AMD PileDriver Discussion thread

[frozentundra123456]

Why is it bad idea? Is it a opinion or a fact? I think it's great idea (opinion). I found this forum very good source for BD information and I am hoping that it will bevery good source for piledriver information.

You have very interesting writing style too. But what that actually has to do with piledriver?

We have been so inundated with marketing hype, rumors, and denial by some about the leaked benchmarks and probable poor performance of bulldozer, that, personally at least, I just dont care to hear more speculation about an AMD product until a definite release date and reliable benchmarks are available. Just give it a rest for a while.

 

[3DVagabond]

I thought I would play around with some numbers. According to Anandtechs review:
Cinabench 11.5 single threaded

2500k --------- 1.47
Ph 2 1100T --- 1.11
BD ------------ 1.02
​

That means the 2500k is about 31% faster. To equal it BD would need a 1.8GHz speed boost to 5.2/6.0 GHz in core/turbo. To equal last generations Phenom II 1100T single threaded performance they would need a clock boost of “only” 9% to about 4.0/4.6 GHz.

In Single Core AIDA64 Queens benchmark

2500k -------- 8141
Ph 2 1100T --- 6091
BD ----------- 5297
​

That means the 2500k is 35% faster. Again, BD still needs about a 1.8 GHz speed boost. To equal the 1100T is needs about a 14% boost to 4.1/4.8GHz.

Did AMD really expect GloFlo's 32nm process to allow a 5.2/6.0 GHz speed? If they had gotten it to the point where it equaled the 1100T (4.1/4.8GHz) in per core performance, would that be considered a win? I am talking end performance here which is IPC plus clock speed. Will process maturity allow 4.1/4.8GHz in six months? Its "only" another 0.5 GHz.

If they had gotten its speed up to 4.1/4.8GHz AND maintained the IPC of the 1100T, that would definitely put them just below the 2500K on single performance, but faster than it in multithreaded. I think that would have been a definite win.

So if AMD can get piledriver's IPC up to 1100T performance, and GloFlo can get their 32nm process capable of producing 4.1/4.8GHz BD CPU, then AMD will be able to claim parity with this gens Intel CPU. Of course, Intels next gen is right around the corner so...

I don't think AMD intended for BD to be as fast clock for clock as SB in single thread operations. I do think they planned for it to be faster than Thuban clock for clock. With 8 "faster than Thuban" cores, and turbo to help make up the difference between it and SB, I do think BD would have been a success. Add to it the ability for a ~5GHz O/C (possibly?) and it would have been a big time success.

 

[frozentundra123456]

I don't think AMD intended for BD to be as fast clock for clock as SB in single thread operations. I do think they planned for it to be faster than Thuban clock for clock. With 8 "faster than Thuban" cores, and turbo to help make up the difference between it and SB, I do think BD would have been a success. Add to it the ability for a ~5GHz O/C (possibly?) and it would have been a big time success.

Yes, but they didnt (make it faster than Thuban) and it wasnt (a big time success).
So what is your point?

 

[3DVagabond]

Yes, but they didnt (make it faster than Thuban) and it wasnt (a big time success).
So what is your point?

Maybe if I break out the part I bolded you'll understand what I mean.

Did AMD really expect GloFlo's 32nm process to allow a 5.2/6.0 GHz speed?

No, I don't think AMD expected 5.2/6.0 speeds. At least not to be released at anywhere near those speeds. 5GHz+ max O/C I think they were hoping for.

If they had gotten it to the point where it equaled the 1100T (4.1/4.8GHz) in per core performance, would that be considered a win?

I think they expected faster than Thuban clock for clock. At least that was what JFAMD said (improved IPC). If it was then the clock speed wouldn't be an issue. I think it's clocked higher than they intended, which is why thermals are so bad. I also think that they expected it to be able to hang with SB in max clocks (~5GHz).




I never claimed it was faster than Thuban, I've actually shown in other posts where it definitely isn't, and I don't expect that it it's a success. At least not an engineering success. Seeing as how I don't think it met any of their design goals. Whether it will be a commercial desktop success or not is yet to be seen, but I'd be really surprised if it is.

Understand that this is conjecture from everything I read leading up to the release of BD. I was just giving my opinion on the bolded part of the post I quoted.

 

[frozentundra123456]

Maybe if I break out the part I bolded you'll understand what I mean.


