AMD chief say BD will offer only 35% not 50% more performance than previous gen

[Imouto]

What's with the immature attitude? I asked a simple question, you fire back like a 13yr old who's dad works for AMD or something.

Check it at the door, you won't be long around here walking around with that chip on your shoulder.

Your mod status allows you to behave in the same way that you're blaming others for? You're calling me AMD fanboy out of the blue only for pointing you out something you should know for years already?

Your attitude has been bad, its not Idontcare, its you that has a problem. And insulting a mod is the best way to get a quick vacation.
Knock it off, or you will be out of here.
Markfw900
Anandtech Moderator.


Check your manners yourself, a moderator should't argue with users like you do really often. Get a fake account or something to post and argue with us but don't show off your powers and threat ppl with incoming bans.

Who's the 30+ year old childish guy now?

BD module design provides 80% of the performance of 2 separate cores.

wow, is that a year old paper? Based on "estimations"? Why don't you stick to the OP's linked info? It's from the same source you know.

These are my alternative calculations:

If Bulldozer is 25% faster in IPC over Phenom II, then 1 BD module = 2 Phenom II cores in IPC (i.e., 0.80x penalty * 1.25 increase in IPC = 1.0x 2-core Phenom II). My feeling is that BD will achieve most of its performance gains from clock speeds and more cores. Its average IPC increase over Phenom II won't be more than 20% at the same clock speeds.

Man, you're saying that IPC increase will be zero over PhII and then you're saying that IPC will increase by 20% out of nowhere.

 

[RussianSensation]

I see what you did there. Apple was never known for performance. Never. They always put chips in their systems that were (1) Not expensive (2) Low power usage. Just how you think they get those insane margins on their products?

Not going to write a 5 page response about how Apple gets high profit margins, but the Core i5 chips in MacBook airs are anything but cheap. In fact they cost more than the i5 2500k CPU. $250 for the i5-2557M. Apple doesn't sell X4 or X6 processors because they are inferior in performance/watt. Also, very few people sit there and do 12 hours of video rendering on their laptops/desktops that they would require 6 or 8 cores. Performance / watt is key and AMD has been behind in this area for 5 years.

Read my initial comment. Also, on the flip side, Intel's GHZ GHZ GHZ GHZ marketing was ... honest? Let's stay out of strawman territory, shall we?

I never supported the Ghz myth. I am always about performance per clock / watt for CPUs. To me the Core myth is the new Ghz myth, just saying.

It's going to target that market because (1) It's way more expensive than other platforms (2) The vast majority of people don't need anything near that amount of CPU power. Thus it's a 'niche' platform.

So in other words for the average consumer a 6 or an 8-core CPU is a waste of $ if it's not faster in 2-4 threaded apps. Agreed.

Wrong. The main reason AMD dominated Intel is because Intell fell asleep and let marketing dictate the orientation of products instead of engineers. You know, the Gigahertz uber alles plan.

Intel had another architecture alongside the Netburst - Pentium M. Intel didn't fall asleep, they just miscalculated the ability for their Netburst architecture to scale with higher clock speeds. Regardless, you aren't giving AMD engineers any credit. The Athlon 64 had 64-bit support, onboard memory controller, among other innovations in the consumer market.

Also, 34% of PC gaming is a drop in a bucket compared to consoles. Let's put things into perspective. While PC gaming has a few anomalies like WoW and SC2, the vast majority of PC game sales numbers are eclipsed by consoles. If you think the PC gaming market is so important, I got a bridge for sale you might be interested in!

What do consoles have anything to do with this discussion?

You still think consumers don't care about gaming on their PCs/Macs?

http://www.apple.com/macbookpro/

First page of MacBook Pros:

"Game-changing graphics. AMD Radeon graphics processors on the 15- and 17-inch models are up to 3x faster....".

Apple even provided marketing charts for HL2 and Portal. Mac is not even a "gaming" platform so to speak and even Apple got Steam support.

