Fudzilla: Bulldozer performance figures are in

[cantholdanymore]

You know the reason Ferrari participates in Formula 1.

Ferrari was building race cars way before started building regular cars

 

[ed29a]

For the average Joes, Intel has the perception of making better CPUs, and in the last 5 years, it was pretty much true, exacerbating the image problem for AMD.

I am not going to bother posting the Google survey video about asking people what the browser is. If people don't know what the hell a browser is, you think they know what a processor is? You are making these outlandish claims that people (dumb consumers) care about Intel Processors, have any numbers, surveys or any facts to back that claim up?

If we examine Pentium 4 and Pentium D, when Intel didn't focus on IPC, performance / watt, performance per core, then we have a situation where a CPU that didn't follow those criteria also didn't do as well in the marketplace.

If you look at marketshare graphs from AMD, you can see that the Pentium 4/D raise was mostly due in servers. You know ... a market segment where people actually know and care about CPUs.

Apple has shown that many users are willing to pay for the experience. They don't necessarily care about specs on paper as long as the product delivers - Intel CPUs do deliver.

My point exactly. Yet you kept boasting that Apple picks performance processors. Tell me again, where are the products using the 2600K, the mobile extreme or hell, the 990x? Right, nowhere. Which brings back to my original point: people don't need these processors, any 2-3 year old quad core is more than enough for the vast majority of users out there.

Notice why Apple carries Intel processors? (1) Better performance / watt (2) Better efficiency --> better battery life (3) Faster performance for most tasks a typical user cares about (4) Ability to meet manufacturing demand (but another reason why Intel has more manufacturing capacity is because there is more demand for its products).

You assume Apple users care about performance and efficiency. They might care about battery life, that's why until very recently all Mac Airs had some really weaksauce processors with great battery life. And I think you're finally getting to one of the real reasons Apple didn't pick AMD (other than battery life): ability to meet manufacturing demand.

Most people would agree that it's much easier to market a great product (i.e., Intel CPUs) than to have great marketing for a mediocre product (i.e., AMD's Phenom I/II).

(1) You assume the target audience knows about the products, they don't.
(2) Mediocre to whom? 'Enthusiasts'? One 'enthusiast's' mediocre product is one average user's speed demon.

As a side note Centrino was a brilliant marketing move by Intel because when the consumer got the product, it was superior to AMD's offerings. AMD couldn't do anything of the sort because to claim that their CPU was superior in performance / watt or battery life would have simply been a misrepresentation of the truth.

I am sounding like a broken record here. The average users doesn't know what is inside a laptop. If you think they can associate Centrino with a great mobile processor, oh boy, I have a bridge for sale you might be interested in.

Notice how the Fusion is taking off for AMD, because it really is the better gaming platform than HD3000. Great product doesn't require revolutionary marketing because it almost sells itself.

And you are sitting on Fusion sale numbers? You know that Fusion is selling like hot cakes because you have access to Dell, Lenovo, Hp, BB, FS, Newegg sales numbers? Oh and, please do a small experiment if you don't believe me. Walk into any BB or FS and ask the salesman for a gaming laptop. I will eat my own Sager if they suggest you a Fusion laptop. They will ALWAYS direct you to the Alienware, high end Toshiba or the Asus ROG laptops. Always.

Steve Jobs ditched PowerPC chips for Intel since Intel offered far superior performance / watt. You are telling me no one cares about performance / watt, IPC and power efficiency?

Duh! Did Apple's consumers knew or cared? My money is on 'No' since they get brainwashed with 'magical' presentations and ads. Apple can sell fridges to Eskimos, their marketing is that good.

I got news for you -- manufacturers and designers who strive to create a killer mobile device/laptops/tablet care because it's a lot easier to sell "10 hour battery life notebook" with a fast processor than to slap a 2 lbs of protruding battery to a laptop and try to sell that as a premium product.

That explains why Dell Adamo line is such a great success. Or why Acer Timeline is a smashing hit too. Oh wait ... Only Apple can sell 1000$+ laptops easy. Again, killer marketing and all.

