Thread: Bulldozer...

my motherboard got fried somehow a few months ago and it would have been compatible with a bios upgrade but i just decided to get a newer chipset that has am3+ on it the 990x to be exact and i mean it didnt brake the bank at all so its not a big deal however i heard today that bulldozer will be officially ready for retail on sept. 19th and that the 7000 series of gpus will be ready in september as well idk if its true or not but i hope so

Well thought I would weigh in my understanding of the new architecture.

Let me preface by saying I have not seen any benchmarks or performance estimates yet, and am basing this completely off of my understanding of software design and instruction set optimizations.....

So let me start by quickly saying how intel's hyper-threading works: So Intel decided a while back that running one instruction per clock on a core was wasteful when a major part of the core isn't being used. For example.... you hypothetically can run one integer operation (say add) at the same time you are doing a multiply, because they use different circuitry. (This might not actually be possible as it is two integer operations... but you get the idea) So Intel put the extra resources there to figure out what operations can be run in parallel and called that a hyper-threading core. It works great when the workload is very diverse and is running a ton of different operations. These extra operations can be handled concurrently and provides a boost almost as if there was a second core. Now this is at the very basic level, and I am sure there are all kinds of tricks being done to make the most of this feature.... just think how many instructions there are in an instruction set and how many instruction sets these modern processors support..... that's a lot of possibility for concurrent hyper-threading and that's what makes them so beneficial to have.

Now..... AMD's new architecture, from what I have seen they are taking a different approach. They looked at all the architectures and said....
'HEY look at this.... integer operations are much much much faster than floating point operations'
and they probably also said.....
'HEY look at this.... because they are so much faster all these instructions in the instruction sets use integer instructions much more commonly than floating points'
Now I could go super technical here and explain how instructions are actually always broken down to smaller stuff till you get to the basic instructions.... but take my word for it.....

(EDIT: ok on a MIPS architecture ADD isn't simply add.... if you look at the underlying hardware it is handling overflow from the registers and making sure that you don't have a negative result from an overflow in the registers.... ok so say we look at something more advanced:

if we look at a BGE (branch if greater than or equal) (bge $rs,$rt,Label) that actually gets translated to:
SLT (set on less than) (slt $at,$rs,$rt)
BEQ (branch on equal) (beq $at,$zero,Label)
Now keep in mind this is just a simple example still, but shows how a single instruction can be translated to two instructions..... (MIPS is very simple compared to intel / amd complexity)
)

So AMD decides, instead of putting a whole nother core right beside the first core... why don't we just duplicate the parts that get used most often. So that's what they did, but they didn't stop there. They made it possible to combine the two cores to execute a higher precision operation, if it was needed....

So in summary while I wouldn't say the new bulldozer chips coming out aren't truly 8 cores by what we are used to.... I would say they are 4 cores with .5x4 extra cores..... so 6 cores? Maybe thats not a good way to explain it. I would say that there are 8 possible integer cores, and there are 4 possible floating point cores. But each of the floating point cores takes a sacrifice of 2 integer cores until the operation is complete.

What I am expecting to see as far as performance: whenever the benches come out is that graphical/encoding/decoding/encryption(64 bit precision)/and most other operations that are highly optimized to utilize faster integer operations are going to see full 8 core operation.

But when you do more floating point operations, or highly accurate (scientific) operations, or intense encryption operations you will notice cores being 'cut' in half as they have to lose the double integer cores to do the floating point operation. How I think this will be represented in task manager for example is when a floating point operation is running on modual 0 will be core 0 and 1 will show the same utilization..... (cause they are the same core).....



For the future this architecture is going to win, I can say just from the way its designed. For instance, say in the future somehow integer operations become less popular and floating points are needed more, we can do cores that are opposite (2 floating point, 1 integer)... perhaps even some dies (processors) that mix 2int1float's with 2float1int's to have equal numbers.... and who says we have to stop at 2.... how about 3 integer cores to every floating point.... Its really a balance that has to be represented by what the software of the day is executing as far as instructions.


I am excited for the performance reviews...... can you tell?

So it seems that just the other day (6th or 7th) the server variant of bulldozer started shipping. But today brings sad news in that supposedly the desktop variant is delayed to october, but they are going to release with more SKU's....

I was all prepared to sell a bunch of hardware to try to get one of them.... now they are giving me another month to loose more value on these parts, lol.....

rush, great write up! I never looked into how bulldozer was going to work. I actually just assumed they were avoiding the efficiency battle and now just trying to cram as many physical cores as possible onto a single die. I am really excited to see benchmarks myself.

Thanks for noticing. It took some time to write that up (and somewhat to show I know what I am talking about), glad someone appreciated it.

If anyone wants an article / official blog post I could clean it up a bit and make it more formal and less like a conversation piece.