AMD Bulldozer Dual-Interlagos Benchmarks On Linux

[[Knuckles]]

Heh:

http://openbenchmarking.org/result/1...KNUC-110322585
http://openbenchmarking.org/result/1...KNUC-110322102

I win

But yeah, what's impressive is that you'll be able to get 4 of these on the same system, for a very reasonable price!

[Raven3x7]

No, it is not. The 2500k is an i5.

You're right that it's an i5 () but it does have turbo http://ark.intel.com/Product.aspx?id=52210

[Sylarr]

You're right that it's an i5 () but it does have turbo http://ark.intel.com/Product.aspx?id=52210

Yea does have turbo, mistake is the article states it has hyperthreading too. Do these articles ever get fixed?

[bbordwell]

Heh:

http://openbenchmarking.org/result/1...KNUC-110322585
http://openbenchmarking.org/result/1...KNUC-110322102

I win

But yeah, what's impressive is that you'll be able to get 4 of these on the same system, for a very reasonable price!

If you have the spare time you should do a scaling test for c-ray, e.g run it with 2 cores enable, then 4 then 8 then 16 then 32, It would be useful in seeing how this result translates into desktop bulldozer performance.

[smitty3268]

They claim +90% performance and +50% extra die space.
All the benefits of an extra core with only +10% extra die space is never going to happen.

Source?

I think you're thinking of Bobcat, which they claimed would provide

90% of today?s mainstream performance in less than half of the silicon area

For Bulldozer, the claims I've seen are 12% die space, and virtually double performance on the right kind of code. On the other hand, the wrong kind of code probably gives no speed up at all.

This source is pretty good at explaining how it works, as well as passing on AMD's 12% die space claim: http://www.anandtech.com/show/3863/a...t-chips-2010/4

[[Knuckles]]

If you have the spare time you should do a scaling test for c-ray, e.g run it with 2 cores enable, then 4 then 8 then 16 then 32, It would be useful in seeing how this result translates into desktop bulldozer performance.

I can try, but I'm not understanding what you're hoping to see. Mine are 12-core G34 opterons, so 16 and 32 aren't even very logical numbers from a test POV.

[Raven3x7]

Source?
On the other hand, the wrong kind of code probably gives no speed up at all.

Actually that would only be the case in single threaded code. The frontend actually supports 2 threads though i suppose there might be some penalty because of the resource sharing in some situations. Also the fp unit can act as 2 128 bit or 1 256bit unit. In short BD has 2 cores that share certain resources which decreases die space but can have a performance penalty in certain situation. That slide btw refers to the die space cost of the 2nd integer cluster only

[schmidtbag]

i'm almost positive the i5 2500k is not multi-threaded.

but really doesn't anyone see the problem with that initial test? 25 seconds is utter crap for 32 cores, no matter what cpu you use. an 8 core sempron or core2 (if they existed) at 2ghz would have more raw processing power than the i5 2500k.

remember, there are plenty of programs and benchmarks out there that are multi-processor compatible but most are limited to 16 cores. if that test had those same limitations, that would mean half of those cores were doing absolutely nothing, and so it doesn't surprise me that 16 1.8ghz cores is about twice as fast as 4 3.2ghz cores.
think about it in this way:
16*1.8=28.8
4*3.2=12.8
12.8*2=25.6
If we were to consider these as theoretical frequencies, proportionally, 25.6 is pretty close to 28.8. Considering the frequency and likelihood of functioning cores, to me, it makes a lot of sense why the amd setup performed twice as fast as the intel setup, and it isn't as crappy as it may seem.

[bbordwell]

I can try, but I'm not understanding what you're hoping to see. Mine are 12-core G34 opterons, so 16 and 32 aren't even very logical numbers from a test POV.

the core counts are arbitrary I just think it would be interesting to see how it scales up, if this test is scaling with 99% efficiency that is not so good for bulldozer, but if there is a large decline as cores go up then that could be good news.

Thought this test may have already answered my question as it is single threaded (http://openbenchmarking.org/result/1...IV-HIMENUBUL97)

That is bad news for bulldozer as that shows about 1/2 the single threaded performance compared to sandy bridge. (2600k gets about 345 in that test where as a 3.6ghz bulldozer would get ~180)

[[Knuckles]]

the core counts are arbitrary I just think it would be interesting to see how it scales up, if this test is scaling with 99% efficiency that is not so good for bulldozer, but if there is a large decline as cores go up then that could be good news.

Thought this test may have already answered my question as it is single threaded (http://openbenchmarking.org/result/1...IV-HIMENUBUL97)

That is bad news for bulldozer as that shows about 1/2 the single threaded performance compared to sandy bridge. (2600k gets about 345 in that test where as a 3.6ghz bulldozer would get ~180)

I think this is really too early silicon:
http://openbenchmarking.org/result/1...KNUC-110322323

Single-threaded new architecture @ 1.8Ghz vs single-threaded old one @ 1.9Ghz and old one wins!? I wouldn't read too much into these results.