Bulldozer Folding @ Home Performance Numbers

[hamunaptra]

You know Im curious, with BD...does this mean that there will be unpredictable thread performance when doing FPU based calculations?

Cuz if the OS scheduler is not module aware, which JFAMD said several times OS schedulers will not be module aware.

Given the case of 4 FP heavy threads. Lets say there are cores 01 23 45 67
OS schedules 2 threads on cores 01 , one thread on core 2, one thread on 4.

Does this mean the 2 threads on cores 01 could be up to 2x slower because of the shared FP...whereas on cores 2 and 4 those threads can claim the entire FP?

So, if someone runs benchmarks improperly...this could really skew the actual FP performance results right?

 

[JFAMD]

Double whammy for servers with both Dirk and the server head-man stepping out in the same business quarter.

Nobody in servers has left.

You know Im curious, with BD...does this mean that there will be unpredictable thread performance when doing FPU based calculations?

Cuz if the OS scheduler is not module aware, which JFAMD said several times OS schedulers will not be module aware.

Given the case of 4 FP heavy threads. Lets say there are cores 01 23 45 67
OS schedules 2 threads on cores 01 , one thread on core 2, one thread on 4.

Does this mean the 2 threads on cores 01 could be up to 2x slower because of the shared FP...whereas on cores 2 and 4 those threads can claim the entire FP?

So, if someone runs benchmarks improperly...this could really skew the actual FP performance results right?

Actually, this is wrong on lots of levels. I am not going to get into this but the idea that something would ever be 2X slower is so wrong that I won't even entertain that discussion.

 

[drizek]

I am willing to bet a lot of people will be surprised and disappointed when BD prices turns out to be not as "value" orientated as AMDs current lineup. I do not see a 8 core BD being under $500. But that is just my opinion.

AMDs current systems are terrible value. A Thuban isn't worth over $200. Bulldozers are smaller chips and I think AMD would like to actually sell some of them, so I doubt they would be a worse value than their existing chips.

 

[Idontcare]

Nobody in servers has left.

Doh! You're right, I had it mixed up from when I read the EETimes article:

Rivet's role in the company had been diminished since the time he served as CFO, Freedman said, and AMD's lost market share in server chips under Seyer's watch left him vulnerable.

http://www.eetimes.com/electronics-news/4212993/AMD-s-COO-leaves-firm

^ I thought that meant he was somehow connected to servers. Now I reread the article and realize there was no org chart connection.

 

[OneEng1]

If an 8c BD trounces a 6c Core i7 980x, expect it to be priced like one too ..... it only makes sense.

It is highly likely that AMD will have more demand than they have chips at first. This is another reason the price will be elevated for a while.... everyone remembers supply and demand right?

I agree that people expecting AMD to price a performace holding chip at a value processor price are dreaming

 

[drizek]

No, it doesn't make sense. The 980x is not a high volume part, Bulldozer is. There is no point in providing something that is competitive in terms of value with an intel chip that literally nobody thinks is a good value to begin with.

They might have an extreme edition that costs a thousand bucks, but they are competing with the 2500K, and if they want to sell these things they need to price them like the 2500K.

If the 2-module bulldozer beats a 2600K, I can see it going for $250, and the 4-module selling for $700, but I doubt that that will happen.

Realistically, the only way they will have super expensive chips is if they literally aren't able to manufacture them fast enough to keep up with demand. Otherwise, they need to take advantage of the next 6 months to sell as many of these things as they possibly can, before Intel whoops them at 22nm.

 

[itsmydamnation]

Realistically, the only way they will have super expensive chips is if they literally aren't able to manufacture them fast enough to keep up with demand. Otherwise, they need to take advantage of the next 6 months to sell as many of these things as they possibly can, before Intel whoops them at 22nm.

if BD is competitive with SB its going to be fine,

1. it will take a long time for intel to move all the product lines that compete with BD across to 22nm.
2. its only incremental power/perf increases, while that means better margins for intel if BD performance is still there it wont make massive difference to AMD.
3. intel will likely be at capacity for a while until more fabs move to 22nm so its unlikely to be a price war.
4. intel 22nm yeilds could alway be s***
5. AMD could always release a 140-150watt part if needed.

assuming that BD can compete with SB i think the bigger thing for AMD will be having a responce for Haswell.

 

[Dresdenboy]

You know Im curious, with BD...does this mean that there will be unpredictable thread performance when doing FPU based calculations?

Cuz if the OS scheduler is not module aware, which JFAMD said several times OS schedulers will not be module aware.

Given the case of 4 FP heavy threads. Lets say there are cores 01 23 45 67
OS schedules 2 threads on cores 01 , one thread on core 2, one thread on 4.

Does this mean the 2 threads on cores 01 could be up to 2x slower because of the shared FP...whereas on cores 2 and 4 those threads can claim the entire FP?

So, if someone runs benchmarks improperly...this could really skew the actual FP performance results right?

What do you mean by "unpredictable"? That benchmark results will vary by 50% during different runs due to OS-unawareness? It's a bit more complex. I have the impression that many think of a CPU as a static device. That's no more the case.

Due to core-hopping threads there will be modules running 2 threads and others could either have only one or even no thread running on them. What would happen then?
2 threads: FPU works like with SMT, but with more resources and flexibility than family 10h, turbo might kick in depending on power consumption
1 thread: FPU, front end etc. is fully available to one thread only, turbo even more likely due to less utilized resources
zero threads: whole module is idle and would be gated off, unused power budget could be used for turbo mode of other modules

Hiroshige Goto published a very detailed schema of BD's FPU here, filled with info from ISSCC as it seems:

High resolution (and readable of course):
http://pc.watch.impress.co.jp/img/pcw/docs/430/801/html/9.jpg.html
Full article in Japanese:
http://pc.watch.impress.co.jp/docs/column/kaigai/20110304_430801.html
Translated:
http://translate.google.com/transla...co.jp/docs/column/kaigai/20110304_430801.html

 

[PreferLinux]

if BD is competitive with SB its going to be fine,

1.it will take a long time for intel to move all the product lines that compete with BD across to 22nm.
2. its only incremental power/perf increases, while that means better margins for intel if BD performance is still there it wont make massive difference to AMD.
3. intel will likely be at capacity for a while until more fabs move to 22nm so its unlikely to be a price war.
4. intel 22nm yeilds could alway be s***
5. AMD could always release a 140-150watt part if needed.

assuming that BD can compete with SB i think the bigger thing for AMD will be having a responce for Haswell.

Intel could do the same. Then where would we be? More power-hungry CPUs is where.

 

[Idontcare]

Intel could do the same. Then where would we be? More power-hungry CPUs is where.

No one forces you to buy them, and the top-power bins would just push down the prices on the lower power bins that it sounds like you are interested in anyways. What's not to be happy about?

 

[IntelUser2000]

They did release a 140W part. It's called the Phenom II 965BE.

 

[bandgit]

Anyway when you buy your LGA2011 Sb rig with 32 GByte just don't forget to order SSDs as well!

You can bet on that! Best one I can get for around $400 at the time!