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

[nyker96]

http://www.xbitlabs.com/news/cpu/di..._No_Design_Wins_No_Launch_Date_Announced.html

We expect to begin shipping our first server platform featuring the Bulldozer this quarter. The Interlagos platform is our first server offering optimized for today's cloud datacenters. The [Bulldozer] [micro]-architecture excels at compute-intensive and HPC workloads, where it will deliver up to 35% performance improvements compared to our current offerings....

some more bad news? no wonder they keep such a lid on the product.

 

[Vesku]

This is a misunderstanding that was cleared up with some helpful math in another thread. The estimate was throughput comparison of the 2.3GHz 12 core Magny-cours versus a 16 core BD. The 35% seems to refer to throughput based on the math but it may refer to overall IPC AND it's comparing to the 2.5GHz 12 core Magny-cours that was launched after the original 50% estimation. Just rumor sites looking for page views and not possessing helpful forum members willing to crunch some numbers.

 

[BD231]

This is a misunderstanding that was cleared up with some helpful math in another thread. The estimate was throughput comparison of the 2.3GHz 12 core Magny-cours versus a 16 core BD. The 35% seems to refer to throughput based on the math but it may refer to overall IPC AND it's comparing to the 2.5GHz 12 core Magny-cours that was launched after the original 50% estimation. Just rumor sites looking for page views and not possessing helpful forum members willing to crunch some numbers.

 

[Regs]

Sounds like Phenom I all over again.

 

[996GT2]

I've gotta say, Xbitlabs' tone in that article seemed a little biased

 

[LOL_Wut_Axel]

The [Bulldozer] [micro]-architecture excels at compute-intensive and HPC workloads, where it will deliver up to 35% performance improvements compared to our current offerings.

Seems to me as if they don't understand what he actually said. He's referring to the architecture having performance improvements of up to 35% in comparison to their current offerings, which are based on the K10.5 architecture. In other words, up to 35% more performance from Bulldozer in comparison to K10.5.

 

[996GT2]

Seems to me as if they don't understand what he actually said. He's referring to the architecture having performance improvements of up to 35% in comparison to their current offerings, which are based on the K10.5 architecture. In other words, up to 35% more performance from Bulldozer in comparison to K10.5.

If that was what they meant, then why did Xbit say this in the same article?

It is also noteworthy that Bulldozer's per-core performance is not projected to be much higher compared to existing microprocessors.

Seems like a pretty confident statement from Xbit.

 

[TakeNoPrisoners]

Oh geez what a fail.

 

[AnandThenMan]

At this point I'll be very surprised if this doesn't turn out to be Phenom launch v2.0. Not looking good at all. Intel will soon have a gun to our collective heads, if they don't already.

 

[LOL_Wut_Axel]

If that was what they meant, then why did Xbit say this in the same article?



Seems like a pretty confident statement from Xbit.

Perhaps because they understood a different thing from what he meant.

If they understood that he meant that Interlagos is 35% faster than Magny Cours that's what they'll report, right?

 

[RussianSensation]

"Unfortunately, amid the massive hype the company not only delayed Bulldozer-based products for a number of times, but at some point it decided not to provide any official performance estimates or expectations.

But the alarming thing is that AMD even decided not to tip any performance numbers or show advantages of its chips when running server applications compared to Intel's already available eight-core and ten-core Xeon chips for dual-socket and multi-socket servers.

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."


========================
Same theme over and over again. All the signs are there. Remember rumors of Fermi being hot and loud? There is always some truth to the rumor.

1) AMD delays Bulldozer multiple times, various sources citing insufficient frequency to compete with 2500k/2600k. AMD goes for a stepping respin. If BD was such an improvement in IPC, they would have no problem releasing lower clocked versions and discontinuing the Phenom II. They would have had plenty of time to release faster clock variants at a later date. Surely, with 1.5x IPC improvement an 8 core BD at 3.0ghz would haved mopped the floor with the 1100T....but yet AMD is aiming for a 4.2ghz Turbo 8 core Processors to compete with the 2600k?

2) Another indication that BD will not be as fast as SB per clock is because they are positioning an 8 core processor vs. a 4 core 2600k (Hyper Threading does not equal 2x as many cores). Also, they are positioning a 6-core processor vs. a 4 core 2500k. So that means there is no way BD will be anywhere close to performance per clock of SB. Otherwise it would smoke the 2500k. AMD isn't in the business to give 50% higher performance at the same price, and never has.

3) Phenom II is 50% slower per clock cycle than SB is. It took Intel Core 2 D/Q (Phenom II) -> Core i5/i7 (1st) -> Core i5/i7 (2nd) generations to achieve such a jump in IPC. For Bulldozer to achieve this in 1 go seems like a fairly tale. If their engineers weren't able to pull it off with Phenom I and II, why should they be able to pull it off now? AMD has continued to be strapped for resourced in the last 5 years. Those engineers probably didn't receive the necessary funding to beat SB.

