For the last ten years AMD have consistently improved on a single architecture, which is an incredible feat all on its own. As a result, consumers have benefited from a cost effective upgrade path, generation after generation. While raw performance of AMD CPUs has started to lag behind because of this, we have seem some amazing bang for your buck platforms from AMD. There has also been return fire from AMD at INTEL in the form of the Phenom X6 Thuban , offering more cores for less dollars. After a few delays and a ton of rumors, we are finally able to show you the nearly mythic AMD Zambezi CPU with Bulldozer Architecture. Let’s see if AMD can strike back at the INTEL empire!
Introducing the AMD Scorpius (FX) Platform
Like the AMD Spider, AMD Dragon and AMD Leo platforms before it, the new AMD Scorpius platform consists of CPU from the AMD FX lineup, an AMD 900 series motherboard, and an AMD HD 6000 series graphics card. While the 6000 series GPUs have been around for some time, it has only recently been joined in the market by the other two pieces of the Scorpius Platform.
The next piece to the puzzle would be the motherboard. The new 900 series motherboards are a refresh of the 800 series with some new goodies and more consistency with certain features. The consistency is for features such as USB 3.0 which could be found on some 800 series boards, while it is present on all 900 series motherboards. Other things would include the AM3+ black socket, FX CPU support, and support for memory up to 1866 MHz.
The major difference between the 800 and 900 series motherboards is the latter also including support for Nvidia SLI as part of the AMD chipset. This is great news for AMD users as it opens up Nvidia to multiple GPU setups.
Now for the crown of the Scorpius Platform, we will be looking at the AMD FX-8150 CPU with an MSRP of $245. The release of Zambezi to the market will also see 3 other CPUs. The FX-8120 at an MSRP of $205, FX-6100 at $165, and FX-4100 at $115 with the usual $20 to $45 markups. These are very aggressive prices for this all new architecture built from the ground up and plenty of new goodies.
The Anatomy of a Bulldozer
Zambezi, which is the name for all FX series CPUs, all use the Bulldozer architecture. In this unique design AMD has taken two independent CPU cores and combined them into two core modules. These modules will still have a separate integer blocks, L1 caches, schedulers, and with the scheduler connected via 4 pipelines. These two integer cores will share a FP scheduler, 2MB L2 cache, two symmetrical 128-bit FMAC (fused multiply-add capability) floating-point cores, decoder, and fetch block.
Some will argue that shared portions means that the Zambezi chips are not “true” octa-core processors. This is not exactly true. The vast majority of CPU tasks are integer based and will thus have eight cores for processing. This is also true for 128-bit floating point calculations as each module has two 128-bit FMAC. The only time Bulldozer will work like a 4 core chips is with 256-bit floating point calculations as the two 128-bit will be combined to run in 256-bit. This is why the FMAC blocks must be symmetrical, which relates to the scheduler, and ensures one does not lag behind when receiving instructions.
The Bulldozer modules are then placed on a single chip in sets of two, three, and four to create quad-core, hexa-core, and octa-core chips. These modules are then interconnected through a single unified shared 8MB L3 cache. This brings the FX-8150 and other octa-core FX chips to just over 16MB of cache.
The Bulldozer architecture does not stop at a new design for the cores, as it also boasts an all new memory controller block. We got to see a preview of this with the APU review, but I will be running the FX-8150 at the new AMD stock spec of 1866 MHz. The obvious difference when compared to older AMD memory controllers is support for higher clock speeds and better memory throughput. Under the surface you will also find that the memory controller adds an additional 8-bits to each channel, bringing it up to 72-bits per channel, but are used solely for error correction. This should help ECC which is normally only used in servers, but should provide no additional benefit for desktop users.
The Zambezi chips also sport the new Turbo Core version 2 from AMD. Turbo Core v2 now has 2 boost states. The first boost state is for all eight cores and the second boost state is for only 4 cores. The FX-8150 has a default clock speed of 3.6 GHz, eight core turbo of 3.9 GHz, and 4 core turbo of 4.2 GHz. This means a task that uses 4 or few threads should see four 4.2 GHz cores for this task and allow the other four cores to drop to idle state. This saves power while providing the best performance for low thread count applications.
Bulldozer also has several bumps in speeds. The FX-8XXX chips will all have a default HyperTransport/North Bridge frequency of 2.2 GHz (200 MHz higher than FX-6XXX, FX-4XXX, and past AMD chips). Since this is a double data rate link, you get an additional 400 MT/s (Mega-transfers per second). This increase in bandwidth between the CPU and Northbridge should reduce GPU bottlenecks, improve device throughput, and allow for more devices to be connected without performance loss.
