Some SMP measurements with Rookie v3

[Modern Times]

Some engines fare better than others on the Piledrivers. Crafty suffers a significant slowdown for example with multiple instances, but when I tried Stockfish I think it it was only about 10% slowdown.

[Modern Times]

I ran some test with my X6 1100T vs the FX8350. I measured the kn/s on both machines. You can clearly see that the FX8350 scales worse.

http://computerchess.org.uk/ray/FX8350.PNG

[mvk]

I ran some test with my X6 1100T vs the FX8350. I measured the kn/s on both machines. You can clearly see that the FX8350 scales worse.

http://computerchess.org.uk/ray/FX8350.PNG

I don't have 1100T data, which surprises me a bit. I owned both the 1090T and the 1100T at the time (and even a 1035T), but for some reason I hadn't run the benchmark on the faster machine. Both systems got a second life with a 8350 now, and another one was added for a total of 3 (and still the 1035T is in a closet somewhere). I like these chips a lot. It figured it doesn't make sense to buy the most expensive hardware as long as improvements must first come from software. My electricity bill is considerable, though. Next systems will be probably be Intel for that reason.

[Modern Times]

I'm a huge fan of the Phenom II X6 1010 T and 1100T, they have been brilliant CPUs for me. At the time when others were buying Intel Sandybridge, with just 4 cores, the X6 just made so much more sense. The one I have left I will never upgrade, I'll run it as long as I can. That machine is running at 3.6 GHz.

The FX8350 are fantastic value for money too and are great general purpose machines. Very happy with mine, and I've run several very long time control matches on it on 8 cores. But as you say use more power than Intel.

[mvk]

Tthe modular design of the AMD Piledriver architecture might be holding back the 8CPU performance a little. I'm not sure of the real world effect of that.

[Note: updating an older thread here]
I have found another cause: It appears that the hardware I used for this test has a problem with scaling to full load. The ASRock 960GM-VGS3 FX motherboard throttles down to 2100MHz above 95W, where 125W is needed. It is a bit disingenuous that ASRock advertises this board as "Support 8-core CPU", this text even appears in big letters on the board itself, when it does this kind of thing to them. I don't have this problem on my other fx8530 systems that have Asus boards. I suppose I will rerun this test later on a better machine, or just upgrade this one. In the meantime I have been using this system with an underclock of 3500MHz and voltage reduction of 1.2V for tests that are not timing critical but where I still like to see some throughput.

The good news is that the simple SMP algorithm must be performing better than I thought at 8 cores.

[Joost Buijs]

I have found another cause: It appears that the hardware I used for this test has a problem with scaling to full load. The ASRock 960GM-VGS3 FX motherboard throttles down to 2100MHz above 95W, where 125W is needed. It is a bit disingenuous that ASRock advertises this board as "Support 8-core CPU", this text even appears in big letters on the board itself, when it does this kind of thing to them. I don't have this problem on my other fx8530 systems that have Asus boards. I suppose I will rerun this test later on a better machine, or just upgrade this one. In the meantime I have been using this system with an underclock of 3500MHz and voltage reduction of 1.2V for tests that are not timing critical but where I still like to see some throughput.

The good news is that the simple SMP algorithm must be performing better than I thought at 8 cores.

This is one of the reasons I always try to buy stuff from reliable brands, in my view ASRock doesn't belong in that category.

You may want to rerun your test on the Intel machine you recently bought, the results will probably be a lot better. It has 8 real cores and not 4 + 4 crippled ones.
A few hundred games will be enough to get a reasonable impression, it is not very interesting to know the difference within a single Elo point.

Since it was a long time ago when I checked the SMP performance of my engine, and I recently made some changes, I decided let it run for 450 games 5-1 with egtb disabled.
I gave the 8 core version an 8 time bigger transposition table because it generates 8 times the number of positions in the same time, so 8 GB vs 1 GB.
The results are 220 wins 10 losses and 220 draws, this is a difference of ~176 Elo points which is completely in line with the expectations.
Although the TTD speedup is not that good (between 4 and 5 on 8 cores) it seems that widening of the search tree also adds some playing strength.