4) Rumored pricing of 8-core versions at only $300 hinted that AMD is aiming to compete squarely against the 2600k, i.e., 8 core vs. 4 core + HT ==> How can this be? That likely means that AMD is relying on highly multi-threaded scenarios. On average, HT provides a very small performance boost, actually about 5-10%. If the IPC per each BD core was really 1.5x faster than Phenom II, we would essentially have an 8-core 2500k for $300!! Seems like another fairy tale to me.

- Finally, SB is a known quantity: it has high overclocking headroom and amazing power consumption while overclocked.

 

[RobertPters77]

Fuck it. I'm dumping all of my amd stock.

 

[tweakboy]

More like 20 percent probably. Sandy Bridge E black will be 20 percent faster then AMD's offering..

just assumptions plz.. Thanks

 

[Kalessian]

I'm hoping the ipc is at least a little better than phII, or if not that then that most chips will reasonably OC to ~5ghz.

If I can grab a 4 core and unlock it to 6 or 8 cores and overclock to 5ghz at 10% ipc over phII then that would be a huge win in my eyes. Even 2 out of those 3 would be good imo

 

[RussianSensation]

I'm hoping the ipc is at least a little better than phII, or if not that then that most chips will reasonably OC to ~5ghz.

I mean let's just think about this:

FX-8150 = 8 core, 4.2ghz (Turbo) vs. 2600k = 4 core + HT, 3.8ghz (Turbo)

But we know 2600k is not that much faster than 2500k, outside of very few select multi-threaded apps. Computerbase found about a 5% performance difference. Hardware Canucks arrived at a more favourable 10-40%.

If we use an average HT increase from Hardware Canucks, we get 18%, or in other words a 2600k 3.8ghz + HT * 1.18 = 2600k 4 core @ 4.5ghz without HT.

vs. an 8 core module BD clocked just 7% lower. How can each Bulldozer core be competitive in IPC under such a scenario? If it was, the 8150 it would literally be about 90% faster!!

AMD claims a Bulldozer module can achieve 80% of the performance of two complete cores of the same capability.

Applying this principle, we get:

FX-8150: 80% x 8 cores x 4.2ghz Turbo clock speeds = 26.88 Ghz
2600k = 4 cores x 4.5ghz clock speeds (HT adjusted for 18% avg benefit) = 18 Ghz

26.88 = 18* X
X = 1.49x

So we still arrive at a situation where SB is 50% faster per clock.

Obviously these are just my rough estimates.

 

[LOL_Wut_Axel]

Paragraph regarding
Bulldozer

What you're saying does have some merits, but I just want to point out that Sandy Bridge has 30-35% higher IPC than K10.5. Also, in multi-threaded applications HyperThreading typically gives a 20% improvement or so. It seems to me as if you're counting HyperThreading as part of the IPC, which it's not.

Remember: Core 2 (45nm): 10% higher IPC than K10.5, Nehalem 10% higher IPC than Core 2, and Sandy Bridge 15% higher IPC than Nehalem.

It does seem odd that they're offering an 8-core CPU for $300-320, but if the die size is smaller than of or similar to Thuban and yields are good I don't see why it'd be impossible. If they can achieve Nehalem IPC I think it's a huge win for them, IMO. That'd mean the FX-8150 would be just slightly faster than a Core i7 990X.

 

[996GT2]

It does seem odd that they're offering an 8-core CPU for $300-320, but if the die size is smaller than of or similar to Thuban and yields are good I don't see why it'd be impossible. If they can achieve Nehalem IPC I think it's a huge win for them, IMO. That'd mean the FX-8150 would be just slightly faster than a Core i7 990X.

The problem is, from the information in the Xbit article it doesn't look like they will come anywhere near Nehalem IPC.

John Fruehe of AMD originally claimed that Bulldozer would bring a 50% increase in performance from the 33% increase in core count:

From a performance standpoint, if you compare our 16-core Interlagos to our current 12-core AMD Opteron 6100-series processors (code named “Magny Cours&#8221 we estimate that customers will see up to 50% more performance from 33% more cores. This means we expect the per core performance to go in the right direction - up," explained John Fruehe, the director of product marketing for server/workstation products at AMD.

It seems like they've now changed that to "up to 35% more performance" for a 33% higher core count. Which essentially means very little IPC improvement over K10.5



Also, Nehalem is already a 30% IPC improvement over K10.5. Have a look at this comparison between a Nehalem i5 without HT and a Phenom II X4 at the same clock speed. The Core i5 is at least 30% faster in most of those benchmarks. Sandy Bridge brings another 15% IPC improvement on top of Nehalem, so it's approaching a 50% improvement over Phenom II.

