Rumour: Bulldozer 50% Faster than Core i7 and Phenom II.

[hamunaptra]

I think the 50% increase in performance with 33% more cores is given the same clock speed as a thuban. Which is easily the case. But you ramp up the clocks and you get a result very similar to the one in the graph.

The reason I believe that is at the same clock speed is because, if it werent (ie it was the equation of speed increase over thuban including clock speed factory) then the BD would literally suck be horrible at IPC.
Heres my math, I dont know if is right or what but....according to the above 50% more performance which is essentially .5 faster with a 33% more core count so add 2 cores.
Using these measurements:
150% = 8 cores is a ratio of 18.75 per core
100% = 6 cores is a ratio of 16.66 per core

If you figure the percentages of what those 2 numbers are you basically get 12% increase in IPC per core over thuban.

Thats where I see how you got the what should be 9000 based number by just multiplying 6100 by 150%. So, that would be purely clock for clock based.

I would also gather that in order for a thuban to be equivilant to BD clock for clock, going with the 50%/33%
thuban would basically need a total of 9cores (to be 50% faster than what it is) compete with a 8 core BD once again 8 is 88% of 9 which equates to a 12% increase per core over thuban once again.

So, Im guessing a clock speed of 4-4.5ghz for BD according to this bench plus some code optimization to squeeze more performance out of it.

According to that graph, as we can tell BD gets 17000, thuban gets 6200. Which equates to a 48% increase in performance per core over thuban according to that graph.
Seeing as I calculated the 12% above clock for clock. Subtract the IPC increase 12% from the 48% increase and we are left with... 36% increase in mhz from an 1100t and that brings us to a total of 4.488ghz to achieve that 17000 score.

So Im guessing if true, this is their highend desktop CPU and its very possible it will be clocked at those speeds given its a highspeed design uarch.
Also, due to the CPU might not meeting its TDP topend under this benchmark load, this may infact be a lesser base clock with some sort of turbo kicking in.

I think I did the math somewhat right somewhere lol, I hate math =P

 

[Elixer]

The BD score looks too good to be true. If they are true, BD is potentially more beastly than previously imagined; by a rather large amount at that.

Most things made up are too good to be true.

There have been no real leaks, and there have been no real benchmarks from any reliable source.

The same kind of hype was made with phenom, and that didn't turn out very well. Phenom II is better than phenom, so we can pretty much guarentee that BD will be better than that series at the same clock speed.

If the lowest end BD can be launched at the highest Phenom II speeds of 3.6 GHz, then that would be a nice start.
It would also be nice if AMD could launch the "first official" 4GHz CPU.

We still have ~4 months of waiting...

Come to think of it, Intel's "re-launch" of their CPU's because of the chipset issue, I can pretty much bet that they will announce a 4GHz CPU.

 

[Mopetar]

I think the 50% increase in performance with 33% more cores is given the same clock speed as a thuban. Which is easily the case. But you ramp up the clocks and you get a result very similar to the one in the graph.

Even if that's true, some basic napkin math indicates that those 8 cores would need to be running at around 6 GHz to turn out those numbers. I know the clock rate is expected to be higher than previous generations, but that's a pretty ludicrous number.

Even if you assume that the quote is a misquote and we're going to get 50% more performance (per core, per clock) and 33% more cores, the clock rate for 8 cores need to be around 4.5 GHz. Not totally out of the ballpark, but still fairly high.

The conditions that need to exist for those numbers to be true are somewhat incredulous. Unless it's a special case where AMD overperforms, it seems unlikely that the number is realistic for an 8 core chip. I'd love to be wrong, but I don't think I am.

 

[drizek]

Besides, JF AMD pretty much said that the graph was fake.

 

[hamunaptra]

Well, of course he would. Im not saying its real, but of course AMD would deny any leaks and dismiss them as fake.

Also, did some more crunching and did points per mhz per core...again if this graph is true and my initial 4.5ghz results are correct:

BD would be .47 points per clock per core
1100t .31 points per clock per core
980x .52 points per clock per core w/ HT advantage(if thats what graph shows)

For the 980x to match BD in terms of per core perform it would need 8 cores running at 4.1ghz.
For 6 cores it would need 5.5ghz

 

[Mopetar]

Well, of course he would. Im not saying its real, but of course AMD would deny any leaks and dismiss them as fake.