The average consumer will walk into the BestBuy store and say, I want a snappy computer for every day tasks, good enough for casual games, doesn't run too hot and consumes little power. What should I get? The answer isn't going to be a 6- or 8-core BD. Those CPUs target guys who know what IPC is what performance/watt is and what a core means.

The reason AMD has been able to sell their CPUs all these years is because they continue to compete in the <$140 price range. Their Athlon II X4 CPUs were excellent budget CPUs. But AMD can't just continue to give us more cores without improving the performance / watt of those cores - that's all I am saying.

 

[RussianSensation]

Man, you're saying that IPC increase will be zero over PhII and then you're saying that IPC will increase by 20&#37; out of nowhere.

I didn't say that at all. You would know that considering I started my statement with "IF".

If you don't understand what I am saying, then please don't put words in my mouth. Also, the number behind 80% efficiency of 2 full-fledged cores came from AMD, not me. I didn't make those #s up.

 

[Imouto]

I didn't say that at all. If you don't understand what I am saying, then please don't put words in my mouth.

then 1 BD module = 2 Phenom II cores in IPC
Its average IPC increase over Phenom II won't be more than 20% at the same clock speeds.

It's perfectly clear for me.

btw with that "same clock speeds" looks like you don't know what IPC means by any rate. IPC doesn't change with clock speed.

 

[RussianSensation]

then 1 BD module = 2 Phenom II cores in IPC
Its average IPC increase over Phenom II won't be more than 20&#37; at the same clock speeds.

It's perfectly clear for me.

btw with that "same clock speeds" looks like you don't know what IPC means by any rate. IPC doesn't change with clock speed.

I don't know how to make myself any more clearer.

Assuming you converted 2 Phenom II cores into a 2 Phenom "module" design, you would instantly lose 20% performance according to AMD. Which means your 2 core BD module design would have 0.8x of the performance of your original 2-core design (independent 2 cores).

You said BD = 1.35x IPC vs. Phenom II.

For this to happen, BD would need to be 69% faster per clock than a Phenom II is today (1.35x / 0.80x = 1.6875x). If you don't understand what I am saying, I don't know how else to explain it. In other words your 1.35x claim is bogus at the same clock speeds. BD will be clocked much higher and have more cores than Phenom II, because its IPC increase won't be anywhere near 35% as you claim.

Actually, in archiving , you took a single threaded apps as WINRAR

Winrar is not single threaded. It even has the ability to toggle single vs. multi-threaded support in its benchmarking mode.

All this time AMD positioned their X4 processors vs. i3s and their 6 core processors vs. i5s, and it still didn't do much for them. Assuming BD has a much improved performance per core, it will be an excellent CPU. However, if it's just a higher clocked Phenom II IPC-style processor with more cores thrown in, it will be a huge disappointment to many. AMD adding more cores didn't really turn out to be the winning strategy they hoped it would be against Intel with Phenom II. I hope this time they made major strides in performance per core metric.

 

[Imouto]

Check topic's title please.

 

[RussianSensation]

Check topic's title please.

I did read the source, did you?

"Insufficient clock-speeds of Bulldozer are probably the reason why AMD now claims that the 16-core offering will be 35&#37; faster than 12-core solution (which is natural, given 33% higher core count) and not 50%, as it initially expected. It is also noteworthy that Bulldozer's per-core performance is not projected to be much higher compared to existing microprocessors."

The bulldozer module apparently only adds 12% more die space. For the server market, this is an excellent way to pack more cores.

However, in the consumer market (i.e., for us), the situation is entirely different. Most of us don't really need 8-12-16 Phenom II style cores. I hope the bolded statement is NOT true, OR

 

[LOL_Wut_Axel]

Assuming you converted 2 Phenom II cores into a 2 Phenom "module" design, you would instantly lose 20&#37; performance according to AMD. Which means your 2 core BD module design would have 0.8x of the performance of your original 2-core design (independent 2 cores).

Only when the cores need to share resources from what I understand. When only one core is being used it should have the same performance as a normal core. When using two cores in the module it's when you lose performance since less resources are allocated to each one of them. Perhaps I'm wrong; however, I think this demonstrates it best:

One core in the module being used:

Two cores in the module in use:

Thanks to AtenRa for posting these images. They're great for visualizing how it should work.

