AMD Zambezi/FX/Bulldozer chess benchmarks

[Gian-Carlo Pascutto]

http://www.hardware.fr/articles/842-19/ ... fritz.html

Summary: even at 8 cores, it does not outperform the Phenom II X6 (let alone Intel chips) in NPS. Real speedup will be lower, and the individual cores are slower as well. It also uses more power on load.

This is not the chip to get.

[M ANSARI]

That is extremely disappointing! I was hoping AMD would get something really competitive out there ... and the idea of having 64 cores seemed very appealing. I guess with more cores, you simply have to accept lower core speed. With chess it is a touch choice since with less cores, scaling efficiency is much better ... so going less cores at higher speed can trump many more cores at lower core speed.

I guess for me the next thing to look for would be Intel's 8 core processor that would allow an overclockable 2 x 8 core platform.

[rbarreira]

According to that page, the Bulldozer FX-8150 does outperform the top Phenom X6 at Fritzmark (11861 vs 11630), despite not being the top Bulldozer chip.

The FX-8170, being clocked at 3.9 GHz instead of 3.6 GHz, should theoretically get around 12,900 Fritzmark.

I've heard rumors that a future stepping of the Bulldozer chip would improve integer performance, but that was a while ago and I haven't kept up with the news about it...

[rbarreira]

it does not outperform the Phenom II X6 (let alone Intel chips) in NPS

In the Houdini benchmark, the Phenom X6 is above the Intel chips (so the "let alone Intel chips" part of your statement doesn't make sense). In the Fritzmark, the Bulldozer chip is above the X6 (so the first part is wrong).

In short I don't understand what benchmark you were looking at to make that statement.

[Gian-Carlo Pascutto]

According to that page, the Bulldozer FX-8150 does outperform the top Phenom X6 at Fritzmark (11861 vs 11630), despite not being the top Bulldozer chip.

1) The Bulldozer FX-8150 is the fastest chip that was released today. AMD did not release any concerte plans for a faster chip yet, let alone you could actually buy such a thing.

2) You're talking about a 2% faster result in NPS, but this is using 2 more cores. Unless you have perfect parallelization speedups, the effective speed of the Bulldozer is hence much lower. And it does so drawing more power under load. And it's probably more expensive, too.
(I said "doesn't outperform in NPS" because I didn't consider a 2% difference to be significant, and its entirely moot because of the above)

I've heard rumors that a future stepping of the Bulldozer chip would improve integer performance, but that was a while ago and I haven't kept up with the news about it...

"Future versions" of chips are generally rumored to be faster than current ones. You can't actually *get* those, though.

[Gian-Carlo Pascutto]

In the Houdini benchmark, the Phenom X6 is above the Intel chips (so the "let alone Intel chips" part of your statement doesn't make sense).

1) Intel is selling faster chips than those in that table.

2) The Intel chips are 4-cores. The AMD's are 6 and 8-cores. Similar NPS with more cores is a losing proposition.

[rbarreira]

2) You're talking about a 2% faster result in NPS, but this is using 2 more cores. Unless you have perfect parallelization speedups, the effective speed of the Bulldozer is hence much lower. And it does so drawing more power under load. And it's probably more expensive, too.

You specifically said "in NPS" in your post, I corrected that.

While you have a point about speedup, these programs are not optimized for Bulldozer, which is a quite unique architecture today (there is no other x86 CPU which shares so much hardware between cores).

"Future versions" of chips are generally rumored to be faster than current ones. You can't actually *get* those, though.

Actually you can, because there are reports that some of the leaked benchmarks were made with versions of the chips that are not the final silicon.

Not that I would be surprised if this particular benchmark is indeed correct (and no the benchmarks are not really exciting at this point).

[Gian-Carlo Pascutto]

You specifically said "in NPS" in your post, I corrected that.

That's right. I was editing my post to clarify: a 2% difference is within measurement error for such benchmarks, hence I said "does not outperform".

Actually you can, because there are reports that some of the leaked benchmarks were made with versions of the chips that are not the final silicon.

I'm not sure what you're referring to, but the benchmarks I linked are official ones - the NDA was lifted today.

And no, AMD hasn't announced any concrete plans for the faster than FX-8150 chips. Last I heard was Q1 2012.

[rbarreira]

In the Houdini benchmark, the Phenom X6 is above the Intel chips (so the "let alone Intel chips" part of your statement doesn't make sense).

1) Intel is selling faster chips than those in that table.

2) The Intel chips are 4-cores. The AMD's are 6 and 8-cores. Similar NPS with more cores is a losing proposition.

What desktop Intel chips are faster than the i7-2600k except the insanely expensive i7 980x which costs 3 times as much as the top Phenom X6?

I know there are 8 and 10 core Intel chips, but those are server chips and cost a lot upwards of $1000.

[Gian-Carlo Pascutto]

What desktop Intel chips are faster than the i7-2600k except the insanely expensive i7 980x which costs 3 times as much as the top Phenom X6?

You already answered your own question (i980, i990, some X and W series 6-core Xeons, and of course the very high end >6 core Xeons).

As for the price difference: blame AMD. As long as they aren't competitive at the upper end, Intel gets to set the price almost where they want. This is very bad news for us. And Bulldozer doesn't seem to improve the situation.