So if he says it's real, the results are real because he confirmed it; and if he says it's not real, he's obviously just trying to cover it up and therefor it's still real?

Is there an outcome under which those results are actually fake? It seems like you're accepting the conclusion and and modifying the steps leading to that point in order to fit that point.

Math:

1100T @ 3.3 GHz gets a 6200.

That's a score of 1033 per core or 313 per 1 GHz per core.

Assuming a 50% BD performance improvement per a 33% core count increase and the same clock speed, BD IPC is 12.7% better than Thuban.

Therefor BD performance per core per 1 GHz should be ~353. Take that across 8 cores and it's 2822 per 1 GHz. In order to achieve a score of 17000 it would need to be clocked at slightly more than 6 GHz.

We also know that when using all cores the chip can increase it's clock 500 MHz. This means that if the test were short enough, that 6 GHz could include the turbo boost, or that the chip can turbo boost all 8 cores beyond 6 GHz.

On the other hand if we assume that there's a 50% increase in IPC irrespective of the core count increase, BD's performance per core per 1 GHz would be ~470 and the clock rate would only need to be around 4.5 GHz, possibly including the turbo boost, meaning the chip could have a clock as low as 4 GHz.

However, those performance numbers still seem too good. We can say that AMD's current generation architecture is seriously out of date and that a 50% improvement is possible. 470 points per core per 1 GHz isn't that out of line considering that the 2500K would be around 515 per core per 1 GHz. For this result to be true you have to believe that AMD has made up serious ground with this new architecture and that as good as SB is, it's IPC is only 9% better than BD. The only other alternative is that this result is a special case where AMD does exceptionally well and that this result is not indicative of general performance.

 

[hamunaptra]

Yeah I just looked at my math guess your numbers are more right, not sure how mines screwy but yeah LOL! anyways, good information...maybe the graph is just an OC'd BD on air muahahah! =P Ill be happy with that too.

 

[bryanW1995]

Uh except for the fact that there are lots of things where there exist no work efficient multi threaded algorithms (you don't happen to have a practically relevent parallel algorithm for sparse graphs in a PRAM model lying around?) and others that just don't scale well above a few dozen cores. So sorry to damp your enthusiasm, but not all problems will profit from more cores - though that's an extremely large research area.


But sure 8 cores will work for most stuff just fine, but I just remember the great discussions we had back about whether someone should get a e8400 or a Q6600. How many people said one should get the quadcore because at least by 2010 we'd be severely limited in new games. Well it's 2011 now and I think we can safely say that precognition turned out to be just completely wrong. So if someone wants to bet with me that by 2013 we'll have several games where a quad core will be severely limited compared to a hexcore, I'm all for it


I don't say that there aren't use cases where more cores are useful (distributed computing, en/decoding, ..) just that that's a minority of all users.. but if someone wants to call me out on that, we can make a poll in this forum and see if we get more than 15% (and the AT crowd is doubtlessly an extreme sample )

there are several games today that are already limited by quadcore, civ5 to name one can be up to 40% faster on a hex core. you are generally correct, however, it just depends on the games that you play.

 

[bryanW1995]

Using a debunked lie won't score you any points...it will just make you look uninformed..and a PR tool:

http://en.wikiquote.org/wiki/Bill_Gates#Misattributed
http://www.wired.com/politics/law/news/1997/01/1484

But kinda funny you should bring up the "Nobody needs more than...".
Becuae that is the what you (AMD) is doing in regards to tesselation, because NVIDA has too "much" tesselation power compared to AMD .

Don't use debunked myths to make a point...you get hit by a hard dosis of reality blowback...

yes, and insulting other forum members makes you look mature and knowledgeable, right?

 

[JFAMD]

However, it does concern me that if you can turbo that high the regular clock rate might be lower than hoped. Then again that's not necessarily a bad thing if the IPC is really good, but most of the technical articles I've read have suggested that BD would have a long pipeline and high clock.

Please go back and read my blog, it explains clock speed.

Besides, JF AMD pretty much said that the graph was fake.

No, I did not. All I said was that I was not sure where it came from because I did not make the slide. Please don't put words in my mouth.

Well, of course he would. Im not saying its real, but of course AMD would deny any leaks and dismiss them as fake.

