http://www.madshrimps.be/?action=get...65&articID=909
Overclocking memory gives much less perfomance increase in Nature test than overclocking the IMC.
However, read performance increases, hinting that it's the write operations that the IMC bottlenecks mostly, at least with DDR2 800, with faster memory the IMC should bottleneck read operations even more.

Well, we are currently using DDR3 for Phenom II, would it be a problem? Bulldozers are DDR3 only AFAIK, only for AM3 mobos.
If you interpolate and make 1066 MHz the basis, anything higher will be the usual 1%-2% improvement from better memory.

Just because it doesn't scale with memory speed doesn't mean it sucks at memory. It simply doesn't need that extra bandwidth.

http://www.madshrimps.be/?action=get...67&articID=909
The Phenom II starts getting memory starved already after 1400MHz.

I think this one explains memory starved better, as it checks scaling at IMC constant 1800 MHZ with memory clocks.
http://www.madshrimps.be/?action=get...65&articID=909

The conclusion: You're not gonna see anymore extra scaling after you overclock memory and IMC beyond.

I would love to see these tests with a 3.6GHz Phenom II with at least 1600MHz memory.
And coupled with some real world game benches.

You can probably see them at Phenom II 965, 975 reviews.

I don't get your post. What problems does Phenom II has with its IMC, apart from unable to run 4 DIMMS without extra volts?

But hey, that gives them a reason for integrating the graphics core : Using up all that extra bandwidth.

EDIT: Oh crap, still not used to firefox 3.6 tabs.

[Behemot]

...

Come on. You play memory write tests on yor computer? Really? If not, why do you think I said in real world apps?

As for Core iX, again, that's what I said: FSB was bottleneck (too thin for both the memory and everyting others - northbridge etc.), now with HyperTransport (or QuickPath) it matters no more. OCed processor has almost the same performance with DDR2-533 and DDR2-1200. The few % performance gain vs. over 100 % frequency gain only proves that

Of course there are some bandwith-hungry apps, but for average use, in 98 % occasions the memory speed has no efect.

[JF-AMD]

what about DDR2 IMC? No, please no...DDR2 is in future death. Yes, i have now DDR2 mobo with 965 BE, but for future not, not not. Or, do u can beat DDR3 2100+ in benhcmarks with ddr2 1066 ?

BD will not support DDR2.

And, believe me, in 2011 nobody is going to want to pay the price for DDR2 because it will be more expensive than DDR3 at that point.

[-Boris-]

The point is, Phenom II do much worse in ram benches than 1156 processors, and at the same time we know that overclocking the IMC gives a nice boost. Of course L3 is responsible for some of it, but it's always nice to have a system without bottleneck. And if you get a multicore monster with bulldozer, Phenom-like ram performance would be a huge bottleneck.

[informal]

The point is, Phenom II do much worse in ram benches than 1156 processors, and at the same time we know that overclocking the IMC gives a nice boost. Of course L3 is responsible for some of it, but it's always nice to have a system without bottleneck. And if you get a multicore monster with bulldozer, Phenom-like ram performance would be a huge bottleneck.

Who told you that BD will have "Phenom-like" memory perf. ??
BTW Phenoms have good memory perf. ,the cores are just not that BW hungry,that's all.

[-Boris-]

Who told you that BD will have "Phenom-like" memory perf. ??
BTW Phenoms have good memory perf. ,the cores are just not that BW hungry,that's all.

I haven't been told that, I was just mentioning that it has to be better than current Phenoms, and that bulldozer for AM3 might be a bad sign, even if the IMC could be completely different, and therefore could be much better.

And if you would be right and it's the cores that isn't bandwidth hungry, it would be the cores that would bottleneck the performance. And if so, it would scale with amount of cores and core frequency. That is not the case. Bandwidth stops scaling with core frequency around 1400MHz. Only under 1400MHz can you say that the cores isn't hungry enough to utilize the IMC.

So, if not the cores are the limiting factor, what could it be?

[ajaidev]

BD will use AM3+, AM3 will be fallback see how x4 9950 does on a AM2 instead of AM2+.

[Sn0wm@n]

Correct, those are server only statements.

could you get a statement from the person responsible for the consumer platform... of course if it isnt under NDA

[AliG]

I don't think there will be an 8 module Bulldozer made on the 32 nm node, not as a single die at least. I expect an 4 module version to be the biggest version for consumers, at least initially.