The final piece to the Bulldozer puzzle is support for a battery of new instruction sets. For instance, a CPU is a calculator and crunches math equations. It is capable of performing simple math equations such as A = B + C at a rate of several million per second. Instructions sets are collections of these simple equations with updates being shortcuts that increase the number of equations that it can perform.
The new instructions include sets for SSE4.1, SSE4.2, AES, AVX, XOP, and FMA. While this seems like a minor update, these instruction sets can make a world of difference in certain tests and applications. It is possible for your CPU to turn 8 clock cycle tasks down to 4 or few cycles. In some cases, especially with AVX, you could gain 4 or 5 times the performance. When you place multiple Bulldozer modules, add the memory controller, L3 cache, and new instruction sets you get an AMD Zambezi processor.
What’s in the Box?
There will be two variations to the FX-8150 retail PIB (Processor in a Box). There will be a standard air cooling version which will be sold in a metal container. The container has a cut out on the side to display the chip in much the same way AMD has done in the past. Even though my FX-8150 is retail, I did not receive a complete retail box. I suspect AMD will use the same packaging and air cooler as the Phenom II lineup.
Test System and Notes
The first thing you want to be sure of when using an FX CPU with a compatible 800 series board is to verify you are using the most up-to-date BIOS. If you use a BIOS that does not have support for FX, you will not get a post screen. It is still a good idea to update to the latest stable BIOS with even your 900 series motherboard, but all should work with their factory installed BIOS with no issues.
Leo Platform Test System
- CPU: AMD Phenom II X6 1100T
- Cooler: Noctua NH-D14
- Case: Rosewill Thor V2
- Motherboard: GIGABYTE 890FXA-UD5 (rev. 3.1 AM3+ socket)
- GPU: HD 6850 Reference 1GB
- RAM: G.Skills Ripjaws 2 x 4GB @ 1600 MHz 9-10-9-27 CR1
- HDD: Seagate Barracuda XT 2TB
- PSU: Antec HCP 850W
- Sound: Onboard Realtek ALC889
- OS: Windows 7 Ultimate 64-bit with CCC 11.10 Preview drivers Version 2
Scorpius Platform Test System
- CPU: AMD FX-8150
- Cooler: Noctua NH-D14
- Case: Rosewill Thor V2
- Motherboard: GIGABYTE 990FXA-UD5
- GPU: HD 6850 Reference 1GB
- RAM: G.Skills Ripjaws 2 x 4GB @ 1866 MHz 9-11-10-28 CR1
- HDD: Seagate Barracuda XT 2TB
- PSU: Antec HCP 850W
- Sound: Onboard Realtek ALC889
- OS: Windows 7 Ultimate 64-bit with CCC 11.10 Preview drivers Version 2
Software
- PCMark07
- 3DMark11
- Sandra 2011 SP5
- Alien Vs. Predator DX11 benchmark
- Resident Evil 5 DX10 benchmark
- Metro 2033
- Cinebench 11.5, Handbrake 0.9.5
- CrystalDiskMark 3.0.1
- OCCT
- FurMark Burn-in
- wPrime
- Hyper PI
- Fraps
As you can very plainly see, we are pitting the outgoing Leo Platform (890FX) against the new Scorpius Platform (990FX). We have even been able to procure a GIGABYTE 990FX-UD5 board, courtesy of GIGABYTE USA, to allow us to do a proper head to head battle with its older counterpart, the 890FX-UD5. For anyone looking to upgrade from Leo to Scorpius, we are about to let you know where you’ll see benefits and allow you to figure out if the jump is worth your money.
BIOS and Overclocking Notes
The GIGABYTE BIOS on the two motherboards in my tests are practically identical. Neither have a UEFI, but instead settle for the tried and true standard BIOS we have come to know and love. There is one major difference in that the 990FX board responds 10 times better to me pressing and holding delete to enter the BIOS. I am not sure what GIGABYTE changed, but it makes the 990FX much easier to work with.
I will not be going into depth for this part one of our review, but I can give you a glimpse of what is to come. After some tweaking, I got my FX-8150 stable at 4.74 GHz and 2.4 GHz Northbridge. More on this in part two of course.