It's 100% the same performance when one core is in use and 80% the performance when two are in use. So no, you don't automatically lose 20%.

 

[Imouto]

Check the "Insufficient clock speeds" at your quote, please. It isn't talking about IPC, it's talking about performance at stock speeds for both 16 core Interlagos and current 12 core Opteron.

If you're lost already is like comparing a 2 Ghz Sandy Bridge vs a 3 Ghz Nehalem.

Given the reduced manufacture process and new architecture desing yeah, I expect a performance improvement. If it won't happen I see a really dark and expensive future for this hobby.

 

[GaiaHunter]

I don't know how to make myself any more clearer.

Assuming you converted 2 Phenom II cores into a 2 Phenom "module" design, you would instantly lose 20&#37; performance according to AMD. Which means your 2 core BD module design would have 0.8x of the performance of your original 2-core design (independent 2 cores).

AMD said nothing of the sort.

AMD said 1 BD module compared to 2 hypothetical BD cores.

"Insufficient clock-speeds of Bulldozer are probably the reason why AMD now claims that the 16-core offering will be 35% faster than 12-core solution (which is natural, given 33% higher core count) and not 50%, as it initially expected. It is also noteworthy that Bulldozer's per-core performance is not projected to be much higher compared to existing microprocessors."

I hope the bolded statement is NOT true, OR

Existing microprocessors, hmm, is SB an existing microprocessor?

Or are you expecting BD to be much faster than existing microprocessors?

 

[lau808]

iirc, its not 2 "true" cores x 80&#37; i thought they said 180% of 1 true core. no link/source as god knows how long ago nor where i saw it. but u have 100% core with 1 thread/module or if running both cores in module then 180% core performance. idk if im even working this correctly so forgive me

 

[Mopetar]

iirc, its not 2 "true" cores x 80% i thought they said 180% of 1 true core. no link/source as god knows how long ago nor where i saw it. but u have 100% core with 1 thread/module or if running both cores in module then 180% core performance. idk if im even working this correctly so forgive me

I'm pretty sure it was covered at some point in the gigantic BD rumor thread. IIRC, the general assumption was that a BD module averaged around 180% scaling, whereas an Intel core with HT provided between 110% and 145% depending on the application.

More generally, HT performance will be worse the better the application can utilize an Intel core. If the program rarely has branch mispredictions, cache misses, etc. there's less of a window for the virtual core to get any work done. With BD's approach, the performance only drops off when the application manages to bottleneck any part of the module. How well an arbitrary application does remains to be seen, but if a module scales on average 180%, it's not overly bad in general.

 

[Phynaz]

Actually, in archiving , you took a single threaded apps as WINRAR
that favour the 2500K , not counting that this apps is Intel optimised.

Why not take a popular multithreaded soft instead ?...

Winrar is multi-threaded.

Got a link to your claim it is "Intel optimized"?

 

[Imouto]

http://www.winrar-tr.com/winrar/Help/ENG/html/HELPCompressionAdvanced.htm

btw paying developers to optimize benchs and **** is old already, just check the new NVIDIA optimized Crysis or Lost Planet for both NVIDIA and Intel. AMD does the same too, don't get me wrong. I just can't recall the last AMD promoted game.


Profanity is not allowed in the technical forums.

Idontcare
Super Mod

 

[lau808]

so theoritcally, ipc of bulldozer 8 core needs approx 66&#37; ipc of i726k to be up and up if app is using all 8 threads for each cpu right? thats realistic cause isnt phenom II ipc already around there? 4 core bulldozer would need 10% ipc advantage over i525k tho to be up and up in same situation tho since i5=no ht :/

 

[RussianSensation]

It's 100&#37; the same performance when one core is in use and 80% the performance when two are in use. So no, you don't automatically lose 20%.

