AMD FX-4100 Bulldozer

[V!NCENT]

Floating point is freaking part of the CPU. Given that there are two, it's dual core. Unless you have four cores and two different ones at that.

The term dual-core was invented for essentialy two CPU's being molten together on one die. In this case it's not any different. Unless it's a six core having four integer and two float cores. But they are not exactly entire seperate, so I'll simply calll this dual core.

[duby229]

By that definition, then a core on BD is not the same thing as a core on every other x86 architecture being used. On BD a so called core doesnt have a front end, or a FP unit, or a retirement stage, or a cache heirarchy.... No it just does not make any sense.

A module is a dual processor core. A module is not a dual core processor.. It may seem like semantics to some, but I think it is a very important differentiation.

[V!NCENT]

A Core is a single CPU. Period. If the design is so interwoven and shared that it can't be split up without evening out all modules; it's a single core.

So this AMD is a dual-core CPU.

Instructions might take more clock cycles to complete the logic operation, but it might be more efficient in that it can do more instructions per clock on avarage.

What we're discussing is nothing but two identical CPU's that have dual integer modules (I still have to read up about that, will do).

Given that most desktop stuff doesn't require insane amounts of floats per integers (less than 0.5); it's great and cheap. It's also great and cheap for home servers. Gaming not that much (if you buy the latest GPU's).

I personally don't like this path, because float is already slower than integer. AMD now cut down the difference even further. This sucks balls. Short term decisions.

[duby229]

I dont mind calling them modules. I just dont like the idea of calling an integer pipe a core. What about the FP pipe then? If an integer pipe is a core, shouldnt the FP pipe be considered a core too?

I think a core should be considered a complete functional unit, which is what a module is.

[AlbertP]

Let's call it a mix of a 4-core and an 8-core CPU. That way we have found a correct term for a lot of CPU's released in the last 5 years.

[bridgman]

???

No, I was saying that the big jumps in CPU complexity happened well before Athlon XP, ie out-of-order execution and superscalar design (which is what allowed everyone to talk about instructions-per-clock rather than clocks-per-instruction).

[duby229]

Please dont tell me that AMD is trying to move towards a "clocks per instruction" model.... That would be bad....

De-emphasizing ILP was bad enough.

I dont really like the term IPC, I think ILP is a more adequate term to use for what we are talking about.

[bridgman]

Yep, although one could argue that the big jump in complexity happened a few years before your Athlon XP came out.

My rule of thumb is that whenever you talk about "instructions per clock" rather than "clocks per instruction" you're talking about seriously complex designs

[AlbertP]

Yes, but a CPU of today is a very complicated thing. At least it is compared to the single-core Athlon XP in my desktop, which has already a lot of transistors.

[bridgman]

But the argument is about what constitutes a "core". If they agree on that then there won't be anything to argue about

The problem is that we're really arguing about rounding artifacts here. A hyperthreading core has some dedicated resources for each thread but sufficiently few that most people round it down to "one core". A CMT module has relatively more dedicated resources for each thread, enough that most people round it up to "two cores". Both simplifications are imperfect, and you really need to go down at least one more level of detail (to the individual execution units after the decoder) before everyone can agree on the terms.

The good news is that the thread is heading in that direction, but so far it's only talking about integer ALU execution units and not the other execution units (load/store etc..).