Table of Contents
Bulldozer. The name is fraught with implications of power. Pure, unadulterated, run-you-over power. Much has been said and debated about AMD’s newest offering over the past year plus. Now, the wait is over and we’re here to bring you the full scoop!
The Bulldozer Architecture
AMD has really reinvented the wheel with Bulldozer. It’s a completely different animal than other CPUs out there (we’ll see later whether that’s a good thing or a bad thing). With Bulldozer, AMD has taken a less-is-more approach. Their engineers put two cores side-by-side and considered what could be shared between them.
They decided to share the front end, floating point cores, and the L2 cache between two cores. It’s not shown in this diagram, but each module has its own dedicated L3 cache in addition to the intra-core L2 cache.
The next four slides just bore a little deeper into the Bulldozer shared and dedicated components, giving a bit more detail to the diagram on the right above. What may give you pause is the shared floating point core. While most applications do use CPU cores, those that use FPU cores would seem to be losing out on this deal.
There are still two 128bit pipes through which to process data. The front end splits 256bit FPU content into two 128bit FPU opperations and sends them down the pipes. These ops have to be processed at the same time. So, in essence, the FPUs can do either two seperate 128bit FPU operations or a single 256bit FPU operation. So, what if your application wants to run more than four 256bit FPU operations? With the loss of the extra FPU core, the lack of ability to split the workload further would indicate potential performance loss.
Bulldozer is the architecture, but the code name for the CPU we’re looking at today is actually Zambezi, the desktop variation of Bulldozer. All of AMD’s new products use Bulldozer cores on both the server and the desktop sides of the market.
Power management is a big deal for AMD, especially when it comes to the server market. It never hurts for the desktop side to reap the benefits though, and that’s what we have here. AMD has reduced idle frequencies to a meager 1.4 GHz with the FX-8150 we’re looking at today and reduced the idle Vcore to match. The changes between states is instantaneous as you would expect.
Consequently, while power isn’t a huge deal to many overclockers in general, it is worthy of a mention. Here is how this setup compares to an Intel setup with a 2600K and the same GPU. There is definitely a strong disparity in power consumption. Sandy Bridge is clearly much more power-efficient than Bulldozer.
|Test Setup||Idle (Watts)||CPU Loaded (Watts)|
|i7 2600K||97 W||158 W|
|FX-8150||121 W||246 W|
They also have advanced Turbo Core, which has a base turbo and a max turbo. All cores can exceed their stock frequency (up to 3.9 GHz in this case) if the chip’s TDP isn’t reached. However, that didn’t seem to happen much in our testing. If a stressful multi-threaded load is applied, chances are it will take up the TDP and the CPU will operate at its base frequency (3.6 GHz on the FX-8150).
Lightly-threaded loads do take advantage of Turbo Core greatly though. With a max turbo of 4.2 GHz, single- up to quad- threaded applications (which comprise a lot of consumer applications, sometimes even games) get a healthy 600 MHz boost over the base frequency.
Eventually there will be eight different Bulldozer iterations to choose from. There are actually seven different models, but the FX-8120 will be available in both 95W and 125W TDP variation.
At launch though, there will be four Bulldozer CPUs to choose from, the flagship FX-8150, the next step down FX-8120, the six-core FX-6100 and the quad core FX-4100. Telling of AMD’s cut-throat pricing (and, potentially, performance), the top of the line FX-8150 is going to be priced at a very reasonable $245.
With the technical demonstration out of the way, we get to the AMD
marketing launch presentation. On the left, they compare the technology between the i7 and i5 family of CPUs with the FX-8150. Let’s have a list, shall we?
- RAM speed you can dismiss most of you know by know that Sandy Bridge has a stellar memory controller that runs DDR3-2133 without changing a single setting other than setting the RAM timings & voltage.
- The CPU spec comparisons can pretty much be dismissed as ‘we’ll see how it performs’.
- The second most important comparison in this chart for overclockers and gamers is the fact that AMD has a full 32 lanes of PCIe graphics capability.
