Bulldozer Benchmarked and Analyzed: Is AMD Back in the Game?

Dan Scharff

AMD finally catches up, sort of

AMD’s newest CPU is perhaps the worst kept secret in the industry. The company telegraphed the microarchitecture seemingly years ago but garnered much attention. That’s no surprise as the chip codenamed “Bulldozer” is considered AMD’s first true redesign of a chip since the original Athlon 64. Truth be told, there’s also a lot hanging on new chip as many are wondering if AMD still has any mogjo to go toe to toe with Intel’s processors. To find out, read on.

Just What is a Core?

The last year has led to pretty blurry lines as to what exactly a core is. Is it strictly x86? Do you count the integrated graphics portions? To add to that Jack Daniel’s blurred and slurred line is AMD’s new Bulldozer. Officially named FX (in a throw back to the glory days of the Athlon 64 FX-51!) the chip makes you wonder if what you thought was a core is still a core.

FX isn’t made up of cores, instead it’s made up of modules. Each module is built using two monolithic “cores.” Each core has its own set of integer schedulers, pipelines and L1 data cache. AMD says compared to Intel’s Hyper-Threading, which splits up the resources of a single core into two virtual cores, FX’s design won’t get as bogged down when it has to deal with multi-threaded workloads. On an Intel chip with Hyper-Threading, the core really only has resources for one core and multi-threaded loads must take turns running if the code calls on the same portion of the chip. That’s not the case with FX.

But AMD didn’t completely duplicate all the resources of a dual core in its module—a single floating point unit services both of the cores that handle integer workloads. Why just integer workloads? AMD says it believes that’s where most of the performance is to be made today.

AMD also says the modular design lends itself to higher performance when say, a single-threaded workload is thrown at a single module. That’s because the cores are so interconnected if only one core is working, some of the second core’s resources can be put toward that single-threaded workload.

AMD will launch four FX chips (two eight cores, one six core and a quad core) ranging from $115 to $245. The company’s top end part is the FX-8150 which is made up of four of dual-core modules on a single die. One potential performance issue AMD has already admitted could crop up on Windows 7 and older OSes is the scheduler inefficiencies. The scheduler should know to throw four threads at four different modules, instead of four cores on two modules for the highest performance. Unfortunately, Windows 7 and anything older isn’t capable of determining how to load an FX for the utmost in performance returns, AMD says. That may not happen until Windows 8 is released. Intel faced similar teething pains when Hyper-Threading was first released too.

AMD's new lineup

Base Clock / Turbo Clock / Max Turbo Clock
AMD FX-8150
3.6 / 3.9 / 4.2
125 Watts
AMD FX-8120
8 3.1 / 3.4 / 4.0
125 Watts
AMD FX-6100
3.3 / 3.6 / 3.9
95 Watts $165
AMD FX-4100
4 3.6 / 3.7 / 3.8
95 Watts

New Max Turbo and New Instructions, Too

AMD first introduced Turbo Modes with its hexa-core Thuban chips, aka Phenom II X6 chips. With FX, the company has refined its Turbo even more with a new Max Turbo mode that, well, maxes out the overclock. On workloads that hits all cores, all cores can be overclocked by 300MHz to 3.9GHz. On lightly threaded workloads, half of the cores can go to sleep while the other half can clock up to 4.2GHz.

Elsewhere in the chip, AMD has brought the new CPU to instruction parity with Intel. The FX processors will have AES instructions to support acceleration of encryption and decryption workloads and Advanced Vector Extensions, which Intel introduced with its Sandy Bridge lineup, is now also present. The old instruction set wars still run hot though as the FX will support the Fused Multiply Add 4 instruction set or FMA4. Unfortunately, Intel is only supporting FMA3 in its upcoming Ivy Bridge CPUs and has apparently cancelled plans to support FMA4. The little stand off may cause problems with developers as to which instruction set they support and how they support it. Who is at fault? Most observers say both companies are playing games. Our normal guidance is not to sweat it over new instruction sets because by the time software support is there, the first chips to support it are usually so old that it’s just easier to upgrade instead.

New chip, Same Old Socket, and Fully Unlocked

There’s one thing AMD gets right year after year—the same old socket. While Intel has shuffled through five sockets, AMD has pretty had just one. The only change though is the electrical underpinning of some of the sockets that’s made them incompatible. Still, many late model AM3 boards should be compatible with AM3+ FX chips (check with your motherboard vendor first of course) and even better, the mounts for the coolers have remained the same too so you can reuse your exotic cooler.

To cater to enthusiasts, AMD says all FX chips will be fully unlocked giving overclockers an all access backstage pass to overclocking. With Intel’s chips, only the “K” versions or Extreme Editions are fully unlocked. AMD has made overclocking a bragging point too and helped fund a team of overclockers to push an FX to 8.429GHz using liquid helium.