Architectural Notes
Before we dive into testing, there is a little something about the design of Bulldozer which can effect its performance. While marketed as an eight-core processor, Bulldozer is more like a cluster of four, dual-core processors due to its modular design. This does not seem to ready effect general performance, but it does create thread scheduling by creating an better way to handle multiple thread application.
As the Operating system still has control over thread execution, there are occasions where the task scheduler will use a non-optimal allocation of threads. This occurs when two interdependent threads are run on different modules. If two threads are interdependent you can gain small improvement in performance if both threads are run on a single module. In this scenario, when thread 1b needs something from 1a, it can access it from the L2 cache. If these are placed on two separate modules, this information would have to be passed through the L3 cache. While the processor can handle this no problem, it is a sub-optimal execution.
This architectural quirk can be resolved, but it would require Microsoft to update Windows 7. According to AMD, the issue with interdependent threads is not a problem for Windows 8 Dev. Improvements in Windows 8 result in 1% to 10 performance gains in games. We will test this in depth in part two as well. I hope to see an update to Windows 7 in the very near future to mimic or improve upon this behavior.
PCMark07 – Total System Performance
PCMark07 is the latest and greatest version of Futuremark’s total system evaluation tool. It runs a battery of smaller tests (21 to be exact) in 3 cycles. These tests mimic daily use tasks like watching videos, live encoding, word processing, web browsing, etc. This should be a good test to see what you can expect from the FX-8150 on a daily basis.
When directly compared to the Phenom II 1100T, the FX-8150 holds a decent advantage in everything especially Computational work. The FX does fall behind in Entertainment which would consist of encoding and HD playback plus post processing. I personally want a more commanding lead here, but I am glad to see improvement.
Cinebench 11.5 – CPU & OpenGL Performance
This is a straight forward test of the CPUs rendering and computational power. I will be testing single threaded, multi-thread, and OpenGL workloads. The OpenGL test mainly focuses on the GPU. Since I am using the same GPU and the chipsets are virtually identical, any changes will be a reflection of the CPU. Also note that these test were done with the AMD 11.9 WHQL drivers as Cinebench would crash the 11.10 preview drivers.
The FX-8150 does perform better in the OpenGL test which should mostly be a result of the improved memory controller. There is also a 4.69% gain in the multi-threaded test; on the other hand, there is a 13.51% loss in the single thread test. It would seem that the FX-8150 is not very efficient in single core workloads. We will re-examine this in Part Two to see what happens if I prevent the thread from changing cores or modules during operation.
Sandra 2011 SP5 – CPU Specific Testing
The Sandra 2011 benchmark suite is a collection of great tools for testing various specifics of your system. While there is a lot I could test here, we are going to focus on CPU centric tests that simulate tasks that will be done during normal use.
When it comes to Processor Arithmetic and Multimedia processing, the 1100T just doesn’t stand a chance. The gains here are 10.9% to 201.3% in these test. As this does not match up with the PCMark07 Entertainment tests, I can only assume Sandra 2011 is using more optimized data paths. This means it is using the newer instruction sets which are often time saving instructions that perform more than one operation per clock cycle. I am sure PCMark07 will update their tools to better take advantage of Zambezi chips in a future revision.
In the Cryptgraphy test, which focuses on encryption and decryption, we get very interesting results. The SHA-256 test shows falling well behind the 1100T to the tune of a 48.3% lose. Then the FX-8150 pulls the AES instruction set support and walks all over the 1100T. For AES-256 test, the FX-8150 has an absolutely unfair 588.34% advantage.
In my final test for Sandra 2011, I just wanted to get an idea of the best memory bandwidth each CPU was capable. While the FX-8150 is using its recommend spec for RAM, I used settings for the 1100T I tested for best performance which for me is 1600 MHz @ 9-10-9-27 CR1. The Phenom II X6 1100T on its best day is over 6 GB/s behind the Bulldozer memory controller’s bandwidth. Even when considering the FX-8150 average of 18.2 GB/s instead of the best throughput being displayed, there simply is no comparison here. This is a 33% improvement.
wPrime & HyperPI– Single Threaded and Multi-threaded Pi Calculation
This time we are going to use two popular programs that enthusiasts use to test the stability of their systems. These programs are based on old code and are not well optimized. This should give us an idea of how the FX-8150 responds to older programs.
In these test the 1100T wins by a landslide. Tackling these older programs comes down to raw clock for clock power and the FX-8150 just can’t keep up. I am sure what the exact cause for this is, but I believe it is more to do with the software than the shared resources in Bulldozer modules being a bottleneck. In this test I simply cannot be sure. Regardless of what the cause, it currently means the FX-8150 falls below the 1100T here.