But if you intend to buy a 4/6/8 core processor, you are most certainly looking to run programs which are multi-threaded. So the scenarios where the 8 core processors will only be running 1 thread is going to be very rare, if at all (and if such, it would be a huge waste of $ to buy a multi-core CPU for such user). Even games now use at least 2 cores. Therefore, the 20% penalty will most certainly be present in 95% of the time as almost all modern code uses at least 2 threads.

AMD said nothing of the sort.

AMD said 1 BD module compared to 2 hypothetical BD cores.

Did you not read what I am saying? If BD = Phenom II in IPC, then the 80% statement applies.

Assuming 2A (Bulldozer) = 2B (Phenom II), AMD said A (module design) is only 80% as efficient as B design. Therefore --> 0.8 (2A).

How are you not following this? In other words, if BD has equal IPC to Phenom II, then out of the gate, its 2 cores are only 80% as fast as 2 equivalent independent cores. In other words, assuming no change in IPC over Phenom II, 2 bulldozer cores will only be 80% as fast at the same clock speeds.

He claimed that BD will be 1.35x faster than Phenom II. In other words, he claims that 0.8 (2A) / 2B = 1.35.

The only way for this to occur at the same clock speeds is if 2A is 1.69x faster than 2B (i.e., 2 core Bulldozer is 69% faster than a 2 core Phenom II).

Now think about how ridiculous that statement is.

Existing microprocessors, hmm, is SB an existing microprocessor?

Or are you expecting BD to be much faster than existing microprocessors?

Wow, you just love twisting information, don't you?

Read the paragraph as it is.

Insufficient clock-speeds of Bulldozer are probably the reason why AMD now claims that the 16-core offering will be 35% faster than 12-core solution (which is natural, given 33% higher core count) and not 50%, as it initially expected. It is also noteworthy that Bulldozer's per-core performance is not projected to be much higher compared to existing microprocessors.

In the above paragraph, the comparison is clearly being made between a new 16-core AMD micro architecture and their modern (i.e., current) 12-core micro architecture. The reference to "per-core performance" is in regard to previous statement (i.e., their own existing microprocessors, and has nothing to do with SB).

iirc, its not 2 "true" cores x 80% i thought they said 180% of 1 true core. no link/source as god knows how long ago nor where i saw it. but u have 100% core with 1 thread/module or if running both cores in module then 180% core performance. idk if im even working this correctly so forgive me

The statement explains that AMD made a compromise between performance and die space. They accepted 80% performance of 2 full fledged cores for an increase of die space of just 12%. It sounds like a great tradeoff, but it also means a 6-core BD with same IPC as Phenom 2 is by their own definition can only be 80% as fast as a 6-core Phenom II at the same clocks. It's no wonder AMD is doing everything they can to launch BD at much higher clock speeds than Phenom II ships at.

 

[Imouto]

So BD will be worse than PhII.

And he's totally serious about it.

This is hilarious.

 

[386DX]

iirc, its not 2 "true" cores x 80% i thought they said 180% of 1 true core. no link/source as god knows how long ago nor where i saw it. but u have 100% core with 1 thread/module or if running both cores in module then 180% core performance. idk if im even working this correctly so forgive me

Basically because of the shared nature of a BD module you won't get the same scaling as a true core (ie. Phenom II). It's like this:

Assuming 1 Core = 100% performance

On Phenom II
1 Core = 100% performance
2 Core = 200% performance (relative to 1 Core)
6 Core = 600% performance

* A Theoretical 8 Core P II would have 800%

On Bulldozer
1 Core = 100% performance
2 Core (1 module) = 180% performance (relative to 1 Core)
8 Core (4 module) = 720% performance

In my previous post I used some math to figure out that approximate IPC that BD should have over Phenom II to make the statement in the OP true "35% faster, with 33% more core"

Doing the math you will come out that Bulldozer should have a 12% IPC improvement over Phenom II. Based on this calculation:

12 core Magny:
1 Core = 100% performance
2 Core = 200% performance
12 Core = 1200% performance

16 Core (8 Module) Interlagos:
1 Core = 112% performance (relative to 1 Core Magny)
2 Core (1 Module) = 201.6% (112% x 1.8 scaling) performance (relative to 1 Core Magny)
** Notice that the 2 Core 1 module has almost the same performance as a 1 core Magny and not the 224% you would expect**
16 Core (8 Module) = 1612.8%

1612.8% is approximately 35% faster then 1200% which validates the OP article that BD (Interlagos) is 35% faster with 33% more cores.

