Quote Originally Posted by Movieman View Post
well said and I agree. Actually I think it's unfair to compare it to a 980X, 2500K and 2600K is more reasonable


Well, if it holds it's own against a 980X that sort of speaks volumes as to how it will do against a 2500k or 2600K yes?

In Dirt 3, the 980x outperforms the Sandy Bridge CPUs (heck, even the i7 920 does...).

This means that the showing from Bulldozer is actually more impressive against the 980X than against a Sandy Bridge CPU...

However, Phenom II does extremely well in the game as well, but not as well as the 980X.

On the Handbrake side, however, most seem to not realize that six-core 3.3GHz phenom II cores lose quite heavily against even the lowly i5 2400. Given the Bulldozer architecture, I'd expect POOR multi-threaded scaling in comparison to the phenom II X6.

Bulldozer's behavior changes under duress:
1. Clock will be at the low end (3.1GHz vs the ~3.6/7 turbo)
2. Internal module overhead and contention will cost around 7~12% *per core* on integer workloads
3. FlexFPU will likely perform poorly in unoptimized code

These design sacrifices should mean that an 8 core Bulldozer is more like a 6.8-7 core phenom II when under heavy load, which is actually about the performance we see... except we still must have some improved IPC to make up the difference... but only about 10-13% or so, which is not making me happy... I hope the i5 was a 2500, that would make it 15-18% (since phenom II loses by a good chunk in Handbrake vs the 2500). That means IPC for the Bulldozer is "only" about 15-20% lower than Sandy Bridge

Well, at least it is faster than phenom II

With turbo running, though, we see a nice boost - the system should run as well as the i5 in the comparison in single threaded tasks, and will be about 20% faster in heavily multi-threaded tasks, and faster STILL when TDP allows turbo to keep things going and thread-balancing between the modules completely removes the module overhead.

So, interestingly, a quad-core load will be about equal between the two, single-threaded loads should be equal, and the Bulldozer should win in multi-threaded work loads, though the margin by which it wins will diminish the heavier the load becomes (big-time diminishing returns in the Bulldozer design ).

This seems to fit the pricing policy perfectly as well, fits the official benchmarks, and conforms with the known design details of Bulldozer.

I will not be surprised to see Bulldozer pulling some upset wins vs Sandy Bridge in single/lightly threaded tasks beyond what would be plainly expected. The performance characteristics will take some time to figure out for a "LOT" of people... Shop by price, and I think AMD has a winner... and a great starting point for improvements!

That said, though, the design is EXTREMELY beneficial to the server environment. There, the work-loads are light, you normally have good process control so you can set affinity to keep threads which share data on the same module, TDP values can be set to self-manage policy between low/high latency, tasks can be sent to separate modules to optimize throughput, clock scales to the TDP so you can get the best performance/$$, you can concurrently execute a thread per CPU thread at the same rate, and much much more design advantages for Bulldozer.

And, some of those same benefits should make the desktop experience much smoother - once the software catches up to it (I believe there will be a Windows driver for Bulldozer to effect the scheduling for optimal behavior depending on policy [max perf / energy savings / etc...]).

God I want to get my hands on one...

--The loon