Are you are still going to deny "no one cares about performance / watt" even after Intel adopted this as their key design strategy for implementing new features into their CPUs? :sneaky:

Again, don't put words in my mouth. I clearly said the average consumer doesn't know/care, Apple as a company knows and cares, and you know why? To add more 'magical' features to their marketing slogans.

If your logic was true, no one would even bother upgrading to faster chips. Laptops would only add features, better aesthetics and slimmer form factors. According to you, Intel could have easily produced Core 2 Duo since 2006 and not done a single thing to increase its performance / watt. Yes, they did exactly that...oh wait.

And the average laptop user needs a Sandy Bridge mobile extreme processor because? Had Intel done that, AMD would have caught up and Intel would have lost marketshare. Duh. Not going to explain Capitalism 101 to you. If Intel doesn't innovate the market will catch up and Intel is in trouble. However, that still doesn't change the fact that the average consumer doesn't need more than a Phenom II or Core 2 Duo. The fact that Windows 8 requirements are going to be lower than 7 just proves that even more so. But keep telling yourself that every average user needs a super duper Intel extreme processor.

But you can go ahead and deny that IPC, Performance / watt, Performance per core don't matter.

Again, don't put words in my mouth. I don't need help expressing my thoughts, thank you.

You can be sure that when Bulldozer launches, its power efficiency will be scrutinized by many people in the server and consumer markets because it directly impacts operational costs and every day usability for mobile devices.

Really? People who buy servers will examine power efficiency. Stop the presses! Next thing you are going to tell me that water is wet and fire is hot!?

Pretty much to anyone who understands where the future of computing is heading - mobile, tablet, smartphone space where (1) performance / watt (2) IPC (3) performance per core is critical. I think only you are being in denial.

Ah yes, we need quad core CPUs for smartphones. Extreme version too. It makes receiving calls and running fart apps so much faster!

Are you insinuating that if Bulldozer was simply an 8 core Phenom II, AMD would be back on track? We know it wouldn't be enough, therefore, successes in the 3 areas I mentioned are critical to win back performance.

Again, no need to put words in my mouth. I am not insinuating anything. I just mentioned two simple things:
(1) Your average consumer is dumb and doesn't know/care about computing performance. See Apple's 'magical' marketing. The iPad/iPhone are the perfect example of a device with low performance components compared to the competition but that eat away the market due to other strengths. AMD needs to focus on something like that.
(2) AMD doesn't give a rat's ass about 'enthusiast' market because it is a tiny segment. And if I were a shareholder I would agree. Invest your energy in a market that can be your bread and butter vs one that is shrinking.

Kapish?

I never said that enthusiast market is the only one that matters. But somehow you keep missing the point of what is necessary for a modern processor to be successful - Excellent power efficiency, great performance / watt, strong IPC and more cores. If you only focus on more cores, you will fail miserably.

You accuse me of ignoring IPC (which I didn't, infact, it wasn't even the goal of my post) yet you do the same thing beating the performance drum. Hypocrisy no?

 

[formulav8]

I've got to admit your definitely right on most of your counts.:thumbsup:

 

[BlueBlazer]

I fail to see how is that impressive.

BD
(7GB/s) / (2 x 12 cores) / 2.4GHz = 0.121

980x
(4GB/s) / (1 x 6 cores) / 4GHz = 0.167

And that is only Gultown not SB-E.

If those AMD "cores" are treated as hardware threads then it would be impressive, however calling them "cores" makes the score unimpressive.

 

[Idontcare]

Some people are just hard to impress I guess

Its a 2.36GHz setup, 2S/24T. And it still manages to thump a 2S/24T Intel rig that is clocked nearly 2x more (4.2GHz).

I consider that to be impressive.

 

[BlueBlazer]

Some people are just hard to impress I guess

Its a 2.36GHz setup, 2S/24T. And it still manages to thump a 2S/24T Intel rig that is clocked nearly 2x more (4.2GHz).

I consider that to be impressive.

If you look at the score here (an Intel 12C/24T 2 socket system)....

