How much is AMD behind.

[ninaholic37]

I think AMD will surpass Nehalem performance this fall with the release of the AMD FX-8350 processor.

Their idle power performance is neck and neck with Intel while the load power usage is slightly higher than Intel.

The real problem comes when you try and overclock an AMD FX processor to be competitive with a modern Intel design and the power usage just goes through the roof.

AMD FX processors are finally priced accordingly to their performance, with the "high" end FX-8150 costing only $200.

I agree that it makes more sense to wait for FX-8350 and compare Ivy/Intel to that instead of AMD's 9 month old chip.

One more comparison using hardware.fr numbers is here .
Compared to top IB 3770K,in applications FX8150 is 183.7/150.7=1.22 or 22% slower ,stock vs stock. This is not bad at all since we have "slow" 32nm FX "8core"(which actually has 4 floating point,8 threaded subunits) versus 4C/8T IB @ 22nm. 22% slower in desktop workloads is peanuts gap and can be closed with Vishera with little to no problems.

I was curious about the actual difference too. 22% difference at stock speeds? So that is what people are complaining about? :awe:

Looks like things get a bit different when they're overclocked though:
http://forums.anandtech.com/showthread.php?t=2255552

 

[Charles Kozierok]

I was curious about the actual difference too. 22% difference at stock speeds? So that is what people are complaining about? :awe:

If AMD had put out a chip 22% faster than Intel's best offering, would you think that was no big deal? 🙂

 

[inf64]

First of all "fat" core is reference to design. It's 4way OO processor with SMT. Each Gulftown/Nehalem/SB core is around 50% larger than each K10 core and more than this than each BD core(note: I'm talking about non-cache die area,the parts that are doing actual computing: integer/fp parts with belonging front ends/backends). This is the way intel and AMD designed the CPUs. AMD went the "narrow" route and intel went the "fat" core route. The latter means more IPC but also more complexity and less clocking ability. Intel combat the clocking issues with their superior process technology.

AMD's approach can now be considered as a "narrow" 2way integer core with no SMT. It can clock rather high but since AMD depends on crappy GloFo's performance they are stuck for now. PD core will address this with some novel ideas and solutions to overcome the power/clock wall current BD iteration has.

SO we have this fat ,very IPC strong core(32nm 2nd iteration of Nehalem) that is 15% faster in single thread workloads and 16.4% faster in MT workloads than 4M/8T Bulldozer chip,both at stock. Die size is intel's advantage since they always had better cache density numbers than AMD (by a large factor). Since cache is making most of the bulldozer die it's natural it's 315mm^2 and yes,it's larger than Westmere. If AMD had access intel's process technology do you all think this would be the case? No of course. It would be roughly a parity between the two.

On to the question about why I use Westmere. It's simple,it's 12T chip and it's STILL faster than 8T IB/SB in MT workloads. Yet, this 12T chip is only 16.4% faster than FX8150 at measly 3.6Ghz(a slim high clocking design,not IPC monster). Who is doing better now? AMD is doing awesome in real workloads with their measly clocked FX chip against Westmere.

Lastly,how much is SB or IB actually faster than Westmere in ST workloads anyway? Let's find out:
i7 980x(ST Turbo @ 3.6Ghz)- 9:32 or 572s
i7 2600K(ST Turbo @ 3.8Ghz) - 8:11 or 491s

2600K is : 491/572s~=0.86 or 14% faster than 980x while clocking 5.5% higher. SO essentially 10% difference at similar clock. Not that much, is it? "Super duper" 3rd gen SB core and all you get is 10% more IPC. It's not small either but this chip is not a Westmere crusher,not even close.

To illustrate how small a difference in single thread performance is, if numbers from THG's Trinity review are replicated in VIshera(I don't see why they wouldn't ,Vishera can only be faster due to L3 and some other core changes Vs Trinity). Vishera will run at 4Ghz starting clock and we can assume that at least it will have SOME single thread Turbo uplift too. 4.5Ghz for Turbo is reasonable although I think we may see 4.6Ghz for single core Turbo. THG article shows us ~10-15% IPC uplift in 3 different workloads(itunes,lame and 3dstudiomax). Average is therefore 12.5%. All summed up : 4.5Ghz x 1.125 / 4.2Ghz ~=1.2 or 20%. In the THG chart this translates into:
FX8350 @ 4.4Ghz Turbo : FX8150 score x 0.8 <=> 673x0.8=538s. Lets round it to 540s.
SB 2600K- 491s
FX8350- 540s

Fat IPC monster core with solid Turbo clock boost(close to 4Ghz) is 491/540= 0.91 or 9% faster than slim IPC weakling with high Turbo clock. Pure integer/FP core sizes are just not comparable at the same node. Each Intel's core is 40-50% larger in die area taken versus each tiny Bulldozer integer core(and belonging half of flexfp unit). It's amazing AMD can compete and they deserve a congrats. Smart engineering.