- The single most important spec is that all FX processors will have unlocked multipliers. Fun for everybody, regardless of budget!
- The slide on the right just reiterates some of these points with bigger text.
Of course, both sides in the CPU battle like to present benchmark graphs, mostly aimed at showing how their products come out ahead. On the left shows a very favorable comparison to the i5 2500K when gaming with Eyefinity. We’ll have to take their word for that one. We’ll be exploring several games but will do so using the more prevalent 1080p resolution.
On the right is a slide you’ve seen floating around already from a leaked presentation. We can show you now it was a real slide and does accurately reflect Bulldozer’s performance. There is a difference in this slide compared to the leaked slide though – in the previous version, WPrime performance was graphed incorrectly. See folks? This is why you take anything – even potentially official leaked slides – with a grain of salt prior to the NDA expiration!
Now we see a little of how Turbo Core can benefit vs. the base frequency. On the right is a gaming comparison between a system running a 980X and an FX-8150. I’m not sure what to think about these tests. Sure, it can game at 1080p right alongside a kilo-buck CPU, but so can Intel’s cheaper Sandy Bridge offerings. So take heart, if you’re a gamer and won’t use that thousand dollars worth of CPU, you don’t need to spend that much on a CPU. Whew, it’s a good thing they saved all that money for you isn’t it?
Now we get tot he exciting part of the presentation – the future! The left slide speaks in generalities – things you’d hope they would do anyway. On the right we get some specifics, and it looks pretty encouraging. You saw in the slide above that the FX-8150 is a worthy competitor to an i5 2500K but didn’t quite catch the i7 2600K. A 10-15% improvement with Piledriver should close that gap to almost nothing. Intel isn’t standing still though, so AMD will have to get that IPC up to boot. As a side note, I like AMD’s architecture naming scheme – large construction equipment!
Well, there you have it folks – the Bulldozer architecture. It’s definitely a big step away from the norm. Can it keep up with Sandy Bridge? We’ll see soon. It is important to consider that AMD has been stressing for a while with reviewers that the FX-8150 is priced to compete with Intel’s i5 2500K, not the pricier i7 2600K.
There are a lot of people out there that expected FX to destroy Intel’s current offering. As you can see above, that’s not where they are positioning this chip. Whether that’s a failure on their part is for you to decide – what we’re going to consider is whether this chip performs to justify its price. No more, no less.
The CPU – Now With Water!
AMD sent a nice, large press kit containing the board on which to bench their new CPU, the CPU in the socket, a box for photos and (for some reason or another) a belt buckle. Not too many people I know wear giant belt buckles, but to each their own.
AMD has switched up their packaging with this generation of processors and will be shipping their FX CPUs in pretty cool looking metal tins. Ours came sans cooler, so we just have an empty box, but this is what they’ll look like on the outside.
One very interesting item is that, for the first time to my knowledge, AMD is going to have a liquid cooling solution to sell with their FX CPUs. They aren’t hiding the fact that it’s an Asetek cooler, but it has a nice, thick radiator to hopefully compensate for it being a single-120mm radiator solution. It won’t get temperatures like you’ll see below using a custom water loop, but will be a fair sight better than the stock cooler. This thing just arrived today so there can’t be any testing, but we did snap some photos to show you what to expect.
The cooler will be rolling out to select markets “soon” (the US is not first, FYI) so don’t expect to buy it right when you buy a Zambezi CPU. Expected retail, subject to change, is in the $100 range.
Now we bring you the main event. Weighing in at ~2 billion transistors in an area of a mere ~315 mm2 is AMD’s Zambezi, the Bulldozer-based FX-8150!
Please hold your applause until the review is complete. No, really. Quit clapping. Thanks!
Overclocking for Stability
The brightest spot about Bulldozer is its ease of overclocking. It’s just begging to be clocked. After dialing in very reasonable numbers…
- 250 MHz bus
- 2500 MHz CPU-NB and HT speeds (with a minor voltage bump)
- Rated RAM speed & voltage
- 1.41 V loaded Vcore
…I started bumping the multiplier and ended up with a very nice 24/7 overclock of 4.75 GHz. That’s 150 MHz higher than I settled on with Sandy Bridge, so consider me pleased.