2 x 2x Intel(R) Xeon(R) CPU X5650 @ 2.67GHz

((4.727GB/s) / (12 cores)) / 2.86GHz = 0.138

"Per-core" the performance still a little higher. If calculated as "per thread" then its a different story.

 

[Idontcare]

If you look at the score here (an Intel 12C/24T 2 socket system)....

2 x 2x Intel(R) Xeon(R) CPU X5650 @ 2.67GHz

((4.727GB/s) / (12 cores)) / 2.86GHz = 0.138

"Per-core" the performance still a little higher. If calculated as "per thread" then its a different story.

Can anyone define for me what a "core" is anymore?

Performance, wall-time, that matters to me.

Power consumption does too, a missing metric for this bench.

Cost matter as well, likewise a missing piece of the puzzle here.

But in terms of deconvoluting the performance into "per core" and "per thread" attributions, not a line of thought that really holds much interest for me at a personal level at this time until we come to see the bigger picture of how a thread differs from a core in bulldozer at the performance level.

We have a good idea how these two terms differ for Intel's use of them, but we remain to be enlightened as to how this differs for AMD.

 

[Abwx]

If you look at the score here (an Intel 12C/24T 2 socket system)....

2 x 2x Intel(R) Xeon(R) CPU X5650 @ 2.67GHz

((4.727GB/s) / (12 cores)) / 2.86GHz = 0.138

"Per-core" the performance still a little higher. If calculated as "per thread" then its a different story.

Why not an Intel 4S Xeon 7560 2267mhz , that is 32C/64T ,
score is 4.506 GB/s.....

(4.506 GB/s)/32/2.267 = 0.062....

On a per thread perfs, that would be 0.031....

http://www.sisoftware.co.uk/rank201...8feed3e0d2e5d1f785b888aecbae93a385f6cbfb&l=en

 

[BlueBlazer]

We have a good idea how these two terms differ for Intel's use of them, but we remain to be enlightened as to how this differs for AMD.

It does become convoluted a bit. Seemingly comparing cores no longer accurate (starting from Bulldozer).

Why not an Intel 4S Xeon 7560 2267mhz , that is 32C/64T ,
score is 4.506 GB/s.....

(4.506 GB/s)/32/2.267 = 0.062....

First off.... Its a 4-socket system, thus some performance penalty is expected (just as I've mentioned earlier here). Second, the Xeon 7560 does not feature AES-NI (reference here) while Xeon X5650 does have AES-NI (reference here). Starting from Thuban to Bulldozer, AMD implemented hardware AES instructions.

 

[Abwx]

First off.... Its a 4-socket system, thus some performance penalty is expected (just as I've mentioned earlier here).

4S doesnt bring inherent penalty provided there s enough bandwith.

Indeed, this "penalty" as you call it is just part of the system
global efficency, wich tell that a 4S 32C/64T X7560 is not
competitive since it s priced ridiculously high-ish.

Second, the Xeon 7560 does not feature AES-NI (reference here) while Xeon X5650 does have AES-NI (reference here). Starting from Thuban to Bulldozer, AMD implemented hardware AES instructions.

If you did read correctly, rather than trying to boost your ego ,
you should have noticed that i specified in a post above that
it was likely that the 7560 lacks the relevant instruction...

Indeed, as pointed by Idontcare , the exemple you re using
is just pointless since the frequency is almost double as the
BD server, and you perfectly know that frequency scale
better than cores for the discussed benchmark, so it s not
only likely, it s completely sure that at equal frequency
the Xeon5650 would be far behind in a theorical
4S variant or with a 2S using a Cpu that would have double
the core count..

 

[Vesku]

Still no server systems released? Does this mean they needed a stepping jump for even the low clocked server chips?

 

[BlueBlazer]

4S doesnt bring inherent penalty provided there s enough bandwith.

Not only bandwidth, but also higher latency and cache coherency traffic.

Indeed, this "penalty" as you call it is just part of the system
global efficency, wich tell that a 4S 32C/64T X7560 is not
competitive since it s priced ridiculously high-ish.