I don't leak and if desktop wanted to do a leak they would have to get my approval because we share the die. It's not in our culture to leak data because it messes up the supply chains for our partners.

 

[maddie]

I think the 50% increase in performance with 33% more cores is given the same clock speed as a thuban. Which is easily the case. But you ramp up the clocks and you get a result very similar to the one in the graph.

The reason I believe that is at the same clock speed is because, if it werent (ie it was the equation of speed increase over thuban including clock speed factory) then the BD would literally suck be horrible at IPC.
Heres my math, I dont know if is right or what but....according to the above 50% more performance which is essentially .5 faster with a 33% more core count so add 2 cores.
Using these measurements:
150% = 8 cores is a ratio of 18.75 per core
100% = 6 cores is a ratio of 16.66 per core

If you figure the percentages of what those 2 numbers are you basically get 12% increase in IPC per core over thuban.

Thats where I see how you got the what should be 9000 based number by just multiplying 6100 by 150%. So, that would be purely clock for clock based.

I would also gather that in order for a thuban to be equivilant to BD clock for clock, going with the 50%/33%
thuban would basically need a total of 9cores (to be 50% faster than what it is) compete with a 8 core BD once again 8 is 88% of 9 which equates to a 12% increase per core over thuban once again.

So, Im guessing a clock speed of 4-4.5ghz for BD according to this bench plus some code optimization to squeeze more performance out of it.

According to that graph, as we can tell BD gets 17000, thuban gets 6200. Which equates to a 48% increase in performance per core over thuban according to that graph.
Seeing as I calculated the 12% above clock for clock. Subtract the IPC increase 12% from the 48% increase and we are left with... 36% increase in mhz from an 1100t and that brings us to a total of 4.488ghz to achieve that 17000 score.

So Im guessing if true, this is their highend desktop CPU and its very possible it will be clocked at those speeds given its a highspeed design uarch.
Also, due to the CPU might not meeting its TDP topend under this benchmark load, this may infact be a lesser base clock with some sort of turbo kicking in.

I think I did the math somewhat right somewhere lol, I hate math =P

Not personal, but I had a flash image of medieval monks calculating the number of angels fitting on a pinhead.

 

[hamunaptra]

lol thanks well Im glad I could provide some laughs at least =) hehe

 

[drizek]

No, I did not. All I said was that I was not sure where it came from because I did not make the slide. Please don't put words in my mouth.

You said that you've never seen a slide like that before. The website say that their source is "AMD-affiliated".

I took your post to mean that you have no reason to believe, at least based on the aesthetics, that the slide is real.

 

[wanderer27]

nm

 

[Ajay]

Please go back and read my blog, it explains clock speed.

This one: http://blogs.amd.com/work/2011/01/31/bulldozer-goes-to-11/ ?

Pretty cool stuff - especially running in TURBO with all cores under load (in a server setup). Sounds like there will be some serious headroom for overclockers (and very high clocks when there are only a couple of threads running and 3 modules can be gated off).

I know you can't confirm this, I sure am looking forward to CEBIT and hoping that some info will come out on performance or speed bins.

 

[Mopetar]

Yeah, that is nice, but it still makes me wonder why there's such a stark contrast with Intel's chips where the most turbo you can get with all cores is 100 MHz. Is Intel not leaving as much headroom for the turbo in their chips? On the other side of the coin, is AMD not being very aggressive with theirs? The only other explanation I can think of is that it's an artificial limitation put in place by Intel that they're using to differentiate their chips.

 

[JFAMD]

I know you can't confirm this, I sure am looking forward to CEBIT and hoping that some info will come out on performance or speed bins.

I can confirm that you will get neither at CeBIT. I have said several times that we disclose both of those at launch.

Yeah, that is nice, but it still makes me wonder why there's such a stark contrast with Intel's chips where the most turbo you can get with all cores is 100 MHz. Is Intel not leaving as much headroom for the turbo in their chips? On the other side of the coin, is AMD not being very aggressive with theirs? The only other explanation I can think of is that it's an artificial limitation put in place by Intel that they're using to differentiate their chips.

I can think of several things. The difference between the two is the derate that happens between different workloads. If you push things closer to the edge, you have less headroom to lean into before you hit TDP. So you get less boost.

