[Rumor] First performance figures Interlagos?

[Particle]

The problem arises in that "core" doesn't really describe the new architecture. We have to stretch the term around to make it fit. We'd be better off just calling these things modules.

A module has two integer units and one big FP unit that can either function as two smaller FP units or one big one. A module is very much two traditional "cores". FP performance is just variable depending on how the module is being used at a given moment. You could say the module is a dual integer + single floating point core or a dual integer + dual floating point core entirely depending on the instructions currently being executed.

[Mech0z]

It is 2 cores in one module, but each module shares some stuff (which saves a lot of space or cost) and it does so without a huge performance penalty per core (Think it was 90-95% performance of 2 individual cores) this is just what I remember reading on JFs blog though, might be kinda wrong.

[DarthShader]

Sounds too good to be true. Anything over 30% single thread boost would be a positive suprise to me. I still think there's some dodgy things about these results. Hope not though ofc. :p JF-AMD where are you? :p

You have to remember that this is only one example of application. (assuming all this is true ofc). This particular software could be a "best case scenario" and average IPC increases could be lower among a variety of workloads.

Similary like SB had a 25% FPS increase in Starcraft c2c, but on average it was more like 7-10% over last generation.

[xBanzai89]

1 module is ~1.8 cores, ONLY if IPC was identical between generations, which is not true.

I am confused what you are saying here. Do you mean that the IPC on single bulldozer core is slower then a single Phenom II core? If so that means a 2 module 4 core Bulldozer is slower then the current Phenom II X4. That would make no sense.

[informal]

I was under the impression from what AMD has said is that the OS system see's them as "cores", but weren't real cores other than they had an integer core...in addition AMD wanted to get away from the "core" terminology with the debut BD. So is that not true then?

Each core is a full core.Each integer unit has own FMAC. Nothing in BD module is half good,or makes 2 cores act as 1.5 core. You need to go back and read up on Bulldozer.

I am confused what you are saying here. Do you mean that the IPC on single bulldozer core is slower then a single Phenom II core? If so that means a 2 module 4 core Bulldozer is slower then the current Phenom II X4. That would make no sense.

No. Module has 1.8x the throughput of a hypothetical Bulldozer based dual core with non shared parts.Nothing to do with previous Family 10h or any other design from AMD.

[Manicdan]

I am confused what you are saying here. Do you mean that the IPC on single bulldozer core is slower then a single Phenom II core? If so that means a 2 module 4 core Bulldozer is slower then the current Phenom II X4. That would make no sense.

informal did a decent job explaining

basically if you have 2 threads on 2 modules, you get 2x perf (or like 1.95x in reality it seems)

with 2 threads on the same module, its 1.8x

so while there are 2 int cores per module, you will only 80% of the second one in terms of expected perf. although we have no idea how much this can vary or what cases it will be noticed.

i added the extra bit about only comparing between the same architecture people some people are worried that 1 module will be slower than 2 thuban/deneb cores, which is probably way, way off.