AMD sheds light on Bulldozer, Bobcat, desktop, laptop plans

[jvroig]

If you say (as in your original statement) "quad-threaded app", then it is assumed the app you are talking about can make very productive use of all four cores. If you then cite a benchmark or game that doesn't quite make very productive use of all four cores, then it doesn't quite come off as consistent to your own statement.

Anyway, regarding that particular scenario of scaling only ~30%... well if that's the scenario, then Bulldozer can realistically beat the Athlon II X4 because of its touted stronger integer performance, and because the app in question does not benefit as much from a quad-core processor.

The incredulity that arose from your original statement is a result of you being unclear with what you wanted. Everybody must have thought you meant that Bulldozer should beat an Athlon II X4 in a heavily quad-threaded app (meaning makes full use of all available cores) such as video encoding, therefore demanding that Bulldozer be more than 2x faster clock for clock than an Athlon II core.

 

[GaiaHunter]

If you say (as in your original statement) "quad-threaded app", then it is assumed the app you are talking about can make very productive use of all four cores. If you then cite a benchmark or game that doesn't quite make very productive use of all four cores, then it doesn't quite come off as consistent to your own statement.

I agree - after all a E8400 was barely slower in GTA IV compared to a Q6600.

Computer Bottleneck, a Dragon Age comparison would be much more relevant - if you can find one.

 

[cbn]

I agree - after all a E8400 was barely slower in GTA IV compared to a Q6600.

FYI Core i3 is ~25% faster than E8400 in GTA IV (see page 10 of that link for the graph).

If you say (as in your original statement) "quad-threaded app", then it is assumed the app you are talking about can make very productive use of all four cores.

Well you are right with what you are saying. GTA IV doesn't count as full quad optimization so it was a bad example on my part.

But take a look at this test:

http://www.pconline.com.cn/images/h...926714_bio5.jpg&namecode=diy&subnamecode=home

In this benchmark Athlon II x4 is beating Phenom II x3 despite having less clock speed and less cache. So I wonder if this test is more quad optimized?

But look at how Core i3 does in comparison to E8400 and the other CPUs?

If Intel hyperthreading scales like this more often in games I really hope for AMD's sake their Bulldozer modules are much more powerful.

 

[Kuzi]

It has been mentioned that individual integer cores in Bulldozer will be much more efficient than Phenom II. I've read recently somewhere (probably AnandTech, either from Johan or Anand, will look) that integer performance in Bulldozer will be about 10-35% better than Phenom II.

Ok, I just read Anand's article released today, and it gave some good info about Bulldozer. First is the way AMD will count cores, an 8-core Bulldozer (4-module) will support a max of 8 threads. This means when comparing with Intel, we should compare that with a 4-core CPU from Intel, because Intel parts support HT, so their quad-cores can run a max of 8 threads.

And as you mentioned, a 4-core Zambezi's (2-module) integer performance will be between 10-35% faster than a Phenom II X4. A 10% gain would be low as that will only allow Zambezi to reach Core2 integer performance. At 15-20% gain, it can be comparable to Nehalem, roughly speaking. But what is really interesting about this, is that each "core" in Zambezi only adds 5% to the die size. Knowing this, and the fact that Zambezi will be a 32nm part, my guess is that @ 4-cores it can easily be under 200mm2 in size, even if it had 8MB L3 cache. A smaller die size would allow higher clocks and lower TDPs, which AMD could really use in order to compete with Intel.

 

[Kuzi]

I've just checked and it seems Intel's soon to be released i9 Gulftown 6-core CPU, should be around ~240mm2 in size. That is smaller than i7 (263mm2) and Phenom II (258mm2), even though the i9 has 6 cores and more cache.

Now it seems "possible" to me that a 4-module/8-core Zambezi, can compete well against a 6-core (12 thread) i9. Although the die size of an 8-core Zambezi may end up larger than i9, depending on cache sizes of course.

The only problem with all this is that Zambezi may only release a year or more after i9, and probably half a year after Sandybridge. We also have no idea how much better will Sandybridge be compared to i9, and I don't believe Zambezi will have any chance competing against Sandybridge, but we can always hope.

