AMD Bulldozer revealed!

[Glaucus]

Ok, so I've been waiting all summer for AMDs Bulldozer as my plan is to put together a brand new rig to do it all: gaming, development, video, photoshop, 3D, etc. The Intel offerings are nice but a tad costly and I've always been partial of AMD. So today I see the benchmarks. If you care you probably have seen them too. Not sure how to describe what I think of them. I think this video sums it up nicely:

 

[the_leander]

On paper the ideas behind bulldozer are quite clever, if you're building a server.

For a desktop however, it's an ill fit - one full core plus an extra integer unit to an fpu causes no end of bottlenecking for real world usage.

Couple that with per clock performance significantly lower than that of Phenom 2 and you're left with a chip that eats electric and doesn't do very much at all with it. Top end hex core Phenom 2's eat this thing alive and that's before we get to intel i7.

In short, this is AMD's netburst moment.

 

[Glaucus]

True. If it were released AS the Phenom II, to compete with nehalem instead of Sandy Bridge, well, that might have been different. But it seems too late now.

I have read elsewhere that this particular chip design maybe just a transition into something bigger. AMD is really big into APU, where they can leverage their GPUs in all sorts of ways. In the future we may see all their chips come with GPUs as standard and all FPU operations offloaded to the GPU. Not sure this would help in single threaded operations, but could make it the only real option for servers - which isn't a horrible niche market to be in. Some argue that the future for personal computing is ARM anyway.

 

[faethor]

Bulldozer is interesting. It appears to be a multi-threaded champ. Though I'd argue at this time I think for most people that's overkill on the desktop. i5/i7 seem to be better balanced for the end user. AMD claims Windows8 is going to have better BD support built in. I'm sure it won't gain much but another 5-10% or so when the next OS comes out is always nice.

Myself? Mostly HTPC uses and webbrowsing. We do some video and photo stuff but that's minimal usage. We didn't see the need to buy a sports car when the ford would do. So, when our 3 year old computer died a bit ago I replaced it with an AMD A8-3850 system. It's only been about a month but I'd say the Llano w/ built in GPU works great for our uses. Really I think the next upgrade would be an SSD rather than a BD or i7 series. ... Which I think kinda echos Glaucus' ARM statement. Lower powered system are adequate for most people.

My expectation is Xbox720 / PS4 will push computing ahead again. I think we'll see something like the BulldozerII w/ built in GPU for either of those systems. So late 2013, at best, likely 2014.

 

[the_leander]

Tbh when this rig goes it'll likely be replaced by a dual core Atom setup. I just don't need all that much power for what I do these days. I don't think I've done a single piece of video transcoding in 5 years and as you say, an SSD or a hybrid drive would be more than adequate for my usage.

As for mobile stuff, I'll probably get an Asus Transformer. I figure by the time I've saved up for it it'll have gone down in price a touch and Android 4 will have been released.

 

[Glaucus]

Lower powered system are adequate for most people.

Well, with cloud computing, this may be more and more true. And in a funny way, because of cloud computing, we may all be using Bulldozers anyway, even if our clients use ARM.

 

[JoBBo]

Bulldozer is interesting. It appears to be a multi-threaded champ.

For AMD´s sake, I wished that was true. Unfortunately, performance is disappointing even for applications that are known to benefit from additional cores / threads although the difference in performance is not as bad as for single threaded applications.

http://www.anandtech.com/show/4955/the-bulldozer-review-amd-fx8150-tested/7

The 3ds max benchmark results pretty much speak for themselves. Champ? Not exactly, as I am sure you will agree.

Though I'd argue at this time I think for most people that's overkill on the desktop. i5/i7 seem to be better balanced for the end user.

Actually, they just seem to be better in pretty much any scenario. Irregardless if you are professional user or an end user, the Intel i5 2500K is consistently faster than AMD´s FX 8150 but costs around 30% less (210 vs 290 USD).

My expectation is Xbox720 / PS4 will push computing ahead again. I think we'll see something like the BulldozerII w/ built in GPU for either of those systems. So late 2013, at best, likely 2014.