In addition, they will run hotter than AMD, so they have less room from a thermal standpoint. They actually allow their processors to go over TDP for short periods of time (I believe ~20-30 seconds) for bursts of performance, but that is not sustainable.

 

[wahdangun]

I can confirm that you will get neither at CeBIT. I have said several times that we disclose both of those at launch.



I can think of several things. The difference between the two is the derate that happens between different workloads. If you push things closer to the edge, you have less headroom to lean into before you hit TDP. So you get less boost.

In addition, they will run hotter than AMD, so they have less room from a thermal standpoint. They actually allow their processors to go over TDP for short periods of time (I believe ~20-30 seconds) for bursts of performance, but that is not sustainable.

btw if we overclock the chip to the max TDP, can we still get turbo activated or its automatically disabled when reaching TDP limit ?

 

[Phynaz]

Yeah, that is nice, but it still makes me wonder why there's such a stark contrast with Intel's chips where the most turbo you can get with all cores is 100 MHz. Is Intel not leaving as much headroom for the turbo in their chips? On the other side of the coin, is AMD not being very aggressive with theirs? The only other explanation I can think of is that it's an artificial limitation put in place by Intel that they're using to differentiate their chips.

Intel grantees that a fully loaded cpu will run one speed bin over it's rating.

AMD has made no such claim.

 

[Riek]

I think the 50% increase in performance with 33% more cores is given the same clock speed as a thuban. Which is easily the case. But you ramp up the clocks and you get a result very similar to the one in the graph.

The reason I believe that is at the same clock speed is because, if it werent (ie it was the equation of speed increase over thuban including clock speed factory) then the BD would literally suck be horrible at IPC.
Heres my math, I dont know if is right or what but....according to the above 50% more performance which is essentially .5 faster with a 33% more core count so add 2 cores.
Using these measurements:
150% = 8 cores is a ratio of 18.75 per core
100% = 6 cores is a ratio of 16.66 per core

If you figure the percentages of what those 2 numbers are you basically get 12% increase in IPC per core over thuban.

Thats where I see how you got the what should be 9000 based number by just multiplying 6100 by 150%. So, that would be purely clock for clock based.

I would also gather that in order for a thuban to be equivilant to BD clock for clock, going with the 50%/33%
thuban would basically need a total of 9cores (to be 50% faster than what it is) compete with a 8 core BD once again 8 is 88% of 9 which equates to a 12% increase per core over thuban once again.

So, Im guessing a clock speed of 4-4.5ghz for BD according to this bench plus some code optimization to squeeze more performance out of it.

According to that graph, as we can tell BD gets 17000, thuban gets 6200. Which equates to a 48% increase in performance per core over thuban according to that graph.
Seeing as I calculated the 12% above clock for clock. Subtract the IPC increase 12% from the 48% increase and we are left with... 36% increase in mhz from an 1100t and that brings us to a total of 4.488ghz to achieve that 17000 score.

So Im guessing if true, this is their highend desktop CPU and its very possible it will be clocked at those speeds given its a highspeed design uarch.
Also, due to the CPU might not meeting its TDP topend under this benchmark load, this may infact be a lesser base clock with some sort of turbo kicking in.

I think I did the math somewhat right somewhere lol, I hate math =P

Yeah the statement was 50% more throughput with 33% more cores compared to to the current 12core offering.

Other way to look at that graph:

SB gets 6800 with 4cores (no HT) @ 3.4Ghz

HT (2600K) brings alot of improvement --> very threaded benchmark with very low ipc. So front end not a bottleneck in BD.

8core BD = 17000

dual 2500 = 13600 -> frequency correction to get ~17000 --> /3.4 * 4.4 = 17600. (600 offset for not perfect scaling.)

= 8core BD = 8 core SB @ 4.4Ghz (no HT).

HT adds around 35% in the bench: so
8core BD = 8core SB wHT @ 3.2GHz
or 8core BD = 4core SB wHT @ 6.5GHz

We know interlagos gets a 500MHz boost over all the cores. So if we assume BD ipc = SB ipc = both @ 4.4GHz -> 3.9GHz 8core BD.
For some reason I don't find the number very far fetched considering i used the boost of interlagos and not the desktop.

We know BD focusses on high average IPC (this is low IPC benchmark)
we know BD has many cores that only slow down a little when front end is the limitation. Front end is not the limitation so both cores can be as efficient as possible.
We know BD has high turbo boosts. Low ipc = low executions = higher boost.

