Originally Posted by azanimefan
and yet the benches aren't showing this at all. As i detailed 19 or so pages ago
the FX is actually "punching" WAY under it's weight class in this fully threaded game. My math spelled it out, this game is clearly fully threaded, and well broken down by thread (evenly). So an 8 core FX should perform a LOT better then it currently does. by a factor of 50% better in fact. However the 8 core FX is performing as you'd expect a 6 core... and we know it's not because it's only 6 core optimized, because the 6 core is performing almost exactly 25% worse then the 8 core piledriver, which is exactly how it SHOULD perform if one chip is using 8 cores and the other is using 6.
in short, BF4 might be fully threaded and even AMD optimized, but it's optimized for JAGUAR, not Piledriver. Apparently the chip structure is too different, because piledriver is actually handicapped when running bf4 (that or it's fully optimized for the core i cpu, which makes no sense, since it's designed for a jaguar game cored console)
admittedly the math in my previous post was based on a lot of supposition and thin information, but the one fact that really stood out was the core i sandy bridge cpus were getting MORE then twice the IPC the piledriver chips were getting. Which we know isn't the case with those two chips. the difference isn't 100%, it's really closer to 25%. So piledriver in BF4 is severely handicapped. (sidenote, intel's i7s, were only getting a little more then 9% boost from their hyperthreading which tells me there is a LOT of work that could be done for BOTH chips to make this game run better)
I had already read it, your calculation was totally wrong with FX.
Did you ever read about how Bulldozer architecture actually works ? The 8 core 8350 at 4.0 GHz doesn't mean that each single core out of 8 is running at 4.0 GHz in Multithreading. This is exactly where you missed. It means each module in FX 8350 runs at 4.0 GHz . All this means that a certain set of transistors that covers a module, runs about 4.0 GHz (combined frequency of each transistor). And there are 4 modules. And a module is distributed into two cores with shared FP/INT resource. Only one of two cores can do FP math at a specific time, and other core is free to do INT math at same time. Though its still unclear what is the ratio of transistor distribution when one core does FP and other core does INT math at same time. I am saying that 8350 at default clock, no two cores of one module can run at 4.0 GHz concurrently. This can't happen. Look into Haswell i7 with 1.4 Billion Transistors, they have been able to manage at 0.35 Billion at 3.9 GHz per core (And this is 22nm we're talking about which is even 45% smaller than FX's 32nm) and AMD Bulldozer has only 1.2 billion transistors, and no way in any world AMD can manage 32nm 0.15 billion transistor set (1.2 / 8 cores) to run at 4.0 GHz switching frequency, at least not till today while retaining temp under 100c with 125watt TDP. Finally you can say when Battlefield 4 runs on a 4.0 GHz 8 Core FX, no single core is running at 4.0 GHz. It should be like 2.5 GHz / 1.5 GHz (1st core / 2nd core in a module), 1.0 / 3.0, 2.0 / 2.0 or similarly according to load. All I want to prove that i7 will obviously run it faster due to its very strong single core architecture, but when it comes to 8 threaded performance FX should handle it more precisely, even then when each core is much slower comparison to i7. And this is not about min,avg and max fps. I am not saying QC i7 will have trouble running BF4, This is about stability in different different complex scenario where dedicated cores might maintain or handle it better than hyper threading which is consuming cycles from same core. Internet benchmarks don't show that. Yes for now you can say that 8 core FX is not performing the way it should have performed because of about 50% less FP throughput than i7. Upcoming steamroller architecture seems to be fixing those issues because both cores in a module will be doing FP math concurrently with full throughput. (100% improvement in FP math over Piledriver ).Edited by sumitlian - 10/7/13 at 5:56pm