It is difficult to say what the next generation of gaming consoles will be like as Sony and Microsoft will be facing an interesting challenge. On the one hand, Nintendo has shown how successful you can be with an ultimately weaker, but more affordable hardware platform that is clearly geared towards casual gaming. On the other hand, gaming consoles face immense competition from tablets and smartphones now especially with regard to casual gamers. Unless the new consoles can sufficiently differentiate itself on the hardware level (read: offer much better performance and more sophisticated input methods), they will likely lose market share as consumers will be unwilling to buy yet another device for marginally better graphics in games.

 

[minator]

It's a completely new architecture so it's entirely surprising it's not that good. I imagine the next rev should improve it significantly.

It's clearly designed for servers though. Be interesting to see how it does on server benchmarks.

Well, with cloud computing, this may be more and more true. And in a funny way, because of cloud computing, we may all be using Bulldozers anyway, even if our clients use ARM.

ARM are looking at the server market these days...

 

[Glaucus]

Well, this particular chip isn't designated for servers. There will be a server specific bulldozer with improved hypertransport.

Also, the bulldozers reviewed here were running on Win7 which is not optimized for the architecture. Win8 is supposedly going to have optimizations for it and AMD says that you could see a 10% improvement. Piledriver which comes out in Q2/Q3 of 2012 should add another 10%-15% per core performance. Add those two up and it suddenly looks a lot better, however, it'll be facing Ivy Bridge by then. Only hope for AMD is for Intel to run into problems with the 22nm fabrication - always a possibility but not something I'd want to bank on.

One other thought. I've only seen reviews of the 8 core BD. The lower core variants are supposed to run at higher clock speeds. This is a chip the needs Hz more then anything, and it's possible that if you want to play games you DON'T want the 8 cores. If the 4 core variants can really pull off the higher frequencies, then those might actually be proper contenders (especially if they cost a lot less).

 

[redrumloa]

Well, with cloud computing, this may be more and more true. And in a funny way, because of cloud computing, we may all be using Bulldozers anyway, even if our clients use ARM.

Cloud computing is evil commie pinko horse crap.

 

[the_leander]

Well, this particular chip isn't designated for servers. There will be a server specific bulldozer with improved hypertransport.

The whole basis for having extra integer units bolted onto the pipes is solely due to it being at it's heart a cpu designed with servers in mind. On a desktop chip it simply doesn't make a whole bunch of sense given the workloads. This is especially apparent when you factor in the frankly appalling FPU performance.

Also, the bulldozers reviewed here were running on Win7 which is not optimized for the architecture. Win8 is supposedly going to have optimizations for it and AMD says that you could see a 10% improvement.

10% on a synthetic benchmark from a dubious source (I take any claim made by a cpu/gpu manufacturer along these lines with a seriously big pinch of salt). I real world performance, you'll be lucky to get 3%. Frankly you'll get more by overclocking ram and cpu speeds.

And even if by some miracle a combination of patching of kernels, BIOS updates etc etc it actually achieves this 10% boost across the board it'll still only be on par with the current hex core Phenom 2 which even now is monstered by top end i5's in some tests and by high end i7's in everything.

Piledriver which comes out in Q2/Q3 of 2012 should add another 10%-15% per core performance.

Which would just about get you to where Phenom 2 is now.

Add those two up and it suddenly looks a lot better

Thing is, you can't simply add them up like that real world workloads, especially games will show up the architecture's inherent weaknesses no end.

One other thought. I've only seen reviews of the 8 core BD. The lower core variants are supposed to run at higher clock speeds. This is a chip the needs Hz more then anything, and it's possible that if you want to play games you DON'T want the 8 cores. If the 4 core variants can really pull off the higher frequencies, then those might actually be proper contenders (especially if they cost a lot less).

Here's the thing, it's not an 8 core chip. It's 4 cores, with 4 extra integer units bolted onto it, these are then both tied into a single fpu which causes all the bottlenecks. The "4 core" variants are 2 full cores, with 2 extra integer units, the same fpu bottlenecks apply. Higher frequencies will help, but only so much, per clock it is significantly poorer than either Phenom2 or any of the i series processors currently on the market.

