Bulldozer FPU performance

[dave343]

I thought, but can't remember where, I read the Bulldozers FPU performance is weak, part of the delay? and amd might be offloading some of the FPU work through OpenCL? Any truth to this?

 

[Tsumi]

The reason why some people consider the FPU performance to be weak is that there is only one FPU core for each two "cores", or for each module. Each module has 2 integer units and 1 FPU unit, whereas traditionally each core had one integer unit and one FPU unit. So relatively, based on those numbers, it will be weaker in a core per core count, since each module is considered as two cores. Whether it actually is remains to be determined when Bulldozer is actually released.

The FPU work through OpenCL as I understand it only applies to Fusion processors, ones that have a graphics chip in them. The upcoming AM3+ Bulldozers won't have this.

 

[CaptNumbNutz]

The FPU work through OpenCL as I understand it only applies to Fusion processors, ones that have a graphics chip in them. The upcoming AM3+ Bulldozers won't have this.

The bulldozer cpu's released next year will have the integrated graphics chips.

 

[Tsumi]

The bulldozer cpu's released next year will have the integrated graphics chips.

Which I doubt will be on the AM3+ platform. That would be FM1 or FM2 as I recall.

 

[Mackintire]

I thought, but can't remember where, I read the Bulldozers FPU performance is weak, part of the delay? and amd might be offloading some of the FPU work through OpenCL? Any truth to this?

This is a rumor, not fact.

What Tsumi said is generally true. What he omitted is that the FPU is double width and can process certain wide instructions in 1 pass where Intel's unit can do the same in 2 passes. The other missing piece of info is that AMD discovered that in most applications the FPU was being under utilized....somewhere around 55%. Therefore AMD technically has a 10% FPU deficiency by design. The trade off was have twice as many integer units. We 'll have to wait and see if AMD's tradeoffs are worth it.

 

[ebduncan]

well on bulldozer will also have a 128bit FPU.

 

[Tsumi]

This is a rumor, not fact.

What Tsumi said is generally true. What he omitted is that the FPU is double width and can process certain wide instructions in 1 pass where Intel's unit can do the same in 2 passes. The other missing piece of info is that AMD discovered that in most applications the FPU was being under utilized....somewhere around 55%. Therefore AMD technically has a 10% FPU deficiency by design. The trade off was have twice as many integer units. We 'll have to wait and see if AMD's tradeoffs are worth it.

Thanks for clearing up some of the technical stuff. I only went by what were on the pretty pictures

 

[JF-AMD]

I thought, but can't remember where, I read the Bulldozers FPU performance is weak, part of the delay? and amd might be offloading some of the FPU work through OpenCL? Any truth to this?

No.

Think of this:

Today's Magny Cours is 12 128-bit FPUs or 48 single precision floating point operations per cycle.

Interlagos will have a Flex FP that can either be 8 x 256-bit or 16 x 128-bit. Either way you are looking at 64 single precision floating point operations per cycle.

By my count that is 33% more.

The vast majority of the FP code out there is going to be 128-bit, few if any apps are going to really run 256-bit FP. That means Intel's SB-EP will have 8 128-bit units running a total of 32 single precision operations vs. the 64 in Interlagos.

I am pretty sure that when it comes to floating point in real world (not benchmarks) you are going to see a huge advantage for AMD.

 

[Devilpup]

Interesting, but how the FPU relates to gaming and fps, and whether that translates into better performance clock-for-clock, is what I would like to know

 

[Flyinfinni]

Thats a mostly Integer deal there Devilpup. FPU is not as big a deal in gaming and fps.

 

[JF-AMD]

Yes, FP is not important for most games. I laugh when the gamers talk about it. They should be thinking about integer and integer pipes.

FP is big for financial, HPC, technical, engineering, etc. Some rendering apps as well.

 

[griff30]

Yes, FP is not important for most games. I laugh when the gamers talk about it. They should be thinking about integer and integer pipes.

FP is big for financial, HPC, technical, engineering, etc. Some rendering apps as well.

It was, back in the 90's when AMD fans were using the K6-3 and trying to keep up with the Pentium3 we were told it was AMDs FPU that sucked. It may not be so much now but back then it was what everyone fed us.
Then Athlon came and smeared Intel in FPU, we were told this was how AMD finally beat Intel in gaming.
I could cite lots of old reviews to prove it, 1999 review of Athlon 600 VS Intel P3 650

With a sinking heart we turn to the built into Athlon FPU block. As all of you probably remember, floating point operations used to be a real Achilles' heel for all the previous AMD processors. The main problem was the non-pipelined FPU block in K6, K6-2 and K6-III. As a result, the general performance was unbelievably low though AMD FPU managed to spend fewer time steps than Intel's processors on most floating point operations, because the FPU couldn't start working on the next operation until the previous one was completed. However, AMD didn't feel like changing anything in their FPU in those times and called to the developers for replacing it with 3DNow!.