We can also expect that BD will have around 12-13% IPC over Phenom II when 1 Core is active, but have 0% IPC over Phenom II when 2 Cores (1 module) is active. Note these calculations are based on 1 module having 180% (=90% per core, or 10% penalty) performance other people in this thread are going with 80% (or 20% penalty) which I dunno where they got it from.

 

[Imouto]

So buying a BD for single threaded apps would be really stupid with SB rolling everywhere and buying it for multithreaded apps would be even more stupid since at 720&#37; theorically max output a 2600K would match it performance wise.

If we take RussianSensation figures the thing gets even worse.

 

[LOL_Wut_Axel]

But if you intend to buy a 4/6/8 core processor, you are most certainly looking to run programs which are multi-threaded. So the scenarios where the 8 core processors will only be running 1 thread is going to be very rare, if at all (and if such, it would be a huge waste of $ to buy a multi-core CPU for such user). Even games now use at least 2 cores. Therefore, the 20&#37; penalty will most certainly be present in 95% of the time as almost all modern code uses at least 2 threads.

AMD will mitigate that 20% loss with higher IPC and base clock speeds, not to mention Turbo CORE. Turbo CORE probably won't be used by most enthusiasts, though, so Bulldozer needs a good IPC increase and high over-clockability.

By sharing resources it already has a good start in single-threaded programs like audio encoding since it's just like a normal core, plus probably higher IPC. In multi-threaded ones it mitigates the 20% loss with more cores and high clock speeds, and if it has good IPC the loss won't be a bit problem.

AMD easily got 6% higher IPC out of K10.5 with Llano. If AMD isn't gonna get anything higher than that with what's supposed to be a higher-end CPU architecture, what's the point? If you're talking IPC, you bet that Bulldozer will be a big increase from K10.5. AMD isn't gonna shoot themselves in the foot by making an A8-3870 BE that over-clocks to 4-4.5GHz and is basically the same IPC as Bulldozer. The FX-4100 needs to have higher IPC than Llano and maybe higher over-clockability. Otherwise, everyone would flock to Llano instead given that it has almost the same performance coupled with the equivalent of a $50 dedicated graphics card, all at a lower price. What that would mean is that only people that use multi-threaded apps would buy the FX-6000 and FX-8000 series. AMD isn't that stupid; they know how enthusiasts view them is at the hands of Bulldozer.

 

[RussianSensation]

Doing the math you will come out that Bulldozer should have a 12&#37; IPC improvement over Phenom II. Note these calculations are based on 1 module having 180% (=90% per core, or 10% penalty) performance other people in this thread are going with 80% (or 20% penalty) which I dunno where they got it from.

How did you get a 10% penalty? The statement about 80% efficiency of a module design vs. 2 full cores came directly from Mike Butler, AMD Fellow and Chief Architect of the Bulldozer core.

If a 16-core BD is 35% faster than a 12-core Phenom II at the same clocks, then the IPC increase has to be much higher than 12%.

0.8x efficiency of a module design =>> 80% efficiency of module design x 8 modules x 2 cores per module = 12.8 "full" equivalent cores.

Therefore we have,

12.8 "full" Bulldozer cores * (1 + X) = 12 full Phenom Cores * 1.35x faster

X = 26.6%

AMD will mitigate that 20% loss with higher IPC and base clock speeds, not to mention Turbo CORE.

But wait, look what happens once you apply a 27% IPC increase:

Bulldozer = 80% (20% penalty due to modules) * (1 + 26.6% increase in IPC over Phenom II) = 1.01x

^ No wonder their statement that the "overall" performance per core didn't change. You get an almost linear performance increase with more cores. The actual increase in IPC seems to be offset almost entirely by the 20% penalty of the module design!!! The math of 16-cores being 35% faster than 12-cores exactly supports this (if all the data provided is true).