 

[JFAMD]

Ok, I just read Anand's article released today, and it gave some good info about Bulldozer. First is the way AMD will count cores, an 8-core Bulldozer (4-module) will support a max of 8 threads. This means when comparing with Intel, we should compare that with a 4-core CPU from Intel, because Intel parts support HT, so their quad-cores can run a max of 8 threads.

For some reason I just don't see people comparing a 4 core Intel product to an 8-core AMD product and calling it a fair comparison because they have the same number of threads.

Let's say, for fun, that the IPC performance is exactly the same. If HT gives ~20% advantage for two threads over a single core and Bulldozer gives ~80% for two threads over a single core, then the math looks something like this:

Intel - 4 cores + HT = (4 x 1.2) = 4.8 cores

AMD - 8 cores = (4 x 1.8) = 7.2 cores

So, that means, in terms of throughput, that even if Intel has ~50% IPC advantage, they should should still be even.

This of course can be thrown off even more if my numbers are conservative (most of our performance statements are vetted through a pretty conservative group of lawyers before they are made) and if my HT numbers are high. Remember that there are places where HT actually provides negative performance gains (anand is giving ~7-10% for integer/FP performance).

I just have to think that nobody is going to compare an 8-core AMD to a 4-core Intel.

 

[cbn]

I think each module of Bulldozer (dual core) needs to be at least as good as Intel dual core w/hyperthreading at running quad optimized tasks.

 

[jvroig]

I sympathize with you, JFAMD.

This situation we are in happened because people - wrongfully or rightfully - started giving a crap about "cores" vs "threads", when in fact nobody should care. Of course, that the companies marketed it that way ("4 cores!" or "Dual core. w/ HT!" instead of just thread-crunching power, but that's a complicated topic of its own) also doesn't help, and is actually the reason people started giving a crap.

I don't know where to lay the blame, but playing the blame game has often been said to be futile, so I won't even try.

So it's just a matter of terminology, for better or worse. Intel says "quad-core", so consumers will likewise compare it with a "quad-core" from AMD, so the hypothetical quad-core Bulldozer will end up fighting with a "quad-core" Nehalem or Westmere or Sandy Bridge or whatever (I've lost track). And of course, if the Intel competitor has HT, then Bulldozer may lose badly, and AMD can rightfully cry foul, since the thread count is way off, and the proper comparison should have been "threads" instead of "cores". The problem with that is Intel will market their processors saying "cores", and so will AMD. Given that, I can't imagine what other conclusion consumers will take from it.

The good news is this: actually, consumers care less about the "core wars"; the real comparison that counts is the price point. If a fair fight is a quad-core Bulldozer versus a dual-core Nehalem/Westmere/SandyBridge/Whatever, then price it accordingly. Don't blame anybody, tech-oriented consumer or not, for comparing a quad-threaded Bulldozer to an octo-threaded Nehalem/Westmere/SandyBridge/Whatever if the aforementioned Bulldozer is occupying the same price point.

Of course, that will happen only when Bulldozer is already available for retail. Until then, Mr. Freuhe, I sympathize with you, but you and I and everybody else plus their dog and their dog's puppies as well will have to endure "unfair" comparisons between quad-threaded AMD processors to octo-threaded Intel counterparts.

 

[GaiaHunter]

FYI Core i3 is ~25% faster than E8400 in GTA IV (see page 10 of that link for the graph).

But that can be related to cache + speed.

See how on this review here the Q9550 is 25% faster than the E8400? But then the i7 920 is barely faster than the Q9550?
http://www.tomshardware.co.uk/build-balanced-platform,review-31726-11.html

But take a look at this test:

http://www.pconline.com.cn/images/h...926714_bio5.jpg&namecode=diy&subnamecode=home

In this benchmark Athlon II x4 is beating Phenom II x3 despite having less clock speed and less cache. So I wonder if this test is more quad optimized?

But look at how Core i3 does in comparison to E8400 and the other CPUs?

If Intel hyperthreading scales like this more often in games I really hope for AMD's sake their Bulldozer modules are much more powerful.

That test is Resident Evil 5, if I'm not mistaken.