So what does this add up to? On paper, it looks like AMD has finally caught up – to some of Intel’s Sandy Bridge chips anyway. But what does it look like in benchmarks? For the answer to that, you’ll have to read on.

Top Chips Compared

AMD Phenom II X6 1100T
AMD FX-8150
Intel Core i5-2500K
Intel Core i7-2600K
Intel Core i7-990X
3.6GHz 3.3GHz 3.46GHz 3.46GHz
Turbo Clock (Max)
3.7GHz 3.9GHz (4.2GHz) 3.7GHz 3.8GHz 3.7GHz
125 Watt
125 Watt 95 Watt 95 Watt 130 Watt
Cores / Threads 6 8 4 4/8 6/12
Volume Pricing $205 $245 $216 $317 $999
Process 45nm 32nm 32nm 32nm 32nm
Total L2 Cache 3MB 8MB 1MB 1MB 1.5MB
Total L3 Cache 6MB 8MB 6MB 8MB 12MB
Die Size 346mm2 315mm2 216mm2 216mm2 240mm2
Transistor Count 904 million
2 billion
995 million
995 million
1.17 billion
Socket Socket AM3
Socket AM3+
Memory Controller Dual Channel DDR3/1333
Dual Channel DDR3/1866 Dual Channel DDR3/1333
Dual Channel DDR3/1333
Tri-Channel DDR3/1066

The Benchmarks

For our testing, we built up matched test rigs for each platform. On the AMD side, we used an Asus Crosshair Formula V board, for the Core i7-990X, an Intel DX58SO2 and for the Core i7-2600K, a Gigabyte GA-Z68X-UD3H-B3. The dual-channel machines featured 8GB of Patriot DDR3/1600, while the tri-channel box had 6GB of Corsair DDR3/1600. Matching GeForce GTX 580 cards, 64-bit Windows 7 Professional and WD Raptor 150 drives were used in all of the tests. Why no SSD? We’ve found odd inconsistencies with SATA 6Gb/s drives on the different platforms so to be safe, we’re sticking with proven technology for storage. We did vet several of our benchmarks using a very fast SATA6Gb/s SSD to make sure they weren’t being bottlenecked by disk I/O though.

PCMark7 is a synthetic test that attempts to measure overall system performance using models of what other applications actually do. It’s apparently not very multi-threaded either. Perhaps that why Intel’s Core i7-990X can’t outpace the Core i7-2600K that’s a third it’s price. Likewise, AMD’s octo-core FX-8150 can’t outrun the hexa-core Phenom II X6 1100T either (higher is better.)

3DMark11 is a popular graphics test that is mostly weighted toward GPU testing. There is a physics test that heavily taps the CPU though. Here, the FX-8150 outruns its sibling, the Phenom II X6, in 3DMark11’s physics test but it can’t touch neither of Intel’s top chips; the Core i7-2600K and Core i7-990X (higher score is better).

Bibble 5 is a heavily multi-threaded test that we’ve long been fans of for showing how you can scale with core counts. AMD has also been a proponent of it in the past. For the test, we take 4GB of RAW files shot with a Canon 5D Mk II and convert them to JPG. The FX-8150 has a good showing here, but so do all the chips. What’s truly amazing is that you can convert so many RAW files to JPEG in as little as two minutes! (lower score is better).

Cinebench 11.5 is a multi-threaded 3D renderer benchmark made by Maxon. Intel’s 12-thread Core i7-990X is the monster here but the surprise is the quad-core (well, not really a surprise to those who follow these chips) runs away from both AMD’s hexa-core and its octo-core. The marginally better score from the FX-8150 also shows that the “dual-core” modules with shared floating point units probably aren’t at their strongest when thrown heavy floating point intensive loads  (higher is better).

We run Far Cry 2 at very low resolutions and very low image quality settings to simulate just what you’d get if you had the ultimate, super-duper graphics card powering your system. If that were the case, both AMD parts would lose handily to the Intel CPUs.

But the truth is that when it comes to gaming at higher resolutions or playing a game that pushes boundaries of what your GPU can do, there is very little difference between a $1,000 Core i7-990X or the lowly Phenom II X6 1100T. Using Unigine 2.5’s DX11 benchmark at just 19x10, you can see that it’s all about the videocard (higher is better.)

HandBrake is a very popular free encoder. The makers of it emphasize quality over speed and thus don’t believe in GPU encoding today. It’s also multi-threaded like a motherfrakker. As much as FX-8150 has a tough time outshining Intel’s top end Sandy Bridge, the really bright spot is encoding is greatly improved. Here, the FX-8150 pulls almost even with the Core i7-2600K. Of course, if we were Intel fan boys, we’d say that it takes 8 cores to equal 4? And then drop in a Nelson Ha ha! (lower is better).