That's why I said it's more realistic that the overall performance increase for BD will come from much higher clocks and more cores. I am glad we finally got here 10 pages later. The IPC advantage over Phenom II is mostly going to be eaten away as a result of the module style design choice.

AMD easily got 6% higher IPC out of K10.5 with Llano.

AMD was promising ~ 6% increase in instructions executed per clock (IPC) for the Llano cores vs. their 45nm Athlon II/Phenom II predecessors. However,

"...On average Anand measured around a 3% performance improvement at the same clock speed as AMD's 45nm parts. Peak performance improved up to 14%, however most of the gains were down in the 3&#8212;5% range." - Source

Looks to me like they barely improved IPC after increasing the L2 cache, reorder and load/store buffers, new divide hardware, and improved hardware prefetchers. All that barely netted them 3-5%.

 

[wlee15]

What will be interesting is Bulldozer's SIMD integer performance with it's two 128-bit ADD and 1 128-bit MUL (or a 128-bit multiply accumulate with the XOP instructions) per module compared to only one 128-bit ADD and one 128-BIT MUL.

 

[LOL_Wut_Axel]

AMD was promising ~ 6&#37; increase in instructions executed per clock (IPC) for the Llano cores vs. their 45nm Athlon II/Phenom II predecessors. However,

"...On average Anand measured around a 3% performance improvement at the same clock speed as AMD's 45nm parts. Peak performance improved up to 14%, however most of the gains were down in the 3&#8212;5% range." - Source

Looks to me like they barely improved IPC after increasing the L2 cache, reorder and load/store buffers, new divide hardware, and improved hardware prefetchers. All that barely netted them 3-5%.

LOL? Not sure if serious. It's the same architecture. The fact they're able to even get a 5% IPC increase from an architecture that's already had everything milked out of it very impressive. The revised K10.5 in Llano is HEAVILY based on K8. The fact that they've been able to milk out so much IPC increase from a 2003 architecture that's had many revisions is downright impressive.

And now you're saying AMD won't get better performance in comparison to what basically amounts to a 2003 architecture? Come on. With Llano AMD already has IPC between Core 2 65nm and Core 2 45nm. It won't take them too much to reach IPC that's within 5% of Nehalem.

 

[bridito]

There is some data, and no, it can't be slower than Atom. That makes no sense; for that to happen it'd have to be much slower than Llano. Let's just put an end to this would-be argument and focus on discussing what performance could be instead.

Ok, then as I stated in my original post which keeps getting rolled around and around, BD is faster on a single module than IBM Watson. http://www.research.ibm.com/deepqa/faq.shtml.

LOL, please try to understand. That BD is either slower than Atom or faster than Watson are extremely unlikely extremes stated only to prove the point that you and I and everyone else on this forum knows exactly ZERO about BD performance. So there is no point confronting Troll Trolling about "BD can't be slower than Atom..." BD could be slower than an 8088 or faster than G-d himself.

There

Are

No

Benchys.

That's the bottom line and that's all there is to that. Everything else is just white noise.

 

[LOL_Wut_Axel]

Ok, then as I stated in my original post which keeps getting rolled around and around, BD is faster on a single module than IBM Watson. http://www.research.ibm.com/deepqa/faq.shtml.

LOL, please try to understand. That BD is either slower than Atom or faster than Watson are extremely unlikely extremes stated only to prove the point that you and I and everyone else on this forum knows exactly ZERO about BD performance. So there is no point confronting Troll Trolling about "BD can't be slower than Atom..." BD could be slower than an 8088 or faster than G-d himself.

There

Are

No

Benchys.

That's the bottom line and that's all there is to that. Everything else is just white noise.

No, I'm not gonna "try to understand". There's nothing to understand, as your would-be argument makes no sense. If you have nothing technical to contribute to the thread apart from flame bait and are gonna keep making up ridiculous would-be arguments for your amusement please get out of the thread. We're here to discuss technical facts.

Sorry if this sounds rude, but you constantly trying to trash or deviate this thread is getting old.