Take a look in here:
http://www.pcgameshardware.com/aid,...nom-strong-Update-Lynnfield-results/Practice/

That game just loves the Core i7 architecture.

Sorry if I seem a bit decided to bash your expectation - don't get me wrong, I will be building a new machine soon. I was thinking of an Athlon II x4, but if the core i3 is faster and around the same price, I will get one instead.

 

[GaiaHunter]

Well if, as JF said, 4 Bulldozer modules are about 60 to 80% faster than one six-core Opteron 6100 CPU in SPECInt_rate, it sounds quite good.

We would expect 16 cores to beat 12 cores by 33%, so there is some improvements all around on the cores.

Now, it also depends on what are the speeds we are comparing - if those numbers are at the same clock and if Bulldozer will be able to reach decent speeds.

Additionally we will have to see how much die size can AMD save compared to its current offers.

If 2 Bulldozer modules that are the equivalent or faster to a Deneb X4 and are quite smaller, then AMD seems to be in a good shape to at least price their CPU aggressively if they can't match Intel offers at the time.

I said it before - this situation seems to be the CPU equivalent of the RV770 vs GT200 fight.

If AMD can stay (at least) in shouting range of Intel in performance but are a fair bit cheaper, they will have a winner.

 

[jvroig]

I said it before - this situation seems to be the CPU equivalent of the RV770 vs GT200 fight.

That would be great for AMD if that were the case. Wasn't the RV770 a huge win for ATi?

Then again, even after that, nVidia still held on to its market-leader position. I guess those two aren't mutually exclusive, anyway. It can still be a huge win for ATi without necessarily taking the market-leader position, as long as it resulted in greater profit and perhaps a good increase in market share.

I hope you are right, Gaia, as that would really be a breath of fresh air for AMD.

 

[ShawnD1]

AMD are way behind Intel with their CPUs mainly because:
*No technology comparable to HT (running more threads per core)

Why would you even want this? We've seen time and time again that hyperthreading often cripples performance because software is confused by it. If a program tries to grab 3 logical cores, it only picks up 2 physical cores. In one test done by Xbit, enabling HT dropped performance by as much as 17%
http://www.xbitlabs.com/articles/cpu/display/core-i7-975-950_7.html#sect0
(scroll to the bottom)

 

[Kuzi]

Why would you even want this? We've seen time and time again that hyperthreading often cripples performance because software is confused by it. If a program tries to grab 3 logical cores, it only picks up 2 physical cores. In one test done by Xbit, enabling HT dropped performance by as much as 17%
http://www.xbitlabs.com/articles/cpu/display/core-i7-975-950_7.html#sect0
(scroll to the bottom)

What I meant by that is the ability for each core to run more than one thread, simultaneous multi-threading, not necessarily Intel's HT method. AMD is taking a different approach for Bulldozer and throwing more transistors to support SMT, by adding an extra Integer Cluster, so there should be no performance loss at all.

 

[Idontcare]

I just have to think that nobody is going to compare an 8-core AMD to a 4-core Intel.

We already compare 6-core opterons to 4-core xeons.

The comparison will be made if the products are remotely close in performance and/or price.

Looking at Intel's upcoming Clarkdale 2C/4T processors they are clearly going to be compared to and evaluated against 4C/4T Athlon II X4 SKU's.

 

[jvroig]

AMD is taking a different approach for Bulldozer and throwing more transistors to support SMT, by adding an extra Integer Cluster, so there should be no performance loss at all.

I think I know what you mean, but just in case... the "extra Integer Cluster" is not AMD's answer to HT. In fact, the "extra Integer Cluster" is not extra at all, in the sense of the word "extra" that we understand (i.e., a bonus, something freely added to the package we bought). That "extra" you refer to is actually a core in itself.

So a quad-core Bulldozer has four cores (four Integer Clusters), and can process a maximum of four threads. Not eight. Intel's quad-core has four cores but can process eight threads if it supports HT.

So AMD is not coming up with its own HT. They are just increasing cores in a very efficient way.

 

[Kuzi]

For some reason I just don't see people comparing a 4 core Intel product to an 8-core AMD product and calling it a fair comparison because they have the same number of threads.

