post #11 of 11
I think it would be better for AMD, and their costumers, to think of an AMD module as a "tandem threaded core" rather than a pair of cores. If an i3 were marketed as a quad core, we would have to spend a lot of time explaining why the performance of the i3 doesn't scale as well beyond 2 threads as a "true" quad core like the i5, and same for the i7. If the i7 were marketed as an octo-core, we would have to spend a lot of time explaining away the deception that such a marketing position creates. AMD has chosen to market their module based architecture in a manner that I believe to be optimistic and deceptive.

The FX-8320 is currently, the best value quad core with tandem threading that money can buy. In parallel computing loads it is absolutely the best performance/$ of anything out there and as a result is worth the efficiency and performance hit in single threaded applications given the price point. Easily one of the highest value chips ever sold. The additional compute throughput offered by Tandem-Threading is exceptional. I would happily point out that an AMD Tandem threaded core offers a low cost alternative to an Intel hyper-threaded core, with a different set of tradeoffs but similar compute-throughput potential.

Marketed as an 8 core (as it is currently), I would not be able to offer a way to "review" the chip with such a positive attitude. The review would be forced to spend too much time unraveling the deception of the marketing choice, and pointing out that it takes 8 AMD cores to compete with 4 Intel cores. I don't see how this marketing strategy has done AMD any good.

Nope. There's still two physical integer cores within a module, and each core is fed by two 128-bit FMACs. These two 128-bit units can combine together as one 256-bit unit to perform the corresponding AVX instructions, but I don't know of any games that would use these.

AMD's implementation of CMT scales higher than Intel's SMT. The performance scaling has nothing to do with the modular approach itself, it's to do with the fact that both integer cores share a single decoder. This decoder is too narrow, and thus when both cores are fully-loaded, there is a performance penalty. If the threads are lightly loaded, or thread A is running at full load but thread B is only a lighter load, there's not really a performance penalty. However, for example, let's take an 8350 here -- a chip with 4 modules and 8 cores. This chip does have 8 FPU's basically, so that's not bottlenecking at all. But you would initially expect at complete 100% processor load for the scaling to be 800% in an application, but in reality the maximum you could achieve is around 650% or 6.5x scaling.

Steamroller addresses and fixes this, but in terms of the current BD and PD, it's a problem.

So AMD's marketing isn't deceptive at all, because they didn't lie about anything. This information is widely available and AMD themselves stated that their CMT implementation would offer most of the performance of CMP.

The reason you see Intel quad-cores matching up with and sometimes exceeding AMD's hexa- and octo-cores is because 1) most games now don't even fully utilize four threads, let alone six or eight and 2) because in that case, four stronger cores would beat four weaker ones, regardless of how many spare cores the weaker chip had in reserve. This is why when Kaveri comes out, despite being a 2M/4C part, it too will surpass the 8350 on many gaming benchmarks. And as has been mentioned, some newer games are still single-thread dependent because of terrible unoptimized and inefficient coding. Planetside 2 is one such game, and due to it having to be re-tooled from the ground-up for its PS4 version, will receive a patch on the PC later to add much more efficient true multi-threading, giving everyone a huge performance boost.