 

[Glaucus]

I agree with most of what you said. But as for the Win8 optimizations, the 10% figure given by AMD was not a per-core performance boost (or at least, I don't think it is). The optimization is in how the OS decides which "core" the thread will be assigned to. Like you said there are bottleneck issues, but these can be optimized out to some degree. It's sorta like how hyperthreading works - if you don't optimize for it at compile time performance can get worse. AMD's multi-core scheme instead requires OS optimizations. Although this link doesn't prove what I'm saying, it does support it: AMD FX "Bulldozer" Review. This guy actually disables the 2nd core in each module (he calls them computation units, or CUs) and notices that performance actually goes up. This supports AMDs statement that the FX chips need OS optimizations to take full advantage of the architecture.

Which would just about get you to where Phenom 2 is now.

For single threaded apps, yes.

Thing is, you can't simply add them up like that real world workloads, especially games will show up the architecture's inherent weaknesses no end.

Of course, and the benchmarks they used to test weren't real world tests either. Perhaps they should run multiple benchmarks at the same time to better test it's multi-threading abilities. Realistically, when transcoding a DVD one is likely also doing something else, like perhaps watching another movie or playing a game. How does the Bulldozer handle that? That might provide very different results, but running such a benchmark isn't easy.

Here's the thing, it's not an 8 core chip. It's 4 cores, with 4 extra integer units bolted onto it, these are then both tied into a single fpu which causes all the bottlenecks.

Yes I know, but talking about is awkward. It's true, comparing cores between Bulldozer and i3/5/7 or even Phenoms is comparing apples to oranges. Perhaps we need to talk about threads but then hyperthreading makes it a funny comparison as well (both the i7 and FX chips can handle 8 simultaneous threads). But yes, per clock it's not that great, but it's a design that was supposed to hit much higher clocks. Some say the fabrication ran into snags and they failed to meet their clock targets. Perhaps later chips will come with significantly higher clocks. No idea. But right now, that's all AMD has left. Is it any surprise that for the first time ever AMD has released a water cooling device for their CPUs - which by the way are all unlocked? It seems to me they really want you to push the clock rate up as high as you can take it. The down side is that water cooling adds about another $100 to the chip making it purely a fan boy product.

 

[faethor]

Didn't the bd just get awarded by Guinness as the highest GHz CPU during over locking?

 

[the_leander]

I agree with most of what you said. But as for the Win8 optimizations, the 10% figure given by AMD was not a per-core performance boost (or at least, I don't think it is). The optimization is in how the OS decides which "core" the thread will be assigned to. Like you said there are bottleneck issues, but these can be optimized out to some degree. It's sorta like how hyperthreading works - if you don't optimize for it at compile time performance can get worse. AMD's multi-core scheme instead requires OS optimizations. Although this link doesn't prove what I'm saying, it does support it: AMD FX "Bulldozer" Review. This guy actually disables the 2nd core in each module (he calls them computation units, or CUs) and notices that performance actually goes up. This supports AMDs statement that the FX chips need OS optimizations to take full advantage of the architecture.

I've read that and other reviews, I've also read the response from the linux kernel developers. Yes, you can get a small amount of performance increase from optimisation on the OS side. But not by much (realistically you'll not likely see 10% extra out of this chip by low level tweaking) and it will still only help with some programs.

For single threaded apps, yes.

Problem is, as the people behind BeOS discovered, there are a lot of tasks that are ill suited to parallelisation and trying to force them actively hurts performance.

Of course, and the benchmarks they used to test weren't real world tests either. Perhaps they should run multiple benchmarks at the same time to better test it's multi-threading abilities. Realistically, when transcoding a DVD one is likely also doing something else, like perhaps watching another movie or playing a game. How does the Bulldozer handle that? That might provide very different results, but running such a benchmark isn't easy.

I've seen games benchmarked and this chip seriously hurts them. Perhaps disabling the extra integer units would help those scores, but you're still stuck with a chip that per clock does less than it's predecessor and still needs the OS to be modded to theoretically get anywhere near...

Yes I know, but talking about is awkward. It's true, comparing cores between Bulldozer and i3/5/7 or even Phenoms is comparing apples to oranges. Perhaps we need to talk about threads but then hyperthreading makes it a funny comparison as well (both the i7 and FX chips can handle 8 simultaneous threads). But yes, per clock it's not that great, but it's a design that was supposed to hit much higher clocks. Some say the fabrication ran into snags and they failed to meet their clock targets. Perhaps later chips will come with significantly higher clocks. No idea. But right now, that's all AMD has left. Is it any surprise that for the first time ever AMD has released a water cooling device for their CPUs - which by the way are all unlocked? It seems to me they really want you to push the clock rate up as high as you can take it. The down side is that water cooling adds about another $100 to the chip making it purely a fan boy product.