Note that CPUz is reading this CPU incorrectly; it truly is an FX-8150 as you can see on the chip photo above. Temperatures are what you can expect from a custom water loop containing a Swiftech MCR-320 radiator, MCP35x pump (with 35x top) and EK HF Supreme Cu. Temps under load at stock were great, running in the low-to-mid 30° C range.
After that I reached 5.0 GHz within ten minutes and ran WPrime 32M without breaking a sweat.
Since this setup was going to be put under liquid nitrogen I didn’t see how far it could be pushed under water. That said, at this overclock it was getting kind-of warm (assuming older architecture temperature rules still apply) and going much farther on this chip without sub-ambient cooling may not be too wise.
Test System, Opponents and Methodology
We’ve got a who’s who of modern mainstream systems for you today. Considering they’re not exactly relevant for current purchasing decisions, we’ve dropped the plethora-of-CPU graphs in favor of easier-to-read-and-decipher graphs with only modern processors.
|CPU||AMD FX-8150||Phenom II x6 1100T||Intel i7 2600K||Intel i5 2500K|
|Stock / Turbo||3.6 / 4.2||3.3||3.4 / 3.8||3.3 / 3.7|
|Motherboard||ASUS Crosshair V Formula||ASUS Crosshair V Formula||ASUS P8P67 WS Revolution||Gigabyte G1 Sniper2|
|GPU||AMD HD6970||AMD HD6970||AMD HD6970||n/a|
The UEFI used was ASUS version 0813. It was the most recent version available when the setup was tested and was supplied by AMD. The OS for testing was Windows 7 x64 with all updates and patches installed. The CPUs don’t appear in every graph below. Notably, the i5 2500K was (very kindly) tested by our esteemed editor EarthDog and he did not have an HD6970, so there are no 3D/game results with the 2500K.
The stock benchmarks were run three times each and the results you see are averaged. The only exceptions being 3D benchmarks, game tests, and overclocked benchmarks, which were run once each.
The results you see below are graphed relative to the AMD FX-8150’s stock performance. This means that results by the FX-8150 at stock all equal 100% and the other results are graphed as a percentage relative to the FX-8150’s performance. So, for instance, if the FX-8150 scored 200 points on a benchmark and the i5 2500K scored 180 points, on the graph the FX-8150 would = 100% and the i5 2500K would = 90%.
Enough talk, let’s bench this thing!
AIDA 64 Benchmarks
First up, we explore the AIDA 64 test suite. These tests were run only at stock to give a comparison of how the chips perform under various testing conditions.
Starting with the CPU tests, it seems AMD’s positioning this chip against the 2500K is accurate. The two chips trade blows throughout the CPU test suite.
Floating point performance isn’t looking so good. Not that it’s bad compared to the professed competition, but that they actually lost ground to the Thuban. I mentioned this might happen based on the sharing of FPU cores rather than duplicating the entire core, which would have given the chip a healthy boost with eight FPU cores instead of four. It seems those fears have been realized and the Thuban – with six FPU cores – out-performs Bulldozer. For better or worse, AMD has put all their performance eggs on the CPU side of their chip.
Memory is one area where Bulldozer does quite well for itself. They are within spitting distance of Sandy Bridge (which has a stellar memory controller) for reads and copies but lag behind a bit with writes. Still, AMD has gained significant ground over Thuban.
We’ll get the bad news out of the way first. AMD gave up on SuperPi. That’s not new, it happened a while ago. SuperPi uses an old x86 floating point instruction set. There it is again – floating point. Not only did they not improve with this benchmark, they lost ground to their older architecture, which isn’t a surprise considering the AIDA FPU tests above. Ground was gained when overclocked, but not enough to say this chip is anything close to good at SuperPi.
