Slashdot Mirror
AMD FX-8350 Review: Does Piledriver Fix Bulldozer's Flaws?
An anonymous reader writes "AMD just officially took the wraps off Vishera, its next generation of FX processors. Vishera is Piledriver-based like the recently-released Trinity APUs, and the successor to last year's Bulldozer CPU architecture. The octo-core flagship FX-8350 runs at 4.0 GHz and is listed for just $195. The 8350 is followed by the 3.5 GHz FX-8320 at $169. Hexa-core and quad-core parts are also launching, at $132 and $122, respectively. So how does Vishera stack up to Intel's lineup? The answer to that isn't so simple. The FX-8350 can't even beat Intel's previous-generation Core i5-2550K in single-threaded applications, yet it comes very close to matching the much more expensive ($330), current-gen Core i7-3770K in multi-threaded workloads. Vishera's weak point, however, is in power efficiency. On average, the FX-8350 uses about 50 W more than the i7-3770K. Intel aside, the Piledriver-based FX-8350 is a whole lot better than last year's Bulldozer-based FX-8150 which debuted at $235. While some of this has to do with performance improvements, that fact that AMD is asking $40 less this time around certainly doesn't hurt either. At under $200, AMD finally gives the enthusiast builder something to think about, albeit on the low-end."
Reviews are available at plenty of other hardware sites, too. Pick your favorite: PC Perspective, Tech Report, Extreme Tech, Hot Hardware, AnandTech, and [H]ard|OCP.
New AMD processor, higher clocks than the last one but no massive improvements performance-wise. Still rocks at multi-threaded, integer-only workloads, still sucks at single-threaded or floating-point performance, still uses a huge amount of power. AMD giving up on the high end, their top-end parts are priced against the i5 series, not the i7. Since Intel's overpricing stuff, they're still roughly competitive. Might be good for server stuff, maybe office desktops if they can get the power down, but not looking good for gaming. Overall mood seems to be "AMD isn't dead yet, but they've given up on first place".
There. Now you don't need to read TFAs.
By rare, you mean all Asus motherboards. You have heard of Asus? I presume they are in the consumer space? Look at any AM3+ motherboard, or even AM3 or AM2+. All of them support ECC.
I can't believe people keep on complaining about the bulldozer's FP performance. Does nobody realize the shared floating point unit is a 256-bit one, TWICE AS WIDE as the usual 128-bit? Also, assuming you're an HPC user, did you run actual benchmarks? Because we did, and for our (only modestly parallelizable) HPC workload, compiled with a bleeding-edge compiler (not Intel) that supports AVX and running on a bleeding-edge Linux kernel, Bulldozer was remarkably competitive with Intel's offerings at the time, with Interlagos and Westmere getting about the same amount of useful work done per clock cycle. There are some HPC benchmarks on AMDs website that seem very unlikely in light of the mainstream press. However, in light of our own benchmark results, they seem quite reasonable (although we never quite could make it look that good for AMD; probably because AMD didn't go to the same lengths to squeeze the maximum performance out of the Intel systems). Either way, AMD simply blew Intel away on a per-node-price basis, even when compared to Romley. All the way, I was the one arguing that "we should try Intels" based on reviews I saw online, but once we got all the benchmark results in, I simply couldn't argue anymore.
Also, if AMD's FP performance is truly that abysmal, please explain this? AMD bribed Dell more than Intel so that they now market Bulldozer-based Opterons as "excellent for oil and gas exploration, scientific and medical research, video rendering and other challenging HPC projects"???
Of course, this is all for a very specific workload and may not hold for all HPC workloads, but I have a strong feeling that even generally spoken, the Bulldozer's FP performance for HPC applications is just fine. It's just that most FP-intensive applications used in most of the benchmarks we're seeing in "end-user" space are not compiled to take full advantage of it and/or not running on an Operating System that takes full advantage of it and/or not very relevant test cases for the Bulldozer's parallel HPC potential. For example, one of the things we found out is that Intel's frequency scaling is more aggressive than AMD's, so Intel suffers badly if you put all the cores on a die to work at once. Also, Intel's improved HyperThreading still ain't worth shit if you saturate the FP units, while AMD's "clustered multithreading" succeeds to squeeze out a significant advantage owing to the fact that not all of our FP code is easily vectorizable so that sharing the 256-bit FP unit between 2 execution threads works better than trying to keep it busy with 1 thread's vectorized instructions.
/rambling rant