I have 2 ASUS M4A89GTD PRO/USB3 that are AM3 socket that I bought last February. There is a beta BIOS (3027) for these boards that adds support for BD. the socket is definitely AM3 and not AM3+. I'm not going to upgrade (downgrade?) to BD as I already have 1090t in both and see no need to change atm. if they improve BD within the next year, then maybe but I don't think that will happen.
My understanding is that an AM3+ chip will physically fit in a AM3 socket, it's up to the vendor to add support for it as AMD will not officially support that option.
So basically you want them to make 2 full cores again... waste alot of die space in the process and call it da day.
(1) Beg for MS to patch Windows 7 ASAP with scheduling optimizations for Bulldozer, this alone could be around 10% gain.
Two of the improved Stars core that are in Llano, the one's that get 3-6% better IPC than the cores already in Thuban, are roughly the same size as one bulldozer module when you include the L2$.
One can make the argument they could swap out 4 bulldozer modules in exchange for 8 Llano cores and Zambezi would remain nearly the same diesize.
1 BD Module + 2MB L2$ = 30.9 mm^2
2 Llano Cores + 2MB L2$ = 35.4 mm^2
The problem with this analysis is that for whatever reasons those Llano cores suck in terms of clockspeed and power-consumption. A mere 2.9GHz and the quadcore gulps down the juice when you do CPU intensive stuff.
So what would an 8-core Llano look like in terms of clockspeed and power-consumption? It would not have been pretty. Something is not right with GloFo's process and I think it shows in both Llano and Bulldozer.
Edit: I'm not sure if they should avoided CMT. I think it is a very good approach in theory. I just don't think Bulldozer is the best execution of the CMT approach atm.. but i think the blame here is the process, the other design choices within BD and not the CMT prinicple. (which actually works very well within BD.. its the only thing that was said by AMD that was on the mark.).
Also am I the only one who doesn't expect alot from sheduler changes? The point AMD wants is that as many threads as possible are crammed in one Module. This will give higher % turbocore than when threads are hopped of other modules. But this also has an impact on the performance due to the mandatory sharing. My guess is that besides a better power consumption we will see performance increases and decreases over the field.
In the year running up to actual Bulldozer launch a lot of people were asking about Windows 7 scheduling. I believe the general AMD attitude was that it didn't matter whether it loaded modules or spread the load since you would get extra clockspeed from loading full modules versus not sharing resources by using threads on different modules. It's as if the software guys at AMD don't really communicate that well with the hardware, because Windows scheduler throws threads around like a bit of a madman since that is what Intel's HT does alright with. It also happens to be the worst way for an OS to treat threads for AMD's CMT design...
Fixed that for all the intel fanboys and investors here...
Is Intel's virtualization really broken in Sandy Bridge?
Well it doesn't help, certainly. But AMD's preference for sharing threads on the same module, assuming low # of threads ofc, isn't a good idea in terms of performance, either -- perhaps for power consumption -- considering the 1-thread-per-module approach seems to be yielding the better results. Consequently the CMT design has to be be scrutinized even more given the disparity between "within module" and "per module" numbers are so fall apart.
The problem with this analysis is that for whatever reasons those Llano cores suck in terms of clockspeed and power-consumption. A mere 2.9GHz and the quadcore gulps down the juice when you do CPU intensive stuff.
AMD should have avoided the whole CMT thing, reduced the core count from 8 to 6 but made them real cores.
Of course we are talking about people who couldn't count transistors correctly, nor realize in advance that their microarchitecture was going to suck if they didn't get MS to tune their scheduler before the chip was launched.
So we can't really hold their feet to the fire, its obvious they put their A-team on Brazos, their B-team on Llano, and the leftovers got shoved onto operation derpdozer.
My point was that you would have to work at it to make a worse scheduler for bulldozer than the one Windows uses now. Not sure why everyone is on this whole "hey look if you run a single thread per module you don't pay the cluster penalty", of course it works that way. The clustering part of the design appears to be working well for the first shipping silicon. It's just AMD obviously missed their thermal+clock targets. (Edit: and possibly their IPC targets, but AMD is being a bit coy regarding that)
In my view there are two main WTFs regarding Bulldozer:
1. Their clock targets seem to be just as aggressive as Intel's for Pentium 4.
2. Working on gate first 32nm and I would guess 28nm seems to be the fab equivalent of fighting during a Russian winter. IBM should be sending Global Foundries chocolates and flowers for all the trailblazing they are doing.
So basically you want them to make 2 full cores again... waste alot of die space in the process and call it da day.
Reducing L2 to 256 is bad. 1MB sounds better. Reduce latency of this cache and implement a L0 cache like structure or other means to bypass branch hits. (think this is already in the pipeline for steamroller).
Decoding width is fine, in combination with some enhancements like SB should do the trick.
Bumping exeuction resources might be an option, although they would be far better of by making their AGU do more basic calculations.
double the fpu is pure nonsense, affecting execution times would have more effect...
In fact, they originally planned a 30% increase in clock speed, which would have put them at the mid 4ghz mark.
The web is saying there will be a 300MHz higher clocked part called a FX-8170. The single core Turbo frequency is at 4.5GHz, while multi core one is at 4.2GHz. That's pretty close.
AMD could do something to save the sinking ship:
(1) Beg for MS to patch Windows 7 ASAP with scheduling optimizations for Bulldozer, this alone could be around 10% gain.
(2) Work on a new stepping (if work hasn't already started) to make small improvements in IPC and/or thermals.
(3) Lower prices of current SKUs, as they are now, they make absolutely no sense.
Lower prices + Windows optimization + small improvements in IPC = we have a decent alternative to SB, especially if priced low.
Bulldozer isn't a server oriented CPU, SpecInt and FP Interlagos outperforms the Xeon X5690 by 14% and 30% respectively.
The 6 core Core i7 3960X, gets 34% and 65% better in the same metric over Core i7 990X, which is essentially the same chip as X5690. Then there's the 8 core Xeon E5's, which should add 10-15% on top of that because of 33% more cores but lower clocks.
What doesn't work in one area doesn't work anywhere else.
Yeah but we also know that a large percentage of, if not nearly all llanos can be undervolted to 1.1~1.15V and still run at 3GHz. At those voltages 8 cores could be had under 120W. It wouldnt be great, but it would be faster than BD. And cheap. Very very cheap.
Its not like some dude at AMD pressed the space bar twice instead of once by accident and because of that mistake all the Llano have excess Vcc.
The yet to appear 8170, if they manage to deliver it at those specs and 125W and ~$250 then it will take some of the sharp edges off of Bulldozer.
The also as yet unseen unlocked Llanos are apparently starting to show up on distributor lists. Perhaps they will actually be purchasable soon. I'm interested in seeing what people can get out of them.