But wait – Bulldozer focuses on multi-core performance, so it should do well in WPrime, right? Wrong. AMD would say this is an older benchmark. I would say phooey. This processor is supposed to be the epitome of multi-core performance. While it still beats its stated competition (2500K remember), it loses ground to the Thuban in a multi-core benchmark.
Overclocking the chip (to 4.75 GHz remember) brings its performance right next to the i7 2600K’s stock performance. To say running this benchmark was disappointing would be very much an understatement. Those that were shown these benchmarks with the Overclockers staff had comments ranging from “there must be something wrong” to simply “this can’t be”.
That said, these results need to come with a bit of a caveat. Believe it or not, in a way they are supposed to look like this. Both SuperPi and WPrime calculations are executed in the FPU cores. There are only four FPU cores among the four modules, as opposed to eight CPU cores. With the loss of two FPU cores, it makes perfect sense that the Thuban out-performs Bulldozer in this benchmark. So while FX is a multi-core powerhouse, as feared the Bulldozer architecture that saved space (and saved TDP) by removing four FPU cores hurt performance when performing these calculations.
I presented that last paragraph to AMD to get their take; this is what they have to say on the matter:
It’s really more than the FPU as the Bulldozer core can handle two 128-bit FP calculations or one 256-bit. These applications use old extensions (really, yesterday’s tech), while Bulldozer is optimized for the workloads of today and tomorrow. We expect that the applications will be updated with new extensions which will dramatically improve performance (or become obsolete to applications that do get updated). As well, when we begin to see other optimizations like the Windows 8 scheduler, Bulldozer will show to be a great building block for our upcoming products.
Rendering, Video Conversion and Compression
Real world performance. That’s where AMD is hoping to have their moment in the sun. In the case of both Cinebench rendering tests and 7zip compression tests, I think we can happily hand it to them.
The FX-8150 beat out its predecessor and the 2500K at stock and put a good lickin’ on them overclocked, also beating the stock 2600K. That said, to require over a GHz overclock to beat a stock 2600K is not what you’d call a good sign.
How about some more rendering with some video encoding thrown in?
Well, PoV Ray is just that – a ray of sunshine through the clouds. Stock just nudged past the stock 2600K and overclocked jolted performance up impressively. x264 was a mixed bag, with Pass 1 requiring the overclock to gain parity with the Intel offerings. Pass 2 is much better for AMD, showing stock performance right on par with the stock i7 2600K and then leaping ahead overclocked. What’s important about Pass 2 is that it’s the actual encoding of the video. Pass 1 is just a scanning pass (see the x264 FAQ). So when you’re actually encoding video, the Bulldozer chip comes out smelling like roses.
Now, AMD supplied reviewers with a new version of the x264 Benchmark late in the game that was supposed to take advantage of the XOP and AVX extensions. Regrettably that didn’t make it into testing because the build was torn down and being insulated for sub-zero benching. If that is a different build than the one we used (which was the most recent build available), you can expect encoding using those extensions to improve even more.
So in real-world use, Bulldozer isn’t so bad – it’s good even. At stock in each test it is either close to or just below the i7 2600K’s performance and overclocked it overshoots the more expensive i7 by a good bit.
Well, can 3DMark save Bulldozer for the benchmarking community, at least those going for globals? We’ll start with the oldest of the three benched here – 3DMark06
Um, err…to answer the question, no, it won’t. Yikes. This was just pitiful, I think the results speak for themselves. 3DMark06 does take advantage of multiple cores, but not as much as newer versions. Benchmarking team member dejo has said multiple times how he gets the same scores with HT off as he does with it on, so maybe this isn’t as bad as it looks. Hey, at least the FX-8150 did okay in the CPU test itself – when overclocked to the moon.
How about Vantage?
Nope, not here either. Things are not looking so hot for the DirectX 10 crowd either. Vantage can definitely take advantage of all available threads, but sheesh, this isn’t even close; and the Thuban continues to beat Bulldozer.
So…maybe 3DMark11 will look better?