So basically these numbers would indicate BD core ~ SB core (without HT)/ clock for these type of application loads.

(note: i'm not saying that slide is true or fake, just that the numbers are not so outrageous as one would assume).

 

[Anexate]

This whole discussion has become convulsive.
Not to mention unpleasant.
Above all, JFMAD is being treated like a squeezed orange, attacked by a thirsty Paris Hilton which wants to check if there is a drop or two still left.

The whole discussion here is a need for overclock race. 200 millions adopters are not going to scream GHz here, Turbo there, overclock with sub-ionic propulsion.
Intel 1155 sold 8 millions in less than a month; the mass of users decides the fate of a product.

I’m not expecting benchmarks or performance data before the release of Bulldozer.
I expected however an indication of the direction of the whole Bulldozer package.
Since the company failed to communicate I ordered a Sandy Bridge 2600K. On the day of the shipment the whole motherboard issue exploded, so I halted the shipment.
Considering that in one way or another I have to wait till late March (for the new motherboards and the revision) I’ll give AMD a chance to present their new product.

If AMD continues like now, playing the fog of pretended ignorance, I’ll take the revised motherboard from Intel and I can assure AMD they are not going to make me feel sorry if their Bulldozer shows wonder.

Do I have issues with Intel 1155?
Yes, huge.
Their “mega spatial” tech supports one x16 PCI-e. What? In 2011 and on the threshold of the PCI-e 3.0 I have only one old as earth (16 lanes) PCI-e 2.0?
Barely enough for two SLI 570 or 6970CF; cards there are out only because of problems (delay) with the 32nm-28nm tech.

28nm tech should enable 2,5 higher performance of Nvidia Kepler GTX 600 or Southern Islands. The 2600K would get saturated with only one GTX “680”.
Right, Intel has a solution, the LGA2011. So, in 6 months one should throw away the 2600K+motherboard and get a future-proof LGA2011.

Customer: Why there is no 2x16PCI-e on the 1155?
Intel: If you have enough money for SLI or CF then throw some of it in our way.
Customer: This concept, x58, proved to be a dead end socket, surpassed by 1155 and 1156.
Intel: Well, you see, the volume for X58 is not that high to warrant the expense for the shrink of the X58 compatible processors.
Customer: But you were marketing the X58 as premium; the CPUs and motherboards were expensive.
Intel: You can rest assured that the prices of LGA2011 will be even higher.
Customer: Ok, I just want a motherboard that supports 2X16 PCI-e lanes. Once, when SLI and CF were an exception, your design offered flexibility, today when SLI and CF are mature and standard you are trying to present such motherboards as high technology meant only for “enthusiast”.
Intel: Exactly. Once it was a mere marketing tool; there was no money to be made. Today we can push the customer in a corner and add additional margin to our products. If you can spend 700$ or more on two standard GPUs, we can request a piece of that cake.
Customer: There is a huge gap between the X58 and LGA2011; how to bridge it today?
Intel: The 2x8PCI-e capability of Sandy Bridge is enough for two high end GPU available today.
Customer: and in 6 months? With the new Kepler and Southern Islands?
Intel: We will have a new LGA2011 platform, just for the event. We recommend upgrading everything, for the sake of superior performance, of course.


AMD: The new revolutionary Bulldozer is on the way.
Customer: Great! Do you have new motherboards in plan?
AMD: We will offer a revision of the good old fossil grade motherboards.
Customer: You were a low-end player. Are the old motherboards on the same level of the new Bulldozer?
AMD: Probably.
Customer: So you plan a new motherboards lineup?
AMD: The future is written in stars.
Customer: Ok. So, since you’re going to offer a new performance oriented Bulldozer, you motherboards will offer SLI for Nvidia technology?
AMD: Who knows. Probably not.
Customer: If Intel throws out CF, to apply you own standards?
AMD: Our lawyers are ready to fill all available and not available legal and regulatory complaints!

 

[IntelUser2000]

We know BD has high turbo boosts. Low ipc = low executions = higher boost.

So basically these numbers would indicate BD core ~ SB core (without HT)/ clock for these type of application loads.

I think you guys are looking at it the wrong way. Take a look at how well it scales with cores.