And remember , these are NOT the workloads AMD targeted their cores in the first place. These client workloads are actually not priority for AMD.

 

[zsdersw]

Price is relatively minimal. Intel simply wins on volume.

In a given wafer size, the smaller process technology makes each chip less expensive to make; more usable chips per wafer. The economy of scale is driven by the smaller process technology and the sheer number of fabs Intel has.

The main thing is Intel can make a chip twice as complex. Meaning they can essentially make and massproduce a chip that AMD cant fabricate.

If by "complex" you mean more transistors, yes. Complex in terms of the actual design, no.

 

[Charles Kozierok]

First of all "fat" core is reference to design.

It's worth emphasizing that while the internals are interesting to geeks like us, the bottom line is what the chips actually do. Doesn't matter how you get there, as long as you do get there.

2600K is : 491/572s~=0.86 or 14% faster than 980x while clocking 5.5% higher. SO essentially 10% difference at similar clock. Not that much, is it? "Super duper" 3rd gen SB core and all you get is 10% more IPC. It's not small either but this chip is not a Westmere crusher,not even close.

Intel didn't need to "crush" its earlier chips, because AMD didn't force them to do so.

You also seem to be omitting some pretty important improvements between the 980X and the 2600K. The former is a 130W "premium" part that launched for $1000. The latter is a 95W mainstream part that launched for about $300. That puts the performance improvement in fairly significant context.

And even if SB isn't that much better than Westmere, it's still far better than what AMD can put up.. 27% using your numbers.

Vishera will run at 4Ghz starting clock and we can assume that at least it will have SOME single thread Turbo uplift too.

Unfortunately, of late AMD's chips look better before they are released than after. We'll see.

And remember , these are NOT the workloads AMD targeted their cores in the first place. These client workloads are actually not priority for AMD.

I think you've got your cause and effect reversed here. It's not like AMD made a conscious decision to avoid "client workloads" -- they are doing so because of lack of competitiveness.

 

[pelov]

Pointing out that the emperor has no clothes is highly discouraged among the peasantry. Please desist.

Pics of said emperor.



Bear in mind AMD has 8 "cores" rather than just core x2 threads (plus a ton of cache) so you'd expect the die size to be larger.

The performance on the other hand sucks.

 

[ShintaiDK]

In a given wafer size, the smaller process technology makes each chip less expensive to make; more usable chips per wafer. The economy of scale is driven by the smaller process technology and the sheer number of fabs Intel has.

Not entirely.

Volume plays a huge role. And this is also why only ~2 companies can make a profit out of 14nm. For the rest they lose money by going 14nm due to lack of volume.

Just watch the sub 28nm complains from nVidia. The transistor cost is around the same as on 28nm with TSMC. Simply because TSMC needs the money for the node progress.

Or simply look at TSMCs capacities at different nodes.

 

[zsdersw]

Not entirely.

Volume plays a huge role. And this is also why only ~2 companies can make a profit out of 14nm. For the rest they lose money by going 14nm due to lack of volume.

Just watch the sub 28nm complains from nVidia. The transistor cost is around the same as on 28nm with TSMC. Simply because TSMC needs the money for the node progress.

Or simply look at TSMCs capacities at different nodes.

Volume = "sheer number of fabs Intel has".

 

[Idontcare]

I think you've got your cause and effect reversed here. It's not like AMD made a conscious decision to avoid "client workloads" -- they are doing so because of lack of competitiveness.

AMD's business strategy 2005-2012

 

[Ferzerp]

AMD is thoroughly insignificant in the server space. We're talking 5.5% the last numbers I saw. Are you really trying to say that they are ignoring their less unhealthy area (client) for their absolute worst market (server)?

Really?