4 socket systems are rare, and usually made for special markets (usually high end database, virtualization, etc). Thus their costs are relatively high.

If you did read correctly, rather than trying to boost your ego ,
you should have noticed that i specified in a post above that
it was likely that the 7560 lacks the relevant instruction...

And yet you went ahead with this.....

Why not an Intel 4S Xeon 7560 2267mhz , that is 32C/64T ,
score is 4.506 GB/s.....

I rest my case.

Indeed, as pointed by Idontcare , the exemple you re using
is just pointless since the frequency is almost double as the
BD server, and you perfectly know that frequency scale
better than cores for the discussed benchmark, so it s not
only likely, it s completely sure that at equal frequency
the Xeon5650 would be far behind in a theorical
4S variant or with a 2S using a Cpu that would have double
the core count..

Did you even checked? My example is 2.86GHz. Tell me, how is that "almost double" of 2.356GHz?

 

[Abwx]

Did you even checked? My example is 2.86GHz. Tell me, how is that "almost double" of 2.356GHz?

Did confuse with a post above, yet , the BD server is still 50%
faster than the wonder you re talking about that is clocked more than 20% higher...

do the maths....

 

[BlueBlazer]

Did confuse with a post above, yet , the BD server is still 50%
faster than the wonder you re talking about that is clocked more than 20% higher...

do the maths....

Already done my maths....

A single Gulftown core > A single BD "core"

Which is why I mentioned this earlier....

It does become convoluted a bit. Seemingly comparing cores no longer accurate (starting from Bulldozer).

 

[3DVagabond]

Already done my maths....

A single Gulftown core > A single BD "core"

Which is why I mentioned this earlier....

This seems like, "Let's find a metric that Intel is better on and critique performance only on that". Intel has > perf/core, they are better, end of story. Who cares how the performance is attained, as long as it performs?

 

[Abwx]

Already done my maths....

A single Gulftown core > A single BD "core"

Done , but innaccurately...

Indeed, you re right about a gulftown core being better than a BD core,
but the interessant question is : by how much ?...

Let s use the exemple we re talking about.

The numbers show that at equal frequencies, the 24C/24T Bulldozer
would be 80% faster than the 12C/24T X5650 based server.

So we can conclude that a gulftown core has at best 12% better
perf than a BD core, and that s not counting that BD has double the
core count and has to give up some efficency due to cores scaling,
as it is unlikely that the xeon would scale more than 80% if its core
count would be doubled.

Beside, each BD core being fully used, there s some ressource
that are shared for each core pair, so it s likely that single thread
performance would be even closer of gulftown one..

All in all, this particular bench look promising as it validates AMD s
approach of clustered multithreading.

 

[BlueBlazer]

Indeed, you re right about a gulftown core being better than a BD core,
but the interessant question is : by how much ?...

Let s use the exemple we re talking about.

The numbers show that at equal frequencies, the 24C/24T Bulldozer
would be 80% faster than the 12C/24T X5650 based server.

So we can conclude that a gulftown core has at best 12% better
perf than a BD core, and that s not counting that BD has double the
core count and has to give up some efficency due to cores scaling,
as it is unlikely that the xeon would scale more than 80% if its core
count would be doubled.

Beside, each BD core being fully used, there s some ressource
that are shared for each core pair, so it s likely that single thread
performance would be even closer of gulftown one..

All in all, this particular bench look promising as it validates AMD s
approach of clustered multithreading.

Basing the performance on this particular benchmark (hardware accelerated encryption) to generalize BD "core" performance is highly inaccurate. And from your own link

There are other numbers of this ES in the same site:

http://www.sisoftware.co.uk/rank201...a69bae88e0dde8ceb68bba9cf99ca191b7c4f9c9&l=en

According to the arithmetic bench, it do 3.18 operations/cycle/core.

The arithmetic bench shows otherwise (example comparison here). It shows the a single Gulftown/Westmere thread slightly faster than a single BD "core", or a BD "core" almost as fast an Gulftown/Westmere thread. If cores compared then it would be totally awkward (Gulftown/Westmere core performance would be at least "twice" faster than BD "core"!). This is right in line with David Kanter's analysis at RWT here....

