AMD Bulldozer server info

[informal]

Im a bit less concerned about dual socket as I am about AMD having a product that will match s2011 sandy bridge. Maybe 8 cores wont keep up, but take a 12 core g34 bulldozer opteron, laser off a hyperT link so its not dual socket capable, unlock its multi and clock it over 3.5ghz then call it fx.
This creates three scenarios
1- It would make mobo makers create oc-rich single g34 mobos (and eventually dual socket that could take opterons and oc them too)
2- It would create a differenciated high perf bulldozer with quad ch. ram that would possibly match anything on s2011 by intel
3- It would not interfere with the server segment since the fx it would not be available in dual socket (hence lasering the HT link)

this is one way to make enthusiast consumers happy and to follow amds policy of staying out of dual socket desktops.

PS: I will further extend my marketing 2 cents by even giving some guesswork figures and a shiny naming scheme (tell me if it makes you feel like buying it)

I like the name that Ive seen floating around for bullzoer to be VISION, then we can simply add the number of cores as a part of the name whch will also work to distinguish the the segment:

VISION 2 (AM3+)- 1 module, 4mb cache, dual ch ddr3. up to 4ghz - 60w
VISION 4 (AM3+)- 2 module, 8mb cache, dual ch ddr3, up to 3.8ghz -90w
VISION 8 (AM3+)- 4 module, 16mb cache, dual ch ddr3, up to 3.5 ghz -105w
VISION GX (??)- 2/4 module, 4/8mb (L2), dual ch ddr3, 3.5 ghz with dx 11 graphics -105/125w
VISION FX (G34)- 6 module dual die, 24 mb cache, quad ch ddr3, up to 3.6 ghz -125w


how do you like my AMD bulldozer lineup?

I highly doubt there will be 6 module(12C) 32nm desktop chip on AM3+/whetever.G34 for client space is a pipe dread...

BUT on the other side,we had a slip of the tongue(now "rectified") in AMD code name table just after the Nov Analyst day.Code name Komodo ,up to 10 enhanced Bulldozer cores,intro date :2012.
So basically they let it slip that Zambezi 2(or ++) will be based on 5module(10C) enhanced BD design and probably be AM3+ compatible. 10 cores of BD++ should be enough to match SB/IB 8C . Heck,I expect Zambezi to be competitive with 8C SB in most real world client tasks while it will trail it in most synthetic and those that may be heavily AVX optimized(but the AVX speedup is under big question mark looking at latest RWT forum discussions-according to Eric Bron,author of Kribibench ,expect ~20-25% speedup from real world optimized AVX 256b apps). Also AVX in intel CPUs in its present form lacks 256b integer instruction support while Zambezi will have that in the form of AMD's advanced XOP extension proposal,so if devs optimize multimedia apps for XOP 256bit AMD may end with an big advantage in such cases. Note that in BD,integer SIMD ops are not done by integer core's ALU units but by the FPU (coprocessor). The thing is that 256b (AVX) floating point instructions are useless for games and media applications..

[Dimitriman]

I highly doubt there will be 6 module(12C) 32nm desktop chip on AM3+/whetever.G34 for client space is a pipe dread...

I meant dual 3 module (12 core which will be on G34 from day one), not 6 modules on one single die, even I know thats insane

but I highly doubt a single die 4 module bd will match an 8 core sb on just about anything at all though...

[informal]

but I highly doubt a single die 4 module bd will match an 8 core thread sb on just about anything at all though...

We are talking about client workloads here . I won't be surprised to see BD matching SB 8C in such situations.If it did that in server FP intensive workloads that rely on AVX I would be surprised though.
On paper,one BD integer core is down just one ALU when compared to SB(Load/store capability is roughly the same).But it has a SIMD integer port(x2) in its Flex FP so it's a wash .Clocks on the other hand are a big unknown but there are rumors that 8C 95W parts are 3.5+Ghz. If this is true then 4Ghz 8C 125W parts are not out of the question.And this is stock,Turbo is on top .

[cky2k6]

We are talking about client workloads here . I won't be surprised to see BD matching SB 8C in such situations.If it did that in server FP intensive workloads that rely on AVX I would be surprised though.
On paper,one BD integer core is down just one ALU when compared to SB(Load/store capability is roughly the same).But it has a SIMD integer port(x2) in its Flex FP so it's a wash .Clocks on the other hand are a big unknown but there are rumors that 8C 95W parts are 3.5+Ghz. If this is true then 4Ghz 8C 125W parts are not out of the question.And this is stock,Turbo is on top .

"Client workloads" for those who want these chips consist of 3 things: benching, folding or video. Gamers obviously have no real interest, outside of the trust fund kiddies who buy alienwarez with the most expensive cpu possible. All 3 of these sitauations are very well threaded for the most part, and AMD has the firepower to even beat Intel's 8 core 16 thread chip if it makes a g34 FX with a dual 4 module chip, or will match it with a dual 3 module chip. It will do poorly with a 4 module chip. And JF, before you come out with your stock "but it takes 5 mil" response, for the love of fsm, look at 1366 sales...