Ahh, now that we have a heavily GPU bound benchmark, the Thuban finally gave way to its younger brother, but just barely. Last up in the GPU benchmark family we have HWBot Heaven, another GPU-intensive activity (the Thuban was not tested for this bench).
Well, at least that – on Heaven, which shows little to no variation between CPUs anyway (which is obvious from the stock-to-overclocked results), the FX-8150 is dead even with the 2600K.
Those looking for global boints from 3D benchmarks will unfortunately have to look elsewhere. It seems the 2600K is still the go-to for that. Bulldozer did make up ground when overclocked, but really not that much. I had really expected a greater-than-one GHz overclock to at least bring the FX-8150 into actual competition with the 2600K but that’s just not the case, especially considering Sandy Bridge’s hilariously easy overclocking.
Now we move on to games, an area where AMD is touting Bulldozer as a top performer. All game tests were run at 1080p with graphics options turned to the max. We’re going for the most average gamer with a good video card and a single 1080p monitor. This is how it will perform for most people; we see no value in running low-resolution benches to show something no one will ever experience.
First up, the Stalker: Call of Pripyat benchmark.
Here again, Bulldozer continues to trade blows with Thuban. It’s not trading blows with the competition, it’s trading blows with AMD’s old architecture. Overclocked, it gets close to the 2600K’s numbers. There is little difference here as you can see, but the difference is definitely there and it doesn’t really improve Bulldozer’s outlook.
We’ll test three more games and be done. Alien vs Predator is the only one with a Thuban result, the other two weren’t run with the 1100T.
More of the same here. I double-checked the results and AvP & Dirt 2 did actually lose a hair of an FPS there, but that’s small enough to be considered just test variation. Put succinctly, overclocking your CPU won’t do much when running those two titles with an HD6970. Hawx 2 does show improvement overclocked, enough to get much closer to the 2600K’s performance.
Truthfully, based on this Bulldozer will be fine in a gaming system. Using one of these chips won’t hurt a gamer in any noticeable way. It also won’t really help a gamer in a noticeable way. If we had a 2500K to test in gaming, it’s safe to assume we would have shown Bulldozer performing right alongside it. I think the best way to describe gaming performance is on par with its stated competition.
Ahh extreme benchmarking. AMD’s saving grace. Before launch they had several members of the reviewing press out to give the tech demo at which they broke the world frequency record. It’s definitely an impressive feat in itself, so I came into this review really excited to get my hands on a chip!
Then, I benchmarked it and was somewhat underwhelmed. It did overclock well on ambient though, so there was hope that it would be a great liquid nitrogen infused experience. I was not disappointed, thus the passion was reignited and the insulation process begun.
Some of you may know Giraffe Pot already. It’s huge. With the extension on, it’s darn near the same height as the thermos that was used to pour liquid nitrogen (LN2) into it! This pic is with the thermos at board level.
Anyway, I got the software installed and pulled the beast right on down to its full -196° C because as we know from the world record attempt, Bulldozer has no cold bug! It got so cold, it formed an icicle where the probe emerged from the pot insulation/tape (see the above pic on the left).
I love the smell of nitrogen in the evening (seriously though, don’t inhale the stuff; it’s odorless and can asphyxiate you). These were taken while running WPrime 1024M with a full pot.
So, how did it do after about three hours of torture? Not too shabby. I had hoped for more but one or two of the modules in this CPU were much weaker than the first two. Multi-threaded benchmarking could “only” be had in the 6.2-6.5 GHz range. WPrime 1024M passed at 6271 MHz and WPrime 32M passed as high as 6528 MHz.
You can’t see it with CPUz, but CPU-NB and HT were both at 2,500 MHz for the entirety of this cold run. Time didn’t allow experimenting with those as my meager hours were all spent trying to squeeze the most out of the CPU.