 

[Vesku]

What's with the strawman, Russiansensations. I don't think anyone claiming equal or better IPC than Sandybridge has gained much traction. Even if it's a 20-30% jump in IPC from Phenom II, that wouldn't bring it into SB territory only Nehalem-ish.

We've known from some of the design choices that BD was designed to hit higher clocks and they seem to be having some issues getting as high as the wanted. That might be the Phenom flashback of this launch, having a whole line of unlocked FX cpus with only 10% headroom.

 

[RobertPters77]

It seems like they've now changed that to "up to 35% more performance" for a 33% higher core count. Which essentially means very little IPC improvement over K10.5

Less than 2% it seems.

 

[RussianSensation]

What's with the strawman, Russiansensations. I don't think anyone claiming equal or better IPC than Sandybridge has gained much traction. Even if it's a 20-30% jump in IPC from Phenom II, that wouldn't bring it into SB territory only Nehalem-ish.

I realize no one is claiming that. The reason I keep coming back to IPC is because SB overclocks to 4.6-4.7ghz+. So, unless BD is able to scale to 5.5-6.0ghz, SB will still be faster. I doubt BD is going to have some magical overclocking headroom beyond SB, considering AMD have had issues with getting BD to hit high frequency in the first place (hence the delays).

 

[Vesku]

http://forums.anandtech.com/showpost.php?p=32032100&postcount=543

Found the post I was referring to at the start of this thread. Thanks go to podspi for trying to bring some analysis into this raging inferno of speculation.

Also, why would it have to hit 5.5-6.0Ghz if it is also offering 8*.8 = 6.4ish cores worth of processing power vs 4, or 4*1.2= 4.8ish with HT. Could Intel's holding back SB-E be in preparation for dulling the BD launch news? Intel loves to do that kind of thing to AMD.

 

[gdansk]

Can't wait until it is released and we can all stop these ... wonderful threads. I don't know if IPC improvements were AMD's aim for Bulldozer. But given that it was given some "speed racer" characteristics and a big bunch of cores, I doubt it was their main aim.

 

[podspi]

Fuck it. I'm dumping all of my amd stock.

I am afraid the time for that is long past. I'm holding on at this point.

John Fruehe of AMD originally claimed that Bulldozer would bring a 50% increase in performance from the 33% increase in core count:



It seems like they've now changed that to "up to 35% more performance" for a 33% higher core count. Which essentially means very little IPC improvement over K10.

Again, you have to remember these statements are about specific products, not IPC. The reason this is important is because when JF initially made that remark, AMD's products were slower than they are now. That is to say, after JF made that remark, AMD released new products which were faster than their old ones. Since their internal projections for BD did not change, the percent change in performance over current top-bin products decreased.

I crunched the numbers in this thread:

http://forums.anandtech.com/showpost.php?p=32032100&postcount=542

And it works out that assuming BD did provide 50% more throughput than AMD's top bin processor when the statement was made, then that same hypothetical BD processor would provide 38% more throughput than AMD's current top bin Opteron. Factor in some conservative rounding, and there you have 35%.


Edit:Vesku beat me to the punch Apparently I took way too long to type out this comment. Glad somebody found the post useful

I remember JFAMD saying at some point (I can't remembre if it was here or SA) that AMD's projections for BD have been ridiculously consistent. I think it is the consumer side that is having clock issues, not server.

 

[Nemesis 1]

This is a misunderstanding that was cleared up with some helpful math in another thread. The estimate was throughput comparison of the 2.3GHz 12 core Magny-cours versus a 16 core BD. The 35% seems to refer to throughput based on the math but it may refer to overall IPC AND it's comparing to the 2.5GHz 12 core Magny-cours that was launched after the original 50% estimation. Just rumor sites looking for page views and not possessing helpful forum members willing to crunch some numbers.

Nothing of the sort was made clear or cleared up. When the 50% better performance figure was made . AMD BD cores that were available were really slow . So no 50% better performance numbers could have been used from that time period . 35 % faster with 25% increase in cores . Gives the performance of BD at 10% less than intels last generation which is exactly the number Bob gave me in Dec . and I reported it here. I think thats pretty good for a company like AMD.

 

[Nemesis 1]

Seems to me as if they don't understand what he actually said. He's referring to the architecture having performance improvements of up to 35% in comparison to their current offerings, which are based on the K10.5 architecture. In other words, up to 35% more performance from Bulldozer in comparison to K10.5.

Ya now your tring to spin it into a core for core comparison . Which is highly unlikely . Same as its unlikely that he is referring to clock for clock between MC and BD. Its likely a higher clocked BD cores against a lower clocked MC cores. Not that bad . But $300 for the topend is a little over priced by $100.Your seronx aren't you.