Maximum PC

If anyone has been keeping up with reviews (rather than just accepting Intel is better), it's nice to understand why AMD's chips are "meh" at lightly threaded tasks and why Intel puts the smack down on them.

Let's look at the 10h architecture first, which was used in the Phenom II: https://upload.wikimedia.org/wikipedia/ ... ch.svg.png . If you look at the top, you'll see that the instruction fetcher dispatches into three different decoders. Of course, this means every core gets three decoders.

Now we'll look at AMD's Bulldozer architecture: https://upload.wikimedia.org/wikipedia/ ... ack%29.PNG . While on the surface it looks like AMD added another decoder, we also have to look at what else they did. The integer cluster, which is where a bulk of the computer operations go to, has been nerfed a bit. There's only a pair of ALUs/AGUs per integer cluster, versus three in 10h. Not only that, if we were to share the instruction decoders fairly between the two, each integer cluster gets two decoders instead of three like in 10h. This means less decoding performance (which impacts branch prediction and other things) and the integer clusters themselves are weaker. You also can't "combine" them, because a thread can only be executed on one integer cluster.

Now let's take a look at Sandy Bridge: http://images.anandtech.com/reviews/cpu ... ontend.png (instruction decoder part) and http://images.anandtech.com/reviews/cpu ... nbexec.png (execution part)
Already we see that Sandy Bridge processors get 4 decoders per core. Unfortunately it looks like Intel organized the execution part so much that I can't directly compare it to how AMD made theirs. But we can at least see they have three integer based units.

This may also explain why if you give Bulldozer a heavily threaded task, it can perform as well as a Core i7. Combining all the resources of a AMD module gives you 4 decoders, 1 FPU, and four integer units. Compared to a Sandy Bridge core, you have 4 decoders, 1 FPU, and basic three integer units (assuming I'm reading that diagram correctly). But when you have one thread of work, AMD falters in the execution area since only two integer units are available.

I'm wondering what AMD was thinking when they thought of this. Maybe since their server procs are highly prized, they thought to cater to them first and sell off "dumbed down" versions to consumers?

AMD's processors are not as bad as most reviews paint them to be. In fact, I think the whole tech site / benchmark scene has devolved into a state they were in back in the late 1990s when it was shown that various vid card / cpu manufacturers colluded together to tweak drivers and even (yes) benchmarks to make their products look better.

So with all that said, I would encourage someone looking at AMD to look at the comprehensive benchmark lists on sites like Tom's Hardware or Anand, and skip reading their commentary.

For example, if you go to the link below and look at a comparison of the FX-8350 vs i5-3470 (these are price equivalents), you'll find that the 'winner' is not all that clear :

http://www.anandtech.com/bench/Product/697?vs=702

I mean, really, in PoV-Ray tracing the Fx-8350 is a solid 40% faster.

In x264 decoding, FX-8350 is between 20 and 40% faster.

The biggest place it 'loses' is in video games, and I suspect that is due to memory bandwidth being a bit lower on FX than Intel.

But then go back and see what kind of rig you need to tell the difference between processors in those games, and you'll quickly find that you'll need a $500+ GPU.

Hint : a GTX 660 Ti won't cut it. Note TweakTown's review below of the FX-8150, there is little discernable difference between that and an i7 using a GTX 580 :

http://www.tweaktown.com/articles/4348/ ... ndex9.html

Then there are some videos wandering the net, where the 8-core FX 8320/8350's are eating up i7-3770's if you attempt to stream video while playing a game. And, in a lot of random games, flat out beat i5's and i7's.

Look here, especially the last half of the video :

https://www.youtube.com/watch?v=eu8Sekdb-IE

For my purposes, running 2-3 VMWare Workstation instances at once to simulate a production environment for testing and development, I'm very happy with the AMD. I say that having come from an i5-2500s; the FX-8320 whips it soundly in this particular arena.

We could go back and forth with benchmarks saying AMD's chips aren't all that bad and other's saying AMD can't even compete against a i5-3570K. And while sure, AMD's processor beats Intel's offering at $199, if you spend $20-$30 more, you get something will beat the AMD chip in all but the most heavily threaded tasks. This post is mostly to analyze why AMD sucks at single threaded tasks. And I know there are applications where multiple cores matter more than raw performance per core alone (and AMD is the go-to for super computers today).

The only thing I don't like about AMD's chips however is they're power hungry and expel more heat (noted by http://www.tomshardware.com/reviews/bui ... 66-14.html ). And if we were to take Tom's Best CPU for the Money articles like the Word of God, there is no $200 segment and AMD is recommended only for the lower end of the spectrum from the latest iteration.

Did you look at the video I linked to? It's interesting, and revealing. The guy doing that is more of a start-up video tech site, but he points out something that I have wondered about for quite some time. It always seems that the vast majority of major online reviewers use the same games for benchmarks... so they always come to the same conclusions.

Specifically, review sites tend to use PCMark7 (which is badly flawed, different conversation), Crysis 2, Metro 2033, Dirt 2 and HAWX. Yet as he demonstrated, in many other extremely popular games an FX-8350 can not only take on and beat the i5-3570K, but the i7-3770K and the socket 2011 i7-3820 as well.

