Testing AMD 1055T versus A10 7850K - wyldckat/wyldckat.github.io GitHub Wiki

Table of Contents

Introduction

Back in 2011 I posted the following on my blog at CFD-Online: Build times for OpenFOAM 2.0.x code with Ubuntu 10.10 with its gcc 4.4.5. This was when I had bought my AMD Phenom 1055T earlier that year and was happily treading along with a neat and low cost 6 core CPU.

Ever since AMD launched their A10-7850K APU back in mid January 2014, I've been eyeing the prices and saving up to a point where I could say "yep, I need a new CPU and more RAM, so it's time to purchase this wonder of technology". That time has come earlier this December 2015, because I wanted to help people on the forum and I needed more than 6GB of RAM than I had with my 1055T build.

Only around Christmas (25th of December), did I finally manage to build the new machine and transfer things around, but I also did some tests to see how well each machine performs. Note: I essentially simply changed the motherboard, CPU, RAM and PSU, while keeping a remaining barebones of the old machine.

In the CPU Specifications chapter I outline the specs for each machine build and what I expect that each one should perform. In the section Estimated performances are my estimated performances, based on online benchmark values and each one of the machines' specs... which is where you will see why I chose the A10-7850K for my personal home computer.

In the chapter Performance tests are the actual performance results.

CPU Specifications

AMD Phenom 1055T

Reference pages:

Specs:
  • 6 cores
  • L1 cache:
    • 6 x 64 KB 2-way set associative instruction caches
    • 6 x 64 KB 2-way set associative data caches
  • L2 cache: 3MB
    • 6 x 512 KB 16-way set associative exclusive caches
  • L3 cache: 6MB
    • Shared 6 MB 48-way set associative cache
  • TDP: 125W
  • Lithography: 45nm SOI
  • Frequency (not overclocked):
    • Max: 2800 MHz (when using all 6 cores)
    • Min: 800 MHz
  • 2 memory communication channels
  • 6 GB RAM DDR2 (3 modules, 1 unpaired) at 800 MHz
    • 512 MB shared with integrated GPU in motherboard

AMD A10-7850K

Reference pages:

Specs:
  • 4 cores + 8 GPU stream cores (8*64 = 512 Shaders total)
    • Note: has AVX and HSA support
  • L1 cache:
    • 2 x 96 KB 3-way set associative shared instruction caches
    • 4 x 16 KB 4-way set associative data caches
  • L2 cache: 4 MB
    • 2 x 2 MB 16-way set associative shared caches
  • TDP: 95W
  • Lithography: 28nm SHP
  • CPU Frequency (not overclocked):
    • Max: 3700 MHz (when using all 4 cores)
    • Min: 1700 MHz
  • GPU Frequency (not overclocked): 720 MHz
  • 2 memory communication channels
  • 16 GB RAM DDR3 (2 modules) at 2133 MHz
    • 2048 MB shared with the integrated GPU in CPU die

Estimated performances

If we compare the two CPU indexes here: http://www.cpubenchmark.net/cpu_list.php - we get the following numbers:

  • AMD Phenom 1055T: 5059
  • AMD A10-7850K: 5568
This seems to simply that the two CPUs should have nearly the same performance. Nonetheless, when I do the math I usually do based on lithography scaling and other details, I get different estimates:
  • AMD Phenom 1055T:
    • 45nm to 28nm implies that the 1055T would have 2800 MHz as an equivalent frequency downscaled to 1742 MHz.
    • Which times 6, would be equivalent to roughly 10452 MHz cumulative frequency.
  • AMD A10-7850K:
    • Has 3700 MHz at 28 nm, therefore the 4 cores should result in a rough equivalent of 14800 MHz cumulative frequency.
In other words, the A10-7850K should be theoretically almost 50% faster, not just roughly 10% faster: 14800/10452 ~= 1.42

So, what gives? Well, if you take a look at the cache L1 and L2 levels, you'll notice that:

  • the 1055T has 6 cache regions, one per core;
  • the A10-7850K has 2 or 4 cache regions, depending on the type or use scenario, which equates to having 2 cores sharing the same cache region in most situations.
Well, this essentially means that the A10-7850K employs a technique similar to Intel's HyperThreading (HT), where not all features have dedicated implementations for each core. Mind you that Intel's HT has it's pros and cons, so much that goes to the extreme with their Xeon Phi in having 4 threads per core, instead of just 2 per core.

