Some Bulldozer and Bobcat articles have sprung up

[jvroig]

Single thread performance and multi-thread performance (or rather "throughput") are very different as JFAMD analogized to traffic, especially given a specific product. A single-thread workload on a core would not need to contend for bandwidth or L3 with other cores, for example, if other cores are idle.

Absolutely. Actually, if you were able to read the whole thing before I edited it out (JFAMD quoted only the first half), I actually "predicted" (not very seriously) >40% single-thread performance increase. Anyway, it's good to have been edited out. I wouldn't want to be blamed for spreading the word that BD will offer 40% improvement in single-thread scenarios over Phenom II, in case that never materializes, and I'm not really a betting man.

 

[bryanW1995]

That's because there's lots of competition in the $200 and under space and zero competition above it. AMD has nothing on a 970 and a 980, thus the price for each. Why are you confused that Intel doesn't price a reedy little dual core similarly to a monster 6 core?

I'm confused because they could make more money if they sold them for $500-$600 and sold 3x as many. instead, they're making cheaper parts that compete vs amd's dirt cheap offerings. do you understand how it makes more sense to keep in a market that you own instead of building more parts to sell on the cheap? maybe their yields aren't that great on 32nm yet so they can just make a LOT more dual cores for now.

 

[Tsavo]

I'm confused because they could make more money if they sold them for $500-$600 and sold 3x as many. instead, they're making cheaper parts that compete vs amd's dirt cheap offerings. do you understand how it makes more sense to keep in a market that you own instead of building more parts to sell on the cheap? maybe their yields aren't that great on 32nm yet so they can just make a LOT more dual cores for now.

I wouldn't be surprised to learn that the market for $500-$600 CPUs is very similar to the market for $900-$1000 CPUs. Lots and lots of tire kickers, but not many people down at the DMV picking up the plates for their shiny new toy.

 

[Scali]

that was more of a shot at your incessant trolling in this thread than a comment about the technical background that amd has published on their upcoming architectures. just because you are qualified to discuss a subject doesn't mean that you can do so in a friendly, non-confrontational way. You've been rarely friendly and/or non-confrontational in this thread, so in spite of your obvious technical expertise you're still trolling.

Guys like you are being confrontational with me all the time, what do you expect?
Perhaps you should look at yourself first, and see if you can try to contribute to a more friendly atmosphere.

 

[Scali]

Um..what?

IPC, single-threaded performance... So don't compare dualcores vs quadcores in multithreaded tests...

 

[jvroig]

IPC, single-threaded performance... So don't compare dualcores vs quadcores in multithreaded tests...

IPC did improve, but clockspeeds went down, so net result was single-threaded performance was hardly desirable, making Barcelona not quite attractive.

Coming from an Athlon X2 5000+, I just got a cheap 7750 (instead of a 9650; the 7750 was just a BIOS update and spare change, it was a no-brainer, unlike even considering a Barcelona) because I needed single-thread performance a lot of the time. From the 7750, I then reused all my RAM (4GB) to get a cheap mobo + Phenom II X4 965, and now I get the best of both worlds (single-thread when needed, multiple threads in other cases). Barcelona was just not worth investing in (this is hardly news, though)

http://www.anandtech.com/bench/Product/34?vs=29
That's a clock for clock comparison as I could make it, and while a few single-thread / dual-thread performance results do verify that there was a clock-for-clock increase (Cinebench single-thread, games), in the long run it doesn't matter. In fact, taking a look at the SKUs I compared, I compared a 4800+ (nowhere near the top Athlon 64 X2) vs a Phenom 9850 (second to the top Barcelona). It just gets worse from there, the top end 9950 gets hammered by the top end Athlon 64 X2 64000+ in single thread performance

That's the problem. They can quote "IPC improved" all they want, but that's all just a part of getting the final single-thread performance, and it's in the final result that Barcelona utterly failed. Barcelona improved IPC, but clockspeeds went down. I wouldn't be surprised if some Barcelona upgraders (coming from high-end Athlon 64 X2's) actually got poorer performance from predominantly single-threaded apps, or apps that only hardly make use of a second thread.