That was multi-threaded and at those speeds Bulldozer made a comeback for competitive benchmarking. Regrettably it’s still not enough to topple an i7 2600K. At 5.4 GHz I managed 4.515 seconds, more than 0.3 sec faster at over a GHz less. Looks like we’ll need to be in the 7.0+ GHz range with Bulldozer to make a dent in Sandy Bridge times. As you can see, that’s not necessarily going to be easy.
Thus, we continued hunting out max frequency with single threaded benchmarks. SuperPi times start getting half-way decent when you start running them over 7 GHz – and run them it did. SuperPi 1M was able to complete (barely stable) at a very impressive 7507 MHz.
If you didn’t look too closely at that screenshot, look again; specifically the memory clocks. That’s right folks, when this memory controller is cold (and it has to be cold!) it can run some really strong memory. In my case it was running at DDR3-2416, but I’ve heard of at least one instance of an FX-8150 running some insane speeds of DDR3-3000 at ASUS HQ.
Last, but not least, we have the maximum speed – the highest frequency CPUz could save a validation at without crashing. It’s the hardest and the simplest “benchmark” all at the same time. Based on my result, I have a feeling this was an ‘average’ chip based on the ease with which they binned the few used in the world record attempt. Alternatively, it could be the fact that I wanted a living, happy chip after benching it cold – meaning I didn’t put 2+ volts through it like they did. Anyway, the maximum frequency at which this chip could validate was a whopping 7623 MHz!
Check out the validation link too – it did it with the memory even higher at DDR3-2452.
While the scores themselves aren’t exactly show-stopping, the clocks definitely are. On top of that, this was a very fun chip to bench. There really is nothing like filling your pot to the brim (less a bit to allow for boiling without splashing!) and going to town torturing the newest CPU to market.
Final Thoughts and Conclusion
Please read this conclusion to the end. I’ll go ahead and spoil the surprise and say this CPU gets an Overclockers Approved. There, now you have no reason to skip to the logo.
This CPU is a mixed bag, start to finish. Like I said, AMD reinvented the wheel and took a big chance. It took years to get to this point. Their engineers poured blood, sweat and tears into this architecture. It wins in some places and loses in others. Overall, it’s an improvement from their previous Stars-core derived Thuban. In a few places it falls short, but where it counts for most people – real-world applications like rendering, compression and encoding – it’s a strong step forward, competing perfectly in the market segment where AMD’s pricing is positioning it – against the i5 2500K.
Which is really the rub with many-an-enthusiast. Bulldozer has been talked about for a long time. The way a lot of people at overclocking forums talk about it, you’d think it was supposed to be the second coming. It has been argued about, hotly debated and speculated upon forever. Our own Bulldozer Rumors thread has been going strong since January 14 of this year. It is definitely not the Intel-toppler many thought / wished / hoped / begged it would be. Was I a bit disappointed? Absolutely. I bought into the hype just like many of you and AMD did not produce what a lot of people thought was going to be the return of top-of-the-hill FX.
That, of course, brings us to AMD’s tried-and-true formula: Price-for-Performance. It’s not priced to topple Intel’s higher-end mainstream i7 2600K. At $314.99 shipped, that chip still remains the mainstream enthusiast king-of-the-hill. Every single consumer Bulldozer FX chip will be an unlocked CPU priced at or below $245. Considering the gains seen against the i5 2500K, it’s worth the $25 difference in price.
With all of that said, the overclocker in me is just screaming so what if it just competes with the i5 2500K and doesn’t topple the i7 2600K?! It clocks like a madman. It does its job at the right price for the performance it offers. So, because we try to review based on what a product is relative to what it offers at its price point – and not based on what people with nothing but hopes and dreams expected and/or hoped it would be – it’s safe to call Bulldozer another solid offering from AMD.
– Jeremy Vaughan (hokiealumnus)
Author’s Note: Please see the first comment below. There is important information regarding the launch prices for these CPUs. I am changing nothing about the review above, but price point was a huge factor in calling this Overclockers Approved. Seeing the price at $280 when these went on sale made me feel duped. Anyway, regarding CPU purchasing decisions, please see the first comment below the article.