The water cooler they're supplying doesn't look like it'd offer much over something like the old Evo33 I had on my Athlon64 back in the day. A good cooler with air pipes and a large fin stack will beat the pants off of an el cheapo water cooler.

And I've had both in my time and took my AMD64 - 2000 running at 2ghz all the way up to 2.4ghz stable and likewise with an older AthlonXP running at 1.6Ghz I managed to get it to crank up to 2ghz with water cooling, the only thing that stopped me pushing further was that the motherboard bios wouldn't let me.

 

[the_leander]

Didn't the bd just get awarded by Guinness as the highest GHz CPU during over locking?

Dunno, if they have, is that by conventional means or by exotic?

 

[Glaucus]

Didn't the bd just get awarded by Guinness as the highest GHz CPU during over locking?

Yes, can't remember the actual clock rate, but I remember reading they did it by disabling 7 of the 8 "cores" to do so. Not sure if that's cheating or just how it's done, but overall doesn't seem practical. But who knows, it might run games really nice.

 

[Glaucus]

Here's an article: AMD Sets CPU Highest Frequency Record by Overclocking Eight-Core FX "Zambezi" Chip to 8.429GHz.

And it seems two cores (or one module) was functional.

 

[faethor]

Thanks Glaucus on that link. Disabling to 2 cores isn't great but even seeing 2 cores above 8.4Ghz is fairly impressive. I can imagine once they drop small nm construction that'll get better.

Also, JoBBo posted the Anandtech link. On that last page is some game testing with the developer's edition of Windows 8. The games see a 2-10% improvement for the Bulldozer.

Though to be fair let's include all the cavots. Win8 isn't fully cooked. Other chips in Windows 8 weren't tested perhaps they gain too. The games were written for Win7. Any testing gives one a fairly good rough estimation differences in various hardware will make different impacts. Etc.

Bulldozer is okay. The pricing seems fair. I'd hoped for better. Though without AMD I wonder if Intel would improve themselves. SandyBridge takes a few hints from Phenom, for example.

 

[Glaucus]

So my dilema now is, which platform to buy? My plan was to buy a high end AM3+ motherboard with a cheap PhenomII CPU, with the roadmap of upgrading to a Bulldozer at some point. However now it seems there's probably little point in that, although I could still upgrade to Piledriver (AMD has promised the next chip will remain AM3+ compatible). Basically I want to build a machine that's not overly expensive but that I can continue to upgrade here and there. My current rig is so old (AthlonXP-M 2500 OCed to 2.2Ghz on an ASUS A7V266 MB w/ Radeon 9800Pro) I'll be pretty much tossing everything except for maybe the case and 550W modular PSU (the PSU is only a few of years old). Going with the Intel i5 2500K would cost me about an extra $100, but I know little about Intel's roadmap regarding the 1155 socket - although I imagine I'll be safe for a while to come. I suppose it's probably worth the extra initial cost as I'll probably save in the long run and still have options for upgrades (there will always be the 2600K after all). The intel route will also give me the option to run a hackintosh if I ever wanted to do something like that... I guess there's little reason to stick with AMD this time around. Darn.

 

[Glaucus]

Dilemma over. I pulled the trigger on an i5 2500K along with the ASUS P8Z68-V Pro motherboard, along with 8GB G.Skill ram and an WD 750GB Black drive (with plans to get next generation SSDs at some point). I slapped it together last night but it was getting late and decided to wait until today before I hook it up to the mains. Ya, no video card yet, I'll toy around with the built in crappy Intel graphics - until I save up for a high end AMD graphics card that is. The ASUS is supposed to overclock very nicely (pretty much does it for you) and I should be able to hit close to 4.5Ghz. Ya I feel cheap and dirty for not buying the AMD, but I'm sure I'll feel much better once I load up Eclipse with the Android emulator running Honeycomb. Plus now I have the opportunity to not just run a pirate windows machine, but a pirate MacOS machine as well!