I'm fairly certain that AMD said they would make 8 module chips, just as MCMs. It's only logical that they're top chips would become the FX line (I doubt we'll see a 12 module MCM, that would be a humongous chip; probably the next generation will come out before they have the manufacturing tech).

[demonkevy666]

phenom II has specif Core:NB ratio that works anything else will show it looking slow.

the limit of architectural speed of 1950mhz is being broken by the Thuban already showing 2,000mhz.

as with all AMD chips timings still work a bit better then higher mhz.

BD will have quad channel IMC see socket G34 was made for bulldozer's needs.

[Manicdan]

so what will motherboards be like with quad channel? 8 dimms or 4? given that by 2011 4GB per stick will become more popular, the next question is if it will even matter, cause right now DDR2 didnt like using all 4 dimms (something the the 1T vs 2T timings that reduces perf, not sure how much that affects DDR3). so when BD comes out, should people expect to run 4x2GB, or 4x4GB, or 8x2GB or 8x4GB, and which one will really matter? the average consumer will like to spend under 150$ on ram, so we need to have prices come down fast, or BD will be one expensive socket to buy

[informal]

MagnyC. boards we have today on the market wil be the boards for Interlagos MPUs(BD based ,2x4 modules,16 cores total). New BD Opties will be drop in compatible.

[JF-AMD]

could you get a statement from the person responsible for the consumer platform... of course if it isnt under NDA

Sorry, I try to stay out of the middle of the client business. The first time that I start to get involved I'll get innundated. Right now I can't even keep up with the server PM's and emails. I just don't have the bandwidth. They live in a different world where people sleep out to be first in line. The enterprise world is not like that so we can disclose more further out.

[JF-AMD]

so what will motherboards be like with quad channel? 8 dimms or 4? given that by 2011 4GB per stick will become more popular, the next question is if it will even matter, cause right now DDR2 didnt like using all 4 dimms (something the the 1T vs 2T timings that reduces perf, not sure how much that affects DDR3). so when BD comes out, should people expect to run 4x2GB, or 4x4GB, or 8x2GB or 8x4GB, and which one will really matter? the average consumer will like to spend under 150$ on ram, so we need to have prices come down fast, or BD will be one expensive socket to buy

4 channel server boards with be in 8 or 12 DIMM configs. If you look at what is available now, that will be an indicator of what most of the market will look like with bulldozer. There will be a few new interesting designs, but I can't talk about them.

[terrace215]

Speaking of BD, roughly when can we expect AMD to give a public "taskmaster demo" of it?

You know, something that confirms that A* silicon exists in-house and can at least boot Windows and show the 8 threads chugging away. Like AMD did for Barcelona back in 2006.

[informal]

Speaking of BD, roughly when can we expect AMD to give a public "taskmaster demo" of it?

You know, something that confirms that A* silicon exists in-house and can at least boot Windows and show the 8 threads chugging away. Like AMD did for Barcelona back in 2006.

They are already sampling Llano to some select partners and have never even demoed a system running a simple thing as task manager.What makes you think they will "show off" the current state of Bulldozer design,the one chip that will be the bread and butter when it comes to revenue and competing with intel for the next couple of years?
BTW we now have a confirmation that Ontario(Bobcat-the new uarchitecture as D.Meyer said it himself a few days ago) is also ready and being tested and this one is even done on a whole different node(40nm bulk;Llano and BD are 32nm SOI HP).

[Manicdan]

4 channel server boards with be in 8 or 12 DIMM configs. If you look at what is available now, that will be an indicator of what most of the market will look like with bulldozer. There will be a few new interesting designs, but I can't talk about them.

i was talking about the consumer boards, i know servers had 8 dimms for a LONG time now.

will consumer boards be 4 channel and 4 dimm?

[terrace215]

What makes you think they will "show off" the current state of Bulldozer design,the one chip that will be the bread and butter when it comes to revenue and competing with intel for the next couple of years?

Running a taskmaster demo gives away nothing to the competition, it just lets investors and interested parties know that schedules are not slipping.

Why do you think they "showed off" the current state of the Barcelona design at the end of 2006?

It was a bit disappointing that no analyst asked, nor did Dirk volunteer, anything about BD on the recent CC. Especially in context of (the admittedly much simpler) llano sampling, the silence around BD was unfortunate.