Now why is that? Is it that the reviewer community is creating its own self-reinforcing feedback loop, like lemmings to the cliff, using the same tools to come to the same conclusions with no real analysis going on?

On the costing side - I picked up my FX-8320 and motherboard from Microcenter for $175 + tax. For that price, you would have to drop down to i3-3220 and a cheap H61 motherboard on the Intel side. Granted for an extra $60, you could get an i5-3470 (non K, slightly slower stock speed). For an extra $90 you could go with the 3570K combo. That's significant, when you're doing a build or upgrade - that $90 can buy a next tier higher video card, an SSD, a good power supply, etc.

I went through over a half a dozen review sites, the only two games that they had in common in their FX-8350 reviews was Metro 2033 and some flavor of Crysis. Otherwise I saw other games like Skyrim, StarCraft 2, World of Warcraft, Shogun 2... enough sample of video games that require different things out of the system and a lot of them used around 5-6 games. And the only way he managed to show that those games could "take on" a Core i5/i7 is if they were streaming. I also didn't like that he was using a source material that less than 1% of the population even uses (I'd like it more if he used 1920x1080).

This is an interesting data point and that's all it is.

Also you're supposed to bench CPUs on games with a high-end graphics card and everything turned down reasonably low. This is to measure how fast the CPU can keep up with the GPU. The higher performing the CPU is, the less the GPU is waiting on the CPU. The TweakTown chart had everything on high... which you might think "Well duh, that's what I'll be playing on." Except since all games are GPU capped, not CPU capped, benching the CPU with graphics turned down is an indication of how long you can live with that system before you need a total system upgrade, versus having to only upgrade the GPU.

You're actually kind of making my point from a different angle.

Also, the video noted that the FX-8350 was faster in almost all streaming situations.

And, most of the games it benched it was faster in as well. They just didn't bench the Crysis / Metro 2033 / Skyrim / Arkham group that every other bench site uses.

Like I said in the original post, you can't tell the difference between an FX-8350 and an i5-3570k in virtually any game unless you have a ~$500 video card. If you look at the difference in say, AnandTech's bench setups and TweakTown, Anand uses a GTX 690 and Tweak uses a GTX 580.

And if reviews are supposed to simulate real life situations, why would one go below 1920x1080 max settings on a GTX 690?

I think you're wrong about typical users playing at 720p. Most play at 1080p. That's what most monitors being sold work at now.

You can't tell the difference between platforms on a 580. I would say most folks don't even have a 660 class card, and a 580 is a sight faster than the 660.

Then when you turn to things like Cinebench and many other common apps, the 8350 is in the running with the i7, passing up the i5 easily :

http://www.bit-tech.net/hardware/2012/1 ... 0-review/3

http://media.bestofmicro.com/Y/Y/357658 ... studio.png

http://media.bestofmicro.com/Y/1/357625 ... /fritz.png

http://media.bestofmicro.com/X/Q/357614 ... reader.png

http://media.bestofmicro.com/Y/F/357639 ... toshop.png

http://media.bestofmicro.com/Y/H/357641 ... emiere.png

http://media.bestofmicro.com/Z/1/357661 ... winzip.png

http://media.bestofmicro.com/X/P/357613 ... l/7zip.png

http://techreport.com/r.x/amd-fx-8350/qtbench.gif

http://techreport.com/r.x/amd-fx-8350/tc-aes.gif

http://techreport.com/r.x/amd-fx-8350/tc-twofish.gif

http://techreport.com/r.x/amd-fx-8350/7zip-comp.gif

http://techreport.com/r.x/amd-fx-8350/7zip-decomp.gif

I think the conclusion here really should be something more along these lines than what we see in the typical commentary : If you havea GTX 580, 670 or better, you'll get better FPS in most of the 'most popular' current gen games, but not all. In everything else, it depends a lot on what you want to do. But when you consider that you can save $60-$100 getting the FX, this is almost a no-brainer in my opinion, at least compared to any i5. The i7-3770k, now that's where AMD really doesn't have a good competitor.

Hahah what the shit... I actually tried to read all of it to see if I can make any sense of it.

Then I tried again.

AMD is bad. 1776 Piers...Murder Pills!

I liked Kleink's reply because it's really deeper than a discussion on some forum can get into cleanly. At some point it will always be like posting something in NFL.com's discussion. Fire the coach, move the right tackle to the left...it's like if any of it was even remotely accurate we'd all be NFL coaches or perhaps an owner. No insult meant, but it's HS football players thinking it's all the same in the big leagues.

I've been at the round tables. I can only sum it up by stating that AMD still has it's place, but for the majority of consumers Intel has what plants crave. The ATI purchase was probably the best thing AMD did to solidify their ability to be relevant at all and have some income. The news of Intel taking a hit is also relative...they have their fingers in many pies and a downfall doesn't mean a single thing about any one product they kick out. With AMD, it basically means one of their 2 products didn't do well. It's just more complicated from the outside trying to look in at the cogs meshing and figuring out why things fail or win. I broke in killing Tejas...think that went well? Pissed 'em all off and out. Things seem well though, right? Yeah...

I think that Tom's hardware is incredibly biased, I've read many of their articles and I always get the sense that they're trying to indirectly bash AMD. If you compare benchmark articles from Tom's to guru3d of the same hardware you'll notice a big difference in how they're represented.