 

[Scali]

IPC did improve

Yes, that's what I said... I estimated about 5% off the top of my head.
But let's take your single-threaded Cinebench....
We have:
K8: 2262
K10: 2536

So we have an improvement of 2536/2262 = 1.12.
Okay, so it was 12% then, not 5%... Still it doesn't change my story (and this is a very favourable test for K10, as it makes full use of the new 128-bit SIMD units, where K8 still had 64-bit ones. In a purely integer or x87 test, it will probably be closer to the 5% I mentioned).

 

[jvroig]

Yes, I wasn't arguing with you. I was corroborating what you said, making a distinction between IPC and single-threaded performance to show why Barcelona failed, since a lot of people here seem to have mixed up the two as interchangeable and talk about IPC as one and the same with single-threaded performance.

 

[Scali]

That's true... as I say, I heard that BD will run at around 2.75 GHz... In which case IPC needs to improve to keep up with the current 3+ GHz CPUs in single-threaded performance.

 

[IntelUser2000]

That's true... as I say, I heard that BD will run at around 2.75 GHz... In which case IPC needs to improve to keep up with the current 3+ GHz CPUs in single-threaded performance.

Not if that's a Interlagos clock.

 

[Triskain]

IPC, single-threaded performance... So don't compare dualcores vs quadcores in multithreaded tests...

Here we have an IPC boost of 15 % on average and here the boost is 18 % comparing the A64 X2 6000+ and Athlon II X2 250.

All CPU's mentioned here are Dualcores, so it's a fair comparision.

 

[Scali]

Here we have an IPC boost of 15 % on average and here the boost is 18 % comparing the A64 X2 6000+ and Athlon II X2 250.

All CPU's mentioned here are Dualcores, so it's a fair comparision.

Well, my point still stands: it falls well short of 50%. It's not even 1/3rd on average.

 

[DrMrLordX]

Can we please not turn another Bulldozer thread into a K8 vs K10 debate? Been there, done that, got the T-shirt. I don't blame CTho for weighing at all, mind you, but that should have been the final word (or if not that, then Triskain's remarks should have).

 

[Scali]

Not my fault that people want to get into a K8 vs K10 debate.
My point is just that we've rarely seen a 50% increase in IPC (or anything even close, K8 vs K10 certainly wasn't it), so we shouldn't get our hopes up for BD pulling that one off.

 

[AtenRa]

You people are obsessed with single-thread performance in an era of Multi-thread and Multi-core CPUs and Software.

 

[Scali]

You people are obsessed with single-thread performance in an era of Multi-thread and Multi-core CPUs and Software.

No we're not, we're just trying to determine what the weak and strong points of BD are.
I think you people are overly touchy about any kind of weaknesses that AMD's products may have. Grow up.
It's getting really REALLY annoying that people keep 'chiming in' with these personal remarks. Someone should do something about that.

 

[Triskain]

This IPC discussion is grating on my nerves, it's time for some factual analysis, instead of conjecture:

Datapoint 1: Yorkfield has an average 7 % IPC advantage over Deneb
proofed here by comparing the Q9650 and the PhII X4 945, both at 3 GHz

Datapoint 2: Nehalem has an average 7 % IPC advantage over Yorkfield (without Turbo or HT)
proofed here by comparing i7-965 and the QX9770

Datapoint 3: Sandy Bridge has an average 12 % IPC advantage over Nehalem
calculated from the leaked benches and the Anandtech preview

(not as exact as the other two Datapoints, because the existing comparisons are suboptimal)

Ergo, to match Sandy Bridge's IPC Bulldozer would need an approximate IPC boost of 30 % over Deneb (1.07 * 1.08 * 1.12 ~ 1.30) and that is indeed an unrealistically high jump.

However, with Barcelona AMD achieved an IPC boost of 15-18 % with relatively few major enhancements to the architecture and considering that the core/module architecture gets a huge overhaul in Bulldozer (as can be seen in this article by David Kanter), an IPC boost of 15-20 % is not an unrealistic proposition for Bulldozer.