AMD Strikes Back! Of course, there were times when the FX-8150 had very nice showing indeed and is able to send the Intel fan boys packing. Here, using MainConcept’s Reference, we take a 1080P MPEG2 file and transcode it to H.264 using the high profile. Normally we do two passes but to save time, we’re doing just one pass here and the FX-8150 hands the Core i7-2600K its hat as it pushes it out the door (lower is better).

The Upshot

To be fair, we don’t think the FX-8150 should be compared to the 990X as that chip costs four times as much. But what about the 2600K? Even there, the FX-8150 has a difficult time and gets beaten up pretty badly by Intel’s top clocked Sandy Bridge. Nope, to be competitive, AMD actually thinks the 8-core FX-8150 is a better match Intel’s Core i5-2500K parts.

How good that is really depends on how you look at it the glass though. In some ways, it’s great that AMD has a part that is at least competitive with some of Intel’s higher tier Sandy Bridge CPUs. Looked at in differently though, How good is it that after all this time and a major redesign that the best AMD can do with an octo-core CPU is compete with a cheaper Intel quad-core chip? But that’s really how some will see it. We know that for people who only pay attention to core counts (like they did to megahertz) the sound of eight cores is really good. But perhaps that’s how we should start looking at it. With GPU and CPU cores starting to blur, does it really matter how many “cores” you have? Just as we once had to keep in mind that a 2.13GHz Athlon XP could kick the crap out of a Pentium 4 clocked up 1GHz faster, perhaps we have to stop looking at CPUs in pure core counts but instead look at, well, the model number.

It’s not all downer news for AMD though, we saw several signs of great performance with the new chip. Up against the Phenom II X6, the FX-8150 offers serious boost in performance in several encoding tests. In fact, in many encoding tests where the Phenom II X6 is road kill against Intel parts, the FX-8150 offers, umm, Sandy Bridge-like performance. In fact, in our MainConcept light test where we only do one pass rendering, the FX-8150 mangles the vaunted 2600K. In other tests, such as POV Ray, Bibble and HandBrake, the FX-8150 pulls pretty damn close too.

As sad as some AMD fans will be that Bulldozer doesn’t run over Sandy Bridge, it’s probably as close as AMD has been in some years.


3.3 Phenom II X6 1100T
3.6 FX-8150
3.4 Core i7-2600K
3.46 Core i7-990X
PCMark7 Score
2,838 3,450 3,332
PCMark7 Lightweight
2,311 2,222 2,612 2,496
PCMark7 Productivity
1,957 2,269 2,227
Cinebench 10 Single Core
4,128 4,080 6,011 5,176
Cinebench 10 Multi Core
18,735 20,277 23,315 28,019
Cinebench 11.5
5.87 5.97 6.84 9.13
POV Ray 4.7 (sec)
227.25 213.08 218.93 163
Bibble (sec)
176 136 137 116
Fritz Chess Benchmark
11,504 11,704 13,065 13,122
Intel Burn Test (Gflops)
Sony Vegas Pro 10 (sec)
ProShow Producer (sec) 1,610 1,171 1,043 1,037
MainConcept (sec)
CyberLink Espresso 6.5 CPU (sec) 420 429 379 316
CyberLink Espresso 6.5 Discrete GPU (sec) 414 412 329 333
CyberLink Espresso 6.5 QuickSync (sec) N/A N/A 311 N/A
7-Zip 12 Threads (MIPS) 18,149 20,400 19,046 30,178
7-Zip Max Thread Load (MIPS) 17,952 20,773 19,288 30,178
wPrime 4-thread 373.8 469 349 318.1
wPrime 6-thread 256.6 357.07 293.1 215.3
wPrime 8-thread 260.5 293.7 248.3 200.8
wPrime 12-thread 277.2 328 271.2 165.5
HandBrake (sec) 383 345 336 283
Sandra (GB/s) 12.93 17.7 17.6 20.7
Valve Particle Test (fps) 119 108 179 243
Dirt 2 (fps) 70 120 189 188.4
Far Cry 2 (fps) 107.1 111.2 202.3 207.1
Unigine 2.5 (fps) 54.2 53.9 54.2 53.8
3DMark2011 Score 5,958 6,138 6,469 6,412
3DMark2011 GPU
6,118 6,167 6,186 6,045
3DMark2011 Physics
5,606 6,426 8,184 9,037
3DMark2011 Combined
5,412 5,569 6,671 6,547

Around the web