As jvroig correctly mentioned, it all comes down to price/performance. Lets say Intel has an 8-core (16 thread) CPU that costs $500, and a 4-core (8 thread) part for $220. And AMD's 8-core (8 thread) CPU costs $250, would you compare the AMD CPU with Intel's $500 part or $220 part?

 

[GaiaHunter]

As jvroig correctly mentioned, it all comes down to price/performance. Lets say Intel has an 8-core (16 thread) CPU that costs $500, and a 4-core (8 thread) part for $220. And AMD's 8-core (8 thread) CPU costs $250, would you compare the AMD CPU with Intel's $500 part or $220 part?

That depends on the performance.

I would say that AMD was being successful (at least for desktop where power consumption is somewhat less important) if their 8 core performed slightly worse than the 8c/16t Intel CPU but was much faster than the 4core/8thread Intel part.

I think that is what AMD is trying to achieve (well maybe what they are trying to achieve being overall faster, and this is just me talking as if there is a rule saying Intel will be faster).

Now if AMD 8c/8t is only slightly faster or is even slower than the 4c/8t Intel part, they will be in trouble.

Another factor we will need is die size.

 

[Kuzi]

I think I know what you mean, but just in case... the "extra Integer Cluster" is not AMD's answer to HT. In fact, the "extra Integer Cluster" is not extra at all, in the sense of the word "extra" that we understand (i.e., a bonus, something freely added to the package we bought). That "extra" you refer to is actually a core in itself.

So a quad-core Bulldozer has four cores (four Integer Clusters), and can process a maximum of four threads. Not eight. Intel's quad-core has four cores but can process eight threads if it supports HT.

So AMD is not coming up with its own HT. They are just increasing cores in a very efficient way.

Yes, and I personally believe this approach to be much better than Intel's HT, and will provide higher performance gains (and no decrease) for multi-threaded apps in comparison. Hopefully the IPC per core for Bulldozer can really be up to 30-35% faster than Phenom II, because a 10% increase for a 2011 CPU will be way too low.

 

[Kuzi]

That depends on the performance.

Exactly

And we all know AMD will price Zambezi or any of their products according to the price/performance of the competition.

Now if AMD 8c/8t is only slightly faster or is even slower than the 4c/8t Intel part, they will be in trouble.

You see, even the performance can change for certain software or situations. Lets say IPC per core is a little slower on the AMD CPU, and we are using software with 4 or less threads, then the 4c Intel part would be faster. But if we take another situation for a highly threaded program, that supports 8 threads or more, the AMD CPU can be much faster.

 

[jvroig]

Forget the performance. Those responsible for 90+% or more of Intel and AMD sales don't care about performance as an isolated factor. They care about price.

If two processors are in the same price point, they'll get compared. Doesn't matter if it's an 8-core AMD versus a 2-core/4-thread Intel. If that's where the price leads us, that's going to be the comparison.

No one, and absolutely no one, in their right mind will compare a $300 processor to a $500 one and expect the $300 processor to win.

 

[ilkhan]

I think each module of Bulldozer (dual core) needs to be at least as good as Intel dual core w/hyperthreading at running quad optimized tasks.

for comparison, 1 module = 1 [intel] core. Both run 2 threads. They're positioning 1 module parts against duals, 2 module parts against quads, and 4 module parts against octos. Hex is pretty much a one time thing that won't help anybody.

 

[piesquared]

It is rumored that AMD will use some kind of clock gating technique to further boost single threaded performance, so maybe they are clocking up cores that are being used, and disabling those that aren't.

 

[piesquared]

.......If Intel hyperthreading scales like this more often in games I really hope for AMD's sake their Bulldozer modules are much more powerful.

That's the problem though; it doesn't.

 

[GaiaHunter]

It is rumored that AMD will use some kind of clock gating technique to further boost single threaded performance, so maybe they are clocking up cores that are being used, and disabling those that aren't.

Yes, one of the things it seems AMD will be introducing is their own "Turbo Boost".

 

[piesquared]

Yes, one of the things it seems AMD will be introducing is their own "Turbo Boost".

Yeah, one of the parts to the rumour was that it would work different than turbo boost.