A boost of at least 12,8 % is already confirmed by the "50 % over Magny-Cours" statement, however, because of the shared architecure the behavior in single thread situations will be different, with a greater boost to be expected.

So while it is very probable that BD won't match Sandy on IPC, an IPC for BD on the level of Nehalem and maybe a bit above is not a totally unreasonable expectation.

Statement: Should the actual IPC of BD "completely suck" [sic], I give you alle every right to knock AMD for it, for a duration of 5 years after this Statement evaluetes to TRUE.

 

[lopri]

No we're not, we're just trying to determine what the weak and strong points of BD are.
I think you people are overly touchy about any kind of weaknesses that AMD's products may have. Grow up.
It's getting really REALLY annoying that people keep 'chiming in' with these personal remarks. Someone should do something about that.

You mean, remarks like this one. (Thought it was funny you could post these two almost at the same time without any sense of contradiction)

http://forums.anandtech.com/showpost.php?p=30376499&postcount=139

I personally do something about things like this. (Hint for you: Open UserCP and edit "ignore list")

Edit: Oh and instead of being touchy feely and crying,

Someone should do something about that.

Maybe you should heed to your own advice.

Grow up.

 

[Scali]

You mean, remarks like this one. (Thought it was funny you could post these two almost at the same time without any sense of contradiction)

Nope, I'm just making an observation about human nature (in a response and agreement to the post above).
I'm not attacking anyone personally. I don't single people out, I don't judge their behaviour, I don't derail or troll. I don't say people are 'obsessed' when I don't agree with them, or when they talk about a subject I'm not interested in. I just avoid the subject/thread altogether.

 

[AtenRa]

First of all, lets be clear that I don’t support AMD over Intel or the opposite, I like both companies and there products.

Now that we clarified that,

I think you people are overly touchy about any kind of weaknesses that AMD's products may have.

Yes we should try and see what’s the strong or weak points in BD architecture but we also have to see the bigger picture, the whole CPU part and not only one modules weaknesses or strong points.

 

[Scali]

Yes we should try and see what’s the strong or weak points in BD architecture but we also have to see the bigger picture, the whole CPU part and not only one modules weaknesses or strong points.

Read the thread more closely... the benefits of the module design have been discussed quite a lot as well (arguments such as better performance per watt, more cores per die etc).

 

[CTho9305]

Well, my point still stands: it falls well short of 50%. It's not even 1/3rd on average.

So, the differnce between 5 and 12% when talking about BD's area is worth flame wars, but not when talking about K8's IPC vs its successor? That difference is only relevant at the "<30%" level of accuracy?

 

[Scali]

So, the differnce between 5 and 12&#37; when talking about BD's area is worth flame wars, but not when talking about K8's IPC vs its successor? That difference is only relevant at the "<30%" level of accuracy?

The exact amount of IPC improvement is not relevant to the point I was making.
Do you understand the point I was making?
I said that BD would require an improvement of about 50% in efficiency in order to maintain the same IPC.
I then pointed out that CPUs rarely make a jump of 50% in efficiency.
So whether K8 vs K10 is 5%, 10% or even 20%, it's not relevant. All these figures are a LONG way from the 50% I was arguing, hence it's not relevant to the point.
Try to keep up, or else just leave the thread. People like you make me have to reiterate the same thing over and over again, while it was already clear from the start to those who bother to pay attention. Don't let it happen again.

That last comment got you an infraction. We don't tolerate personal insults here.
Markfw900
Anandtech Moderator

 

[Tsavo]

Try to keep up, or else just leave the thread. People like you make me have to reiterate the same thing over and over again, while it was already clear from the start to those who bother to pay attention. Don't let it happen again.

You've brought nothing positive to this thread specifically and to this forum generally.

:thumbsdown:

And this one also, Same as above, personal attacks are not tolerated
Markfw900
Anandtech Moderator

 

[Scali]

You've brought nothing positive to this thread specifically and to this forum generally.

:thumbsdown:

Many will beg to differ with you on that matter.