BD, for various reasons, is even less attractive in the server space than in the desktop space. Part of it is due to the impact the artificially inflated core count has on certain product's licensing, part of it is the underwhelming performance, and part of it is the processor price savings becomes so thoroughly diluted in the overall cost of a server that it makes very little sense to go with Opterons.

 

[khurios2000]

AMD's business strategy 2005-2012

they meant to do that 😎

 

[Jeff7181]

I remember saying that about my core 2 duo E8400 four years ago. And I compile code and run processor-heavy apps regularly. I still can't justify upgrading, but I probably will next yearish when Haswell pops out. I consider it a testament to [Intel] CPU development that I can skip 2 full tick-tocks and a halving of the process size without needing to upgrade.

I'm in the same boat. I wish I would have picked up a Q9550 before they disappeared from retailers. Would be nice to have a couple extra cores when I'm transcoding movies.

 

[grimpr]

AMD's business strategy 2005-2012

That takes balls.

 

[Nemesis 1]

Pointing out that the emperor has no clothes is highly discouraged among the peasantry. Please desist.






























😀 😉

Good one!

 

[Obsoleet]

Time permitting there is another elephant in the room Apple. Intel with medfield pits itself against Apple. I would say apple buys AMD.
Yet another elephant is making its way into the room . China

Over time, Samsung, someone from China, or a middle eastern company will acquire them.

 

[Obsoleet]

If AMD had put out a chip 22% faster than Intel's best offering, would you think that was no big deal? 🙂

It's a double standard, and there should be IMO. David vs Goliath.

 

[moonbogg]

The gap that seperates AMD from Intel runs deep and wide, like the marianas trench!

 

[Puppies04]

2600K is : 491/572s~=0.86 or 14% faster than 980x while clocking 5.5% higher. SO essentially 10% difference at similar clock. Not that much, is it? "Super duper" 3rd gen SB core and all you get is 10% more IPC.

No it isn't "all you get" SB also runs cooler and clocks higher. Shame bulldozer couldn't pull this feat off for AMD or we might not be having this conversation....

 

[Jeff007245]

Around what time frame will 2nd Generation Refresh AMD FX CPU's be released?

 

[AtenRa]

Even more embarrassing for AMD, FX-8150 versus i5-750:

http://www.anandtech.com/bench/Product/434?vs=109

Bulldozer wins on the heavily multi-threaded benchmarks. Intel's desktop R&D department could have taken a three-year holiday and they would probably still be ahead of the curve.

Im not sure you realized that some graphs say "lower is better".

I believe you should recheck because there are only 2-3 benchmarks from the above link that i5-750 is faster than FX8150.

Even the Core i3-530 doesn't have a bad showing against the FX-8150:
http://www.anandtech.com/bench/Product/434?vs=118

Being 2-3 times slower is not bad showing ?? then FX8150 being 10-20% slower than Core i7 2600K is nothing to talk about, don't ya think ??

 

[AtenRa]

around what time frame will 2nd generation refresh amd fx cpu's be released?

q3-q4 2012

 

[nenforcer]

I doubt anyone thinks a Phenom III X8 sounds bad, given that the cores perform as well as Phenom II cores did. Simply throwing full speed L3, AES, AVX, and SSE4 support onto K10 would've helped out significantly - the 32nm die-shrink probably would've boosted clocks above 4.0 GHz, as well.

 

[Rifter]

They are at least 3 years behind.

 

[Haserath]

You can't really base how far they are behind by time. They could come out with their own core 2 efficiency/performance jump and blow Intel away.

http://www.tomshardware.com/reviews/fx-power-consumption-efficiency,3060-13.html

I'd say they're a little over 100% behind efficiency wise, which is a very significant metric to measure by.

I saw these and couldn't believe BD didn't improve at all on efficiency against older 45nm designs. Shows me just how bad this first implementation is.

 

[Riek]

You can't really base how far they are behind by time. They could come out with their own core 2 efficiency/performance jump and blow Intel away.

http://www.tomshardware.com/reviews/fx-power-consumption-efficiency,3060-13.html

I'd say they're a little over 100% behind efficiency wise, which is a very significant metric to measure by.

I saw these and couldn't believe BD didn't improve at all on efficiency against older 45nm designs. Shows me just how bad this first implementation is.

Put llano next to it and a 4core phenom and you'll notice the biggest hurdle is the 32nm which only gave a higher density but nothing on power or frequency spectrum.