The most striking thing about the results is the sheer variation. A Bulldozer core is anywhere from 0.6X to 1.3X the performance of Istanbul.

BD "cores" performs somehwat more like HyperThreading (performance seems to vary from applications to applications). That's why its looks seemingly convoluted (with AMD opting to call them "cores")

 

[Imouto]

Maybe only daydreaming here because it isn't out yet but BD "module" in a light threaded workload behaves like a single hyperthreaded Intel core.

Intel: Bump 2 threads into a single core for complete computing power optimization.
AMD: Bump 1 thread into a module splitting load into 2 cores. Obviously this is available only with at least one core idling.

At least that's what I understood taking in mind all that schedule features and shared components.

Next thing you are going to tell me that water is wet and fire is hot!?

The sky is blue, water is wet, women have secrets.

 

[Dresdenboy]

@BlueBlazer:
David Kanter did his Analysis 5 monts ago based on a set of benchmark results of one single ES system, which was not fully functioning in some aspects. I think at this point in time it's better to look at more recent results or wait for official numbers for discussing them.

And then there are simple rules (as usually with some exceptions) for SMT/CMT-performance.

For one SMT-capable core supporting 2 threads:
1. Performance/core with 1 thread is lower than performance/core with 2 threads. (e.g.: 1.0 vs. 1.2)
2. Performance/thread with 1 thread is higher than performance/thread with 2 threads. (e.g.: 1.0 vs. 0.6)

For one CMT-capable module containing 2 cores:
1. Performance/module with 1 thread is lower than performance/module with 2 threads. (e.g.: 1.0 vs. 1.8)
2. Performance/core with 1 thread is higher than performance/core with 2 threads. (e.g.: 1.0 vs. 0.9)

 

[poohbear]

These BD cpus are supposed to ship next week and not ONE review? why's AMD so tightlipped about em? Not a good sign when's the NDA finally over?

 

[Imouto]

NDA date is under NDA too?

 

[nenforcer]

These BD cpus are supposed to ship next week and not ONE review? why's AMD so tightlipped about em? Not a good sign when's the NDA finally over?

They are shipping to their partners who will then begin assembling and certifying systems for shipment in mid / late September.

I actually think most people will be impressed with BD and say Intel at least has some competition now at the mid and high end. The chip was supposed to release in June and AMD delayed probably because they knew it wasn't exactly up to par.

They know they can't duplicate the Phenom 1 launch and have to impress when they finally do show it.

 

[AtenRa]

@BlueBlazer:
David Kanter did his Analysis 5 monts ago based on a set of benchmark results of one single ES system, which was not fully functioning in some aspects. I think at this point in time it's better to look at more recent results or wait for official numbers for discussing them.

And then there are simple rules (as usually with some exceptions) for SMT/CMT-performance.

For one SMT-capable core supporting 2 threads:
1. Performance/core with 1 thread is lower than performance/core with 2 threads. (e.g.: 1.0 vs. 1.2)
2. Performance/thread with 1 thread is higher than performance/thread with 2 threads. (e.g.: 1.0 vs. 0.6)

For one CMT-capable module containing 2 cores:
1. Performance/module with 1 thread is lower than performance/module with 2 threads. (e.g.: 1.0 vs. 1.8)
2. Performance/core with 1 thread is higher than performance/core with 2 threads. (e.g.: 1.0 vs. 0.9)

You should enjoy the Cretan sun and the Women and stop posting about BD :hmm:

VAI beach

Bulldozer Engineering Samples and an Analysis of Orochi's Die Size

 

[Rhezuss]

Been bleeing from all my orifices above the belly button reading this page...

Is BD due for september, october or later?

Anyways, eager to see whats under the hood for this BD and hoping to put it in my rig as my next upgrade.

Can't wait to see the prices and model range for this too.

 

[formulav8]

Starting from Thuban to Bulldozer, AMD implemented hardware AES instructions.

I wasn't aware that AMD added AES to Thuban?