Only working at 650MHz Pentium III can manage to catch up with a new super FPU of AMD Athlon. This situation strikes as very unusual but seems to be extremely promising especially for 3D games. At last AMD processor will be able to give it hot there!

 

[pxc]

It was, back in the 90's when AMD fans were using the K6-3 and trying to keep up with the Pentium3 we were told it was AMDs FPU that sucked. It may not be so much now but back then it was what everyone fed us.

Something changed, namely hardware T&L with the GeForce 256 and the next year, on the ATI Radeon. Games started taking advantage of the performance available on video card hardware directly and indirectly instead of writing CPU code to do those tasks.

CPU SIMD performance can affect gaming performance (software-based physics and certain types of AI). It really depends on the game.

 

[savantu]

No.

Think of this:

Today's Magny Cours is 12 128-bit FPUs or 48 single precision floating point operations per cycle.

Interlagos will have a Flex FP that can either be 8 x 256-bit or 16 x 128-bit. Either way you are looking at 64 single precision floating point operations per cycle.

By my count that is 33% more.

The vast majority of the FP code out there is going to be 128-bit, few if any apps are going to really run 256-bit FP. That means Intel's SB-EP will have 8 128-bit units running a total of 32 single precision operations vs. the 64 in Interlagos.

I am pretty sure that when it comes to floating point in real world (not benchmarks) you are going to see a huge advantage for AMD.

Shouldn't you be comparing Interlagos with SB-EX or at least Westmere EX?

16x128 for Interlagos or 16x256bit in FMA mode
20x128 for Westmere
16x128 for Sandy Bridge EX or 16x256bit in AVX mode

 

[Topweasel]

Shouldn't you be comparing Interlagos with SB-EX or at least Westmere EX?

16x128 for Interlagos or 16x256bit in FMA mode
20x128 for Westmere
16x128 for Sandy Bridge EX or 16x256bit in AVX mode

Depends on where it is priced. It could be that they would expect that Westmere and its 10 cores, and SB-EX as the ultra high-end server CPU from Intel. Just looking at prices a 6 core Westmere starts at $150 cheaper then the most expensive MC available, and goes up epically from there.

Chances are that Intel will be pricing their larger counts out of the range AMD will be pricing their 16 core options.

 

[tangoseal]

you guys do realize that bulldozer is a server part by definition right? Hence the need for more FPU performance.

Everyone thinks its going to be a gaming processor, well even though it may be used for games, just as some use Xeon to great success, the Dozer is a 100% server designed chip. I intend to build a bulldozer for the performance of video encoding if it proves to give a run for the money on current Xeons and Sandy server chips.

So we need to stop pre-comparing a server segmented product to current desktop technology. In a server environment I guarantee an 8 core bulldozer is going to absolutely smoke all the 2600Ks you can throw at it. Maybe not so in games. However wait until AMD releases the Dozer with integrated GPU for the gaming/desktop segment early next year and then we will be able to start talking about it like a gaming chip.

Just my 2 cents and I do have to laugh at all the misinformed or self-assumed experts on these matters always trolling how Intel Sandy is going to kill bulldozer in Crysis2 and yada yada yada

 

[pxc]

In a server environment I guarantee an 8 core bulldozer is going to absolutely smoke all the 2600Ks you can throw at it.

The 2600K is not a server part and AMD throwing 2 BD die in one package for Interlagos suggests an 8 core part won't necessarily be competitive with single die SB Xeons when those are released.

 

[Devilpup]

To those that answered my question without being asshats, thank you. Some people can be real jerks when a guy is just trying to learn something

 

[sirmo]

The 2600K is not a server part and AMD throwing 2 BD die in one package for Interlagos suggests an 8 core part won't necessarily be competitive with single die SB Xeons when those are released.

To make such a speculative claim you'd at least need to know where the Interlagos would be priced.

But you don't know that, or anything about the actual performance of the BD.

 

[JMccovery]

The 2600K is not a server part and AMD throwing 2 BD die in one package for Interlagos suggests an 8 core part won't necessarily be competitive with single die SB Xeons when those are released.

Having a dual-die MCM says nothing about Interlagos' performance vs single die Xeons. Interlagos' main competitors are Magny-Cours (more cores, higher FP performance, same power consumtion, same socket), Westmere-EX and Sandy Bridge-EX/Ivy Bridge-EX. Valencia (single-die) will compete with Sandy Bridge-EN and -EP Xeons.

Besides, aren't there already Sandy Bridge Xeons?

 

[Pieter3dnow]

you guys do realize that bulldozer is a server part by definition right? Hence the need for more FPU performance.

Everyone thinks its going to be a gaming processor, well even though it may be used for games, just as some use Xeon to great success, the Dozer is a 100% server designed chip. I intend to build a bulldozer for the performance of video encoding if it proves to give a run for the money on current Xeons and Sandy server chips.

So we need to stop pre-comparing a server segmented product to current desktop technology. In a server environment I guarantee an 8 core bulldozer is going to absolutely smoke all the 2600Ks you can throw at it. Maybe not so in games. However wait until AMD releases the Dozer with integrated GPU for the gaming/desktop segment early next year and then we will be able to start talking about it like a gaming chip.