1x Xeon X5680: 9721
2x Xeon X5680: 18565(+91&#37

1x Opteron 6176SE: 8203
2x Opteron 6176SE: 15197(+85%)

Core i7 975: 7031
Core i7 980X: 10482(+49%)

Phenom II X4 955: 3938
Phenom II X4 965: 4270
Phenom II X6 1090T: 6199(+57%/+45.2%/avg: +51.3%)

SpecCPUINT_Rate2006 only shows 36% gain from 975 to 980X, yet PassMark is showing near 50%. And SpecCPUINT_Rate2006 has been criticized for being a horrible benchmark itself. What is that mean? It isn't a relevant benchmark at all. It's like Dhrystone, Coremark, Sisoft(add your favorite synthetic benchmark here).

It could very well be the score for the 16 core version, or a dual CPU 8 core version. Divide 17000 by 1.9, then you get 8950, or close to what AMD claimed with Interlagos back last year. That is, 50% faster than the last generation.

Since we know the 2600K isn't 50% faster than X6 1100T in multi-threaded apps, that's another indication that its a synthetic benchmark. Synthetic benchmarks sometimes exaggerate the differences between architectures.

 

[Riek]

I think you guys are looking at it the wrong way. Take a look at how well it scales with cores.

1x Xeon X5680: 9721
2x Xeon X5680: 18565(+91&#37

1x Opteron 6176SE: 8203
2x Opteron 6176SE: 15197(+85%)

Core i7 975: 7031
Core i7 980X: 10482(+49%)

Phenom II X4 955: 3938
Phenom II X4 965: 4270
Phenom II X6 1090T: 6199(+57%/+45.2%/avg: +51.3%)

SpecCPUINT_Rate2006 only shows 36% gain from 975 to 980X, yet PassMark is showing near 50%. And SpecCPUINT_Rate2006 has been criticized for being a horrible benchmark itself. What is that mean? It isn't a relevant benchmark at all. It's like Dhrystone, Coremark, Sisoft(add your favorite synthetic benchmark here).

It could very well be the score for the 16 core version, or a dual CPU 8 core version. Divide 17000 by 1.9, then you get 8950, or close to what AMD claimed with Interlagos back last year. That is, 50% faster than the last generation.

Since we know the 2600K isn't 50% faster than X6 1100T in multi-threaded apps, that's another indication that its a synthetic benchmark. Synthetic benchmarks sometimes exaggerate the differences between architectures.

Well that was the whole point of my calculations.
A) that is it highly multithreaded. Stated by HT gaining alot.
B) that it has a relative low IPC. Again stated by HT gaining alot (alot of free execution resources).

Thats why i compare the score to an hypotetical 8core SB without HT and with HT and thats why i said the difference isn't all that far fetched given the application. So i'm reading you say the same thing as me , but then i don't know what you mean with "I think you look at it the wrong way"

 

[IntelUser2000]

Well that was the whole point of my calculations.
A) that is it highly multithreaded. Stated by HT gaining alot.
B) that it has a relative low IPC. Again stated by HT gaining alot (alot of free execution resources).

Thats why i compare the score to an hypotetical 8core SB without HT and with HT and thats why i said the difference isn't all that far fetched given the application. So i'm reading you say the same thing as me , but then i don't know what you mean with "I think you look at it the wrong way"

Low IPC applications scale with the I/O and caches, and whatever that's not cores well. High IPC applications scale well with core only changes, but Passmark is nothing, its synthetic. I assume the code fits within the L1 caches perfectly well and never stress the core routers, the L2 and L3 caches, or even the memory subsystem. Because it fits so easily within the caches, there's almost no bottleneck or contention with Hyperthreading. The gain is big. I suggest you to take a look at Dhrystone benchmarks and see how well they scale with cores and Hyperthreading.

Somehow you got the notion that with Low IPC=Hyperthreading gains are big, High IPC=Hyperthreading gains are low. While it might make sense in the beginning, Low IPC means IPC<1, High IPC means IPC=1.5-2.5, still far from the 3 or 4 capable on modern CPUs.

I'm saying the results don't make sense. 17000 is way too high. There's extremely few applications where the 980X show 49% gain over the 975.

 

[Ajay]

I can confirm that you will get neither at CeBIT. I have said several times that we disclose both of those at launch.

Oh fudge, yes you have said that many times