The pros and cons of such a strategy are usually these:

  • Pros:
    • Fewer transistors needed when sharing integrated components, such as cache.
    • Tasks are essentially scheduled and executed simultaneously whenever possible.
  • Cons:
    • Nearly impossible for each independent core to be able to perform its own mathematical operation on its own.
In HyperThreading, this usually means that the CPU is only able to output at a max of 30% more equivalent cumulative frequency, based on its maximum stock frequency. So if we assume the same goes for the A10-7850K, the math results in:
  • 2 cores at 3700MHz have an equivalent 7400MHz cumulative frequency;
  • which boosted in 30% with the duplicated core components, equates to 9620 MHz...
  • OK, if we do the opposite math, 10452 * 1.10 = 11497 MHz, which means the 4 cores in the A10-7850K deliver around 55% boost over the base 2 cores.
OK, so far so good, the A10-7850K as a conventional x86_64 Simultaneous multithreading (SMT) processing unit is probably only around 10% faster than the 1055T, so this seems like a somewhat bad upgrade. Well, no quite. Here's why:
  • Remember how CUDA has all the buzz nowadays of being the go-to compilation methodology for employing NVidia GPUs for number crunching?
  • Well AMD/ATI cards have better number crunching capabilities than NVidia, so much that Bitcoin mining had made AMD/ATI cards more expensive some time ago, because they were the best bang for your buck when it came to bitcoin mining: Non-specialized hardware comparison - Note: this was before the ASIC became the comercial miners.
    • For example, the AMD Radeon HD 6990 outputs around 800 to 860 Mhash/s, while the comparatively the NVidia GTX 680 outputs around 110 to 120 Mhash/s.
    • More details: Why a GPU mines faster than a CPU
Now, bitcoin mining asside, leads me to the following calculation:
  • GPU at 720 MHz with 512 shaders results in an equivalent cumulative frequency of 512 * 720 = 368640 MHz
  • Comparing with the 1055T, this is roughly 35 times more potential mathematical power, but mathematical only.
This essentially equates to the key advantage here that the A10-7850K has got, namely that it is practically sporting a massive oversized Floating Point Unit, which can access the same RAM space as the CPU. This essentially removes one of the largest bottlenecks in GPGPU cards: the transfer of data between the card and the main RAM on the motherboard.

But it's not all a sea of roses:

  1. Today's GPU cards have GDDR5 RAM at 7000MHz.
  2. The A10-7850K using the maximum stock clock for DDR3 RAM interface, is only at 2133MHz.
  3. The A10-7850K can only use this feature if an HSA compiler is used, along with updated/adapted source code can use the HSA features.
Therefore, I have two main reasons for purchasing the A10-7850K:
  1. It has got a videocard benchmark index of 1005, versus 84 on the integrated Radeon HD 3200 the motherboard for the 1055T had.
    • Which means I can play with higher fidelity (3/5-year old) games.
  2. I can try and build HSA support into OpenFOAM and check how it compares to the GPGPU implementations :)

Performance tests

Building OpenFOAM 3.0.x source code

Building OpenFOAM 3.0.x (commit `195caf7479b2`), with ThirdParty-3.0.x already built:

  • AMD Phenom 1055T - Ubuntu 12.04 (GCC 4.6.3):
    ~/OpenFOAM/OpenFOAM-3.0.x$ time ./Allwmake -j > log.make 2>&1

    real  50m48.133s
    user  236m13.122s
    sys   9m18.259s
  • AMD A10-7850K x4 - Ubuntu 15.10 (GCC 5.2.1):
    ~/OpenFOAM/OpenFOAM-3.0.x$ time ./Allwmake -j > log.make 2>&1

    real  63m52.444s
    user  217m52.936s
    sys   6m25.872s

There are a few details at hand here:

  1. GCC 5.2.1 does more optimization and complaining than GCC 4.6.3.
  2. Nonetheless, even though the A10-7850K took longer to build in real time, but it took lesser time in user and sys timings, making it more efficient than the 1055T.
  3. Both builds are slower than the 36 minutes it took to build OpenFOAM 2.0.x back in 2011, but the base code is now considerably larger and more template based, which requires more processing by the compiler to properly compile and optimize things.