Just my 2 cents and I do have to laugh at all the misinformed or self-assumed experts on these matters always trolling how Intel Sandy is going to kill bulldozer in Crysis2 and yada yada yada

It is one of the things where AMD put a lot of effort behind it for several years now, most of us look forward to this CPU.

The server part is where it comes down to shutting down cores which aren't needed when compared with say a Phenom II, there is much more but I'm not sure that you should say that because of these features it is a server part or it won't perform with gaming , what use would AMD have releasing a CPU that is for consumers but only useful for servers ?

That doesn't fly for me because what would Joe consumer do with a server part for his desktop computer ? Why not sell it for businesses only ?

The advancements over the Phenom II clearly are there otherwise they wouldn't have bothered for the past 5 years now ? Everytime we get the same story , will AMD release something that will kill Intel , the answer is no (there is all kinds of drama involved in here with people saying that AMD sucks and so on)

Will it be able to get close is another question.

 

[sirmonkey1985]

It is one of the things where AMD put a lot of effort behind it for several years now, most of us look forward to this CPU.

The server part is where it comes down to shutting down cores which aren't needed when compared with say a Phenom II, there is much more but I'm not sure that you should say that because of these features it is a server part or it won't perform with gaming , what use would AMD have releasing a CPU that is for consumers but only useful for servers ?

That doesn't fly for me because what would Joe consumer do with a server part for his desktop computer ? Why not sell it for businesses only ?

The advancements over the Phenom II clearly are there otherwise they wouldn't have bothered for the past 5 years now ? Everytime we get the same story , will AMD release something that will kill Intel , the answer is no (there is all kinds of drama involved in here with people saying that AMD sucks and so on)

Will it be able to get close is another question.

all AMD parts since opteron processors were first released were server parts first, consumer parts second. the difference between the opteron version and consumer version are typically server only features that consumers don't need are disabled which typically gives better performance for the parts the consumer side uses.. just as a recent example since i'm to tired to list them all.. Istanbul/phenom II x6's are the same processor, shanghai/phenom II x4 and lower variant's, Barcelona/phenom I and lower variants. the opteron versions of those processors typically come out 4-6 months before the consumer processors which is why i was surprised they even tried to say they would release the consumer based bulldozers before the opteron's, since that makes absolutely no sense profit wise for them.

as far as it beating intel performance wise, LGA-1155 most likely yes, LGA-2011 not a chance. but the difference is AMD can easily compete on price. LGA-2011 will be so ungodly expensive that its really only going to cater to that very small group of enthusiast that have more money then they know what to do with and Intel will be able to charge that price due to the much larger performance difference between bulldozer and LGA-2011 on the consumer side, but AMD doesn't actually need to compete with LGA-2011. it just needs to compete with LGA-1155. but on the server side it may be a totally different story since AMD does have the G34 socket which comes in 1p/2p/4p with quad channel DDR3 which all G34 opterons support unlike Intel which will have a 2p and 4p version of each processor with a huge price difference between them.

Having a dual-die MCM says nothing about Interlagos' performance vs single die Xeons. Interlagos' main competitors are Magny-Cours (more cores, higher FP performance, same power consumtion, same socket), Westmere-EX and Sandy Bridge-EX/Ivy Bridge-EX. Valencia (single-die) will compete with Sandy Bridge-EN and -EP Xeons.

Besides, aren't there already Sandy Bridge Xeons?

The 2600K is not a server part and AMD throwing 2 BD die in one package for Interlagos suggests an 8 core part won't necessarily be competitive with single die SB Xeons when those are released.

just because they are a dual die design doesn't mean its going to lose any performance. its just more cost effective to do it that way since its a hell of a lot easier to put dual dies on a single processor then it is to fit dual IMC's on a single die. they are interconnected within the processor and then connected to the other processors with the hypertransport link. and yes there are lga-1155 xeon sandy bridges limited to 2p if i remember correctly. they also have lga-1756 or what ever the hell it is with the 10 core/20 thread westmere-EP processors running quad channel DDR3 but limited to a maximum of 2.4Ghz and are slower than snot when compared to say a W3690 6 core/12 thread @3.4Ghz(3.7Ghz turbo).

 

[Mr. Bluntman]

Interesting, but how the FPU relates to gaming and fps, and whether that translates into better performance clock-for-clock, is what I would like to know

Thats a mostly Integer deal there Devilpup. FPU is not as big a deal in gaming and fps.

Yes, FP is not important for most games. I laugh when the gamers talk about it. They should be thinking about integer and integer pipes.

FP is big for financial, HPC, technical, engineering, etc. Some rendering apps as well.

A very strong FPU used to be important back in the days of early software 3D rendering. I remember Quake in particular loved the FPUs of the old Pentium Pros and PIIs. Now with the GPU doing most of the heavy lifting, Flyinfinni and JF-AMD are right in saying the floating point unit isn't as important for games.