Support NeoGAF

[Pimpwerx]

You have to consider the system as a whole in the real world when discussing throughput, not some mythical scenario where the system is magically running at 100%. (IE the whole peak number vs real world thing that some people love to ignore when comparing hardware specs generated by marketing.)

What Allard is saying is perfectly valid in the scope that he's referring to. For example, if your single thread is doing something which is dominated by memory latency (not completely unlikely), you can easily construct cases in which having two threads in SMT could in fact double throughput over a non-SMT core. You haven't exceeded 100% "per construct" but you have doubled throughput, which is in fact the only thing that a developer in the real world cares about.

It's a rigged comparison. You could then say you are running at 100% efficiency, as long as you make the caveat that it's 100% effcient for the code you wrote. You have to take the ideal situation for these comparisons, otherwise there's no realistic baseline for comparison purposes. I'm the tallest person...in my apartment. I'm the strongest man...in my school. See what I'm getting at? You're doubling capacity...of inefficient code. You're not doubling capacity base on the actual underlying hardware. The hardware can only work on one thread per core, so your realworld performance is like 1.5x increase, never 2x. I'm not gonna badmouth Allard, since it's just PR. But I don't think it needs justification either. PEACE.

 

[aaaaa0]

You have to take the ideal situation for these comparisons, otherwise there's no realistic baseline for comparison purposes.

The problem is the ideal sitation is useless except for flame wars on TEH GAF.

A developer doesn't care about some benchmark against 100% utilization. The developer cares how fast the platform can run his piece of code, and if it is fast enough for what he wants to do. If it's not, he optimizes and rewrites it until it is. If it can't be optimized enough, then he cuts features until it can run at the speed he needs it to run. If he's got CPU to spare, he adds features until it runs too slowly again, or he runs out of development time.

You're doubling capacity...of inefficient code.

First of all, it's not necessarily inefficient code. The code could be completely optimal but still limited by (just for example) memory latency, and there is nothing you can do about that. And it's a fact that SMT will give you a nice performance boost in those situations that can nearly double throughput for that operation against a processor without SMT.

SMT doesn't double your throughput, regardless of what that douchgebag Allard alludes to.

And the point is it can, sometimes.

 

[jett]

Is using 1 core the new beta kits?

 

[Pimpwerx]

The problem is the ideal sitation is useless except for flame wars on TEH GAF.

A developer doesn't care about some benchmark against 100% utilization. The developer cares how fast the platform can run his piece of code, and if it is fast enough for what he wants to do. If it's not, he optimizes and rewrites it until it is. If it can't be optimized enough, then he cuts features until it can run at the speed he needs it to run. If he's got CPU to spare, he adds features until it runs too slowly again, or he runs out of development time.



First of all, it's not necessarily inefficient code. The code could be completely optimal but still limited by (just for example) memory latency, and there is nothing you can do about that. And it's a fact that SMT will give you a nice performance boost in those situations that can nearly double throughput for that operation against a processor without SMT.



And the point is it can, sometimes.

No one knows what hw can do until they get it in front of them. Until then, they look at the ideal, period. The statement was flawed, period. What can or cannot happen in framed situations ultimately means nothing. Otherwise, the PS1 is the most powerful piece of hardware ever, so long as you qualify it enough. That's all I have to say on this issue. PEACE.

 

[aaaaa0]

No one knows what hw can do until they get it in front of them.

Precisely.

Until then, they look at the ideal, period.

No. The ideal can only give you a vague and not particularly useful idea of what the machine can do. Most developers will sit down, write some engine test code and benchmark that to get an idea of what they can do with their game before they enter full production.

Like I said, ideal numbers are only good for fanboys and flamewars.

 

[Hurricane_Chris]

Vince, (or any one in particular) has there been any info disclosed about the PPE on cell ?
Just wondering about the structure (number of execution units etc) and changes from DD1 to DD2.

link

 

[Doc Savage]

I was under the imperssion that each core can execute two threads so in total it is possible to have six threads being run in parallel to one another and each thread can be used for something different like A.I., physics, etc.

 

[ThirdEye]

It's not that they don't use 3 cores...
Can't use more than 1-core theoretical processing power = bottleneck