7-Zip Benchmark – Operations Per Second
For the 7-Zip benchmark we are going to do some real world computational work. As we have seen in Sandra and PCMark07, the FX-8150 should do well in test involving this kind of workload from software that is kept up-to-date.
As expected the FX-8150 holds a commanding lead here. The Zambezi chips seem to do much better in real world programs that are either new or kept up to date. It seems only when the software being used is fairly dated do we get relative under performance. This is a double edge sword as there are a few very popular programs that don’t get updates very often such as Handbrake.
3DMark11 – Synthetic Graphical Performance
Now lets look into the one thing we are all concerned about: Gaming performance. I will start with 3DMark11 to get an idea of what should be expected from FX-8150 now and in the future.
Much like PCMark07, the FX-8150 holds a nice lead in both Entry and Performance testing. It also has a big impact on the Physics score. Note that the Physics score is shared with the GPU in 3DMark11 with the GPU handling soft body object, with the CPU working the rigid body physics.
Real World Graphical Performance (Resident Evil 5, AvP benchmark, Metro 2033)
Now lets take a look at real game performance from the FX-8150. AvP was tested at DX11 with its default settings. Resident Evil 5 was tested with vsync off, frame rate unlocked, 8xAA, and high everything else. Metro 2033 on its “Very High” preset which included 4xMSAA and 16xAF. These were all tested at 1920 x 1080.
In these tests the FX-8150 and 1100T go back and forth. DX9 shows the FX-8150 falling behind, but it has a very small advantage in DX10 and DX11. This is not unexpected seeing as other older apps have not performed as well as hoped.
In Metro 2033 both have the same minimum frame rate. The FX-8150 manages to have a higher maximum frame rate, but falls 1 FPS behind in average. While playing the exact same section for benchmarking with FRAPS running, my personal character movement could account for this difference. As such, I say to only consider the minimum frame rate which amounts to no gain for either CPU.
Power Consumption
This is a simple test of the power from from the wall using a Kill-a-watt meter. I then adjusted the scores based on the tested efficiency of my Antec HCP 850W PSU to get a better idea of what the PC system is using.
As expected of the FX-8150 with more components, there is an increase in power consumption. The FX-8150 alone is doing pretty well with idle power considering its high clock speed. I noticed there were three major voltage points for the FX-8150. 0.84V for idle, 1.2375V for standard, and 1.4125V for turbo. This is for my chip in particular as these numbers will vary slightly for each chip.
Final Thoughts (For Now)
I am honestly not sure what to make of the FX-8150. I did not really lose or gain any performance in games. While this is disheartening, such is not the case with Sandra 2011. In every benchmark I ran in Sandra, with the exception of the SHA-256 results, shows FX-8150 far far ahead of the 1100T. The FX-8150 seems to shine brightly in computational workloads.
When you turn on an application or perform a certain task, I am not sure if I will get i7 Sandy Bridge or Phenom II 965 performance. Newer applications and benchmarks that take full advantage of the new instructions sets, perform between the INTEL i7 2500K and 2600K. Older software and single thread applications perform below the 1100T and closer to my AMD Phenom II X4 955. While it is not bad performance it is inconsistent and does not encourage a sense of an upgrade in these cases.
My gaming performance is being limited by my GPU making most of the results superfluous, but trends do arise. I notice the older the titles core engine, the worse the FX-8150 performs. I will examine this further with a more recent title in Part Two of our evaluation.
The other drastic change is AMD has improved memory bandwidth. This should prove to translate to great performance in multiple GPU setups and memory intensive programs. While this does not help with the HD 6850, we will look at a few of the issues I ran into here. In Part Two of the FX-8150 review, we will look at Overclocking for both the FX-8150 and 1100T, controlling the affinity for single threaded apps to prevent the “first available core” scenario. Well also be testing with newer titles like Deus Ex: Human Revolution and give a comparison to INTEL’s Core i7 2600k, plus other specialized tests to help you make an informed purchasing decision.
For now, I can see the Bulldozer design and the Zambezi chips being something great with a few updates to Windows 7 and software that understands how the chip is designed. I just hope I can see some of these improvements soon through software updates, BIOS updates, and/or AMD drivers. The problem here is a lot of the optimization that would help Zambezi is out of AMDs hands.
Help Us Improve Our Reviews By Leaving a Comment Below!