Comparision C2/Phenom2/i7

[IWB]

Hello all

As I am not happy with the standard Benchmarks to compare a CPU I run a bit of a more extensive Test with 20 Single Engines. This is the result:



The colored numbers at the end might help calculation your desired speed.

The AMD 6 core 1055T in this test is running with 0.1V below the specification by AMD. With a 29€ air cooler it is at 40°C when all 6 cores are running! Looking at the price of this it isnt the worst choise!

Bye
Ingo

[Milos]

The AMD 6 core 1055T in this test is running with 0.1V below the specification by AMD. With a 29€ air cooler it is at 40°C when all 6 cores are running! Looking at the price of this it isnt the worst choise!

1055T you can run with a stock cooler at 3.7GHz (and default voltage) without any problem.

[jpqy]

If you do same test and running with all real cores..the AMD Phenom II X6 1055T with a overclock like Milos say @3.7Ghz will get the highest nodes!

This cpu is the best Chess/Performance/Price cpu from today!

Next one maybe the Bulldozer...price?

JP.

[IWB]

Hi

If you do same test and running with all real cores...

I already have problems doing it with one core. You see that I have diffeences in speedup of +13% to -9% for an engine. I tried doing this with 2 cores and found the differences even bigger. Comparing a CPU like this seems to be like comparing apples and oranges to me. Actually this comparision, with all cores, makes only sense if you tend to use only ONE engine. If you want to use multiple engines, your favorite engine will always be faster on that other CPU brand (Murphy!)

Nevertheless, the popular Fritzmark favors Intel by 10%, This means that as long as people are comparing CPUs with Fritzmark, AMD has to be 10% better to get equal ...

Bye
Ingo

[michiguel]

Hi

If you do same test and running with all real cores...

I already have problems doing it with one core. You see that I have diffeences in speedup of +13% to -9% for an engine. I tried doing this with 2 cores and found the differences even bigger. Comparing a CPU like this seems to be like comparing apples and oranges to me. Actually this comparision, with all cores, makes only sense if you tend to use only ONE engine. If you want to use multiple engines, your favorite engine will always be faster on that other CPU brand (Murphy!)

Nevertheless, the popular Fritzmark favors Intel by 10%, This means that as long as people are comparing CPUs with Fritzmark, AMD has to be 10% better to get equal ...

Bye
Ingo

I have seen serious accusations that the intel compiler chooses, on purpose and deceptively, not to optimize code running on AMD with the same aggressiveness that it does on intel. There are apparently some court battles. So, you have to be very careful with which compiler some of the programs have been compiled.

Miguel

[beachknight]

Hi

If you do same test and running with all real cores...

I already have problems doing it with one core. You see that I have diffeences in speedup of +13% to -9% for an engine. I tried doing this with 2 cores and found the differences even bigger. Comparing a CPU like this seems to be like comparing apples and oranges to me. Actually this comparision, with all cores, makes only sense if you tend to use only ONE engine. If you want to use multiple engines, your favorite engine will always be faster on that other CPU brand (Murphy!)

Nevertheless, the popular Fritzmark favors Intel by 10%, This means that as long as people are comparing CPUs with Fritzmark, AMD has to be 10% better to get equal ...

Bye
Ingo

I have seen serious accusations that the intel compiler chooses, on purpose and deceptively, not to optimize code running on AMD with the same aggressiveness that it does on intel. There are apparently some court battles. So, you have to be very careful with which compiler some of the programs have been compiled.

Miguel

I have run Intel and AMD compiles of same engine on Intel computers
and found out that AMD ones are faster than Intel ones (around 10 %).

Strange?

[IWB]

Here is the same list with 4 core engines:



What is intersting:

While the Phenom was 6% slower with the same clock it is 7% faster with 4 cores! Even the distance to the i7 is geting closer by 2%.
When assuming the C2 as a gold standard, the differences on the Phenom for a quad are up to 46% and for the i7 even 78% ...

These differences in speed between the Core2 and the Phenom and even more the i7 is huge (78% is more than another doubeling of the cores)! The CPUs are so different, that an adaption of different speeds for a rating list by changing a time level is impossible when MP engines are tested vs different CPU architectures!

Even for single engines on different hardware there is a gap between engines of 20% which can not be addapted with an adaption of time, but 20% might be accepted as "noise" by someone.

Bye
Ingo

[ernest]

Hi Ingo,
That's a lot of work!
But the multicore non-reproducibility shows its ugly head...
Look at the "different" results of DS 12 x64 and DS 12 32b, which should be the same (same engine, as the single-core results show). Have you run them several times, in order to get an average?
So I think one has to be very cautious with those multicore data...

[IWB]

Hi Ernest.

Hi Ingo,
That's a lot of work!
But the multicore non-reproducibility shows its ugly head...

Not really as this is pure NODE count. This count will be always more or less identical in the starting position.

Look at the "different" results of DS 12 x64 and DS 12 32b, which should be the same (same engine, as the single-core results show). Have you run them several times, in order to get an average?

No need to. The difference is there on 1 core as well. I will send you the excell sheet.

So I think one has to be very cautious with those multicore data...

See above - I think it is quite accurate as I ONLY count nodes, not time to depth ... or find a solution or something. Pure nodecount!

Bye
Ingo

[Milos]

So I think one has to be very cautious with those multicore data...

See above - I think it is quite accurate as I ONLY count nodes, not time to depth ... or find a solution or something. Pure nodecount!

Never rely on this if you use multiple threads. Depth can differ even for 2 plies and node count can differ for more than 100% in two consecutive runs of the same engine from the same (starting) position. Usually differences are less drastic, but still quite visible.