Bulldozer (V1 and V2) - The (partial) Scoop

[4P_Bulldozer]

Public Demo of Working System. Ref: http://www.xbitlabs.com/news/cpu/di...s_Off_Working_16_Core_Opteron_Interlagos.html

 

[PP Mguire]

Perhaps you liked the Phenom 2 better (@ 7.02GHz) (Featuring Chew*):

I was there and watched them do this. A few days later (while still at Quakecon) the 06 record was beat again and they couldn't top it. I lold.

 

[Wildside]

The bulldozer chips may be secret for now in the clocks, but i think i will wait for the chips in the summer. Those ones will have AMD's own "hyper-threading" tech implemented.

 

[Apokalipse]

Leaked specs suggest the FX-8150 runs at 3.6GHz, with a turbo of 4.2GHz

Anyway, BD modules don't work the same way HyperThreading does, but the goal is the same - a small increase in die area giving a large increase in multithreaded performance
You could compare 1 BD module with 1 hyperthreaded core that way.

AMD says one BD module gets 80% more multithreaded performance than single threaded performance, while having 12% larger die area than a single core, 256-bit FPU and all (this is not compared to Deneb/Thuban, but hypothetical single BD core)

In comparision, Hyperthreading gets about 20% more multithreaded performance than single threaded performance, and has 5% larger die area than a core designed without hyperthreading.

 

[Puddle Jumper]

If I had to guess, I'd say it's IPC will be somewhere near SB's IPC.
Will it be higher? lower? I don't know.

I wouldn't hold your breath for IPC matching Sandy Bridge, the gap between it an K10.5 is massive and I don't see AMD closing it in a single generation. Currently it's like the netburst days in terms of the gap in IPC except this time Intel is the one on top.

It is also worth keeping in mind that Ivy Bridge isn't that far away and Intel even had early samples of Haswell back at IDF.

 

[Apokalipse]

There's an incredible amount of work gone into BD; architecturally, it looks like it will be at least as fast as Nehalem in single threaded IPC (probably better)
So yes, I think it will be in the ballpark of SB, though it may well be lower.

I think AMD's BD modules are better than HyperThreading - it's a lot closer to two full cores, but without the die area
In particular, I like the FlexFP design; a 256-bit FPU that can process two 128-bit FP instructions simultaneously from two threads - Almost all FP instructions are 128-bit or less still, with 256-bit AVX instructions only just being made. So making 8 full 256-bit FPU's wouldn't really make much sense. It would be a massive increase in die size with a tiny increase in performance. You want to do the opposite.
Sharing a 256-bit FPU like BD's FlexFP means those execution resources are actually used a lot more, and also means it can still process the (small number of) 256-bit AVX instructions in one cycle instead of decoding it into 2x128-bit micro-ops and processing one after the other.

Now, if IPC isn't higher than SB, how will it compete single threaded?
Actually, it looks like AMD is designing it to clock very high, with a streamlined pipeline.

But wait a minute, didn't Intel try that with Netburst? How is this different?
With Netburst, Intel pretty much ignored IPC altogether, and tried to rely on imaginary improvements in manufacturing processes.

AMD is streamlining BD's pipeline for higher frequency, but they're not making it their sole focus. It's just part of the picture.

IPC isn't the whole picture. Frequency isn't the whole picture. Core/thread count isn't the whole picture. You need a good balance of these things.
That's basically what AMD is trying to do with BD. Not focus on one specific thing, but improve as many areas as they can without over-engineering them.

I'd say they're definitely ahead of Intel in some ways. Maybe not every way, but we'll have to see some results to know exactly how successful they've been.

 

[PP Mguire]

The 8 core dozer chip according to the released AMD grain of salt slides puts it on par with the 2600k in multi-threaded apps and in class with the 980x in games. It makes me assume that SB eats it alive in single threaded apps meaning core to core AMD still can't compete. This is just an assumption based on press slides released by AMD. Usually in these the product at hand will be "better" than the competition, but in this case they put dozer marginally better/worse than the 2600k/980x so that makes me assume for the price point that performance wont be all that we hoped. I would be willing to bet slightly better than X6 in single thread and marginally better at multi-thread depending on clock speed.

Lets not forget that Sandy-E will be released soon so it's a safe bet that Intel will still hold the performance crown with the enthusiast market holding dear to that quad channel goodness.

 

[gurusan]

From what I've seen it's less than impressive, and those are AMD's own benchmarks. The price, however, is pretty darned good. Not good enough to compete with the already old Sandy Bridge though.

Also SB-E is stupid expensive and quite frankly foolish to get for a gamer rig...it's not really in the same price bracket as Bulldozer.

 

[PP Mguire]

In all fairness neither i7 enthusiast platform has been in the same price bracket as AMD because they simply can't compete with Intel's enthusiast line. It's stupid expensive, but people were paying 1000+ for FX chips back in the day too.

I wouldn't call Sandy Bridge old, considering it's not even a year old yet and Intel nor AMD has released anything new prior to the release of SB. It's also still the top dog, and looks like it will be till its bigger brother is released.

 

[gurusan]

I meant old as in we are getting closer to Ivy Bridge early next year.