No, I don't think AMD expected 5.2/6.0 speeds. At least not to be released at anywhere near those speeds. 5GHz+ max O/C I think they were hoping for.


I think they expected faster than Thuban clock for clock. At least that was what JFAMD said (improved IPC). If it was then the clock speed wouldn't be an issue. I think it's clocked higher than they intended, which is why thermals are so bad. I also think that they expected it to be able to hang with SB in max clocks (~5GHz).




I never claimed it was faster than Thuban, I've actually shown in other posts where it definitely isn't, and I don't expect that it it's a success. At least not an engineering success. Seeing as how I don't think it met any of their design goals. Whether it will be a commercial desktop success or not is yet to be seen, but I'd be really surprised if it is.

Understand that this is conjecture from everything I read leading up to the release of BD. I was just giving my opinion on the bolded part of the post I quoted.

First, I apologize if my original post seemed rude. I know what you said, and you are correct. What I am trying to say, is I dont know if the problems are the fault of GF or AMD design or whatever, but I dont see the point of speculating about what AMD wanted or what GF did, or what the performance would have been if they had met this or that condition. The chip "is what it is" and nothing will change that until AMD or GF or both make some improvements.

 

[lifeblood]

I think they expected faster than Thuban clock for clock. At least that was what JFAMD said (improved IPC). If it was then the clock speed wouldn't be an issue. I think it's clocked higher than they intended, which is why thermals are so bad. I also think that they expected it to be able to hang with SB in max clocks (~5GHz).

Now that I think about it, they must have been hoping for around a 4.6/5.2 clock speed. Remember, many of the reviews quoted AMD as saying they were shooting for the same IPC as Phenom II. Given that, they would have to get BD up to about 4.6/5.2 to match a 2500k in single threaded performance. To match a 2600k would have taken even more.

I don't think Piledriver will even equal what AMD hoped BD initially would do. Looking at the speed improvements in GloFlos 40nm process over time as a reference I just don't see a 1GHz increase in speed from 32nm. I also don't see tweaks getting IPC up to 1100T levels. I don't think AMD believes they can do it either. They're claiming 10%-15% over BD for Piledriver. That won't be enough to match the 2500k in single threaded performance.

 

[lifeblood]

First, I apologize if my original post seemed rude. I know what you said, and you are correct. What I am trying to say, is I dont know if the problems are the fault of GF or AMD design or whatever, but I dont see the point of speculating about what AMD wanted or what GF did, or what the performance would have been if they had met this or that condition. The chip "is what it is" and nothing will change that until AMD or GF or both make some improvements.

The benefit of examining where BD failed is it helps you/me understand if Piledriver will be able to achieve what BD didn't since Piledriver is derived from BD.

It also helps answer who's fault it is. At the moment its looking like both AMDs and GloFlos fault.

Its like speculating which football team will win, just for geeks rather than wannabe jocks.

 

[Vesku]

Doubt they were looking to match Sandybridge in single threaded, they had a pretty good idea of retail BD performance when Sandybridge launched. More likely 3.8-4.0/4.4-4.6 turbo with the cores the way they are now with CMT penalty. The current clocks would be OK if they could match Phenom II IPC in current applications as they also bring features the Phenom II's lack.

 

[pmccall2]

IMO, Llano is the only thing AMD has going for it right now. I for one think AMDs top priority will be to get the bd architecture into trinity, and piledriver will reflect that.

 

[3DVagabond]

First, I apologize if my original post seemed rude. I know what you said, and you are correct. What I am trying to say, is I dont know if the problems are the fault of GF or AMD design or whatever, but I dont see the point of speculating about what AMD wanted or what GF did, or what the performance would have been if they had met this or that condition. The chip "is what it is" and nothing will change that until AMD or GF or both make some improvements.

No apology needed.

You are right, it is what it is. You have a lot of people who can't understand how or why something like this can happen. I think it's pretty good when something like this doesn't happen, personally. There was a post by "Dr Who/Francois" on another forum where he was actually defending JFAMD. What he said (paraphrase) was that people need to understand that designing CPU's is incredibly difficult. The "most difficult thing there is" I believe was the quote. You have people who think, "They should've just shrunk Thuban". Like they don't have to do anything for that to happen and in June, when it bacame obvious there was a problem, they could've just ordered up some 32nm Thubans.

I'm going to blame AMD for the final product and GloFo for the pitiful yields. There's plenty of blame to go around, after all.