hcl64 (mario.smarq@gmail.com) on 5/1/12 wrote:
---------------------------
>David Kanter (dkanter@realworldtech.com) on 4/30/12 wrote:
>---------------------------
>>hcl64 (mario.smarq@gmail.com) on 4/30/12 wrote:
>>---------------------------
>>
>>>IF i'm not mistaken the majority of XOP is Integer (3 >operand) total pack has more
>>>than one hundredth instructions defined, i think.
>>>
>>>Perhaps integer fusing could be also a good thing.
>>
>>Doubtful. The latency would also kill you.
>>
>>David
>
>Perhaps there are ways to mitigate the latency upon a kind of decoupled access-execute
>which BD with its so called co-processor oriented Flex ?>Units seems to trail, and instruction fusing...

Latency kills performance for CPUs, in a major way. Higher latency means your instruction window is effectively smaller, forwarding is more challenging, etc.

There really isn't any good way to hide the latency. Just think about code where you are branching based on the result of an FP computation.

Moreover, the idea of offloading from an integer core to a GPU is especially silly when you consider the frequency difference. Bulldozer can run at around 3-4GHz. AMD's integrated GPUs run at around 400MHz or so. The clock domain crossing along would add a ridiculous amount of latency.

The Bulldozer FPU uses the load/store pipes from the integer cores, because it fully obeys x86 semantics. That's not the case for GPUs. There's really no reason you'd want to share the LSU between GPUs and CPUs. The address calculation is totally different, the TLBs are different, the memory ordering model is totally different, etc.

You cannot replace the x86 FPU with a GPU. x86 needs x87 FP, which isn't available in any GPU. It also needs lower latency than is available through a GPU and support for all MMX, SSE, AVX instructions. With branch prediction, the FPU also has to be able to rollback to an earlier state. GPUs don't really have any facilities for speculative execution, since branches are either predicated or halt execution (while the GPU core switches to another thread). GPUs simply don't have all these features, nor would it be wise to add some of them.

So the CPU core still needs a regular FPU. At that point, you are better off with the GPU as a separate block. It runs at a different clock, with different voltage, with it's own load/texture pipelines, power management and so forth.

i.e. it's really just another part of the SoC...not a coprocessor really.


[snip]

Are you just copying and pasting from a patent? What was that text?

David