4 versus 2 cores

The next day I tested building OpenFOAM-dev with only 2 cores on the A10-7850K and I got this: 

    $ time ./Allwmake -j 2 > log.make 2>&1

    real   95m43.559s
    user   177m2.468s
    sys    5m9.592s

Which is we do the math gets us: 96 / 64 = 1.5 is essentailly the estimated ratio, based on the aforementioned 2 cores at 3700MHz have an equivalent 7400MHz cumulative frequency and if we do the opposite math, 10452 * 1.10 = 11497 MHz, which means the 4 cores in the A10-7850K deliver around 55% boost over the base 2 cores.

OpenFOAM 2D Cavity tutorial with 300 kcell

The case used is shared here: https://github.com/wyldckat/wyldckat.github.io/tree/1055T-vs-7850K

The total number of cells is 550*550*1 = 302500 cells. The decomposition method was `simple` and only over X, out of simplicity.

Edit the script `DoALap` to set the desired number of cores and then run it like this: 

    ./DoALap

Summary of results:

  • 1055T:
    • 1 core: 203 s
    • 2 cores: 151 s
    • 4 cores: 123 s
    • 6 cores: 118 s
  • A10-7850K:
    • 1 core: 117 s
    • 2 cores: 82 s
    • 4 cores: 78 s
Very curious results, aren't they? Well, there are several details at play here:
  1. Where the 1055T 6 cores left off at 118 s, the A10-7850K picked up with a single core at 117 s. Reasons:
    1. RAM speed matters: 800 MHz vs 2133 MHz equates to nearly 2.7 times faster memory access.
      • Although keep in mind that the unbalanced RAM distribution in the machine with the 1055T CPU might also have influenced the result.
    2. 5 to 10% performance boost can be expected from the differences in GCC versions.
  2. The 1055T had a very hard time scaling up with the core count. This can easily have been due to a bottleneck from the RAM interface, either due to the 800 MHz speed or the unbalanced memory module pairing.
  3. A10-7850K did 1.4 times faster with 2 cores than 1 core.
    1. But using 4 cores was only marginally better, so much that it's within a plausable margin of error.
  4. The A10-7850K with 2 cores was rougly 1.4 times the performance of the 1055T with 6 cores.
  5. Assuming the two machines were operating at top speed with all cores, the performance ratio between them is 118/78 = 1.51
Conclusion? Well... if the RAM is the bottleneck here, then another test should be conducted, namely using more cells.

OpenFOAM 2D Cavity tutorial with 1 Mcell

The case used is shared here: https://github.com/wyldckat/wyldckat.github.io/tree/1055T-vs-7850K

The total number of cells is 1000*1000*1 = 1000000 cells. The decomposition method was `simple` and only over X, out of simplicity.

Summary of results:

  • A10-7850K:
    • 1 core: 318 s
    • 2 cores: 223 s
    • 4 cores: 220 s
Well, it's consistent! Did the same performance ratios as in the 300 kcell case. It was not possible to test with the 1055T machine in this case, because it was not operational at the time of this test and writing.

Also tested with 2x2 and 1x4, instead of 4x1 decompositions and the results were the same with the 4 cores, no improvement... well, actually, it was a bit slower, around 230 s.

Conclusions

  1. For hard-core CFD mathematics, the x86_64 processing units in A10-7850K act as Intel's HT technology, where only the real cores really count.
    1. This means that the 2 extra cores simply work as schedulers for mathematical operations, which is why there is a slight performance increase when using all 4 cores.
    2. In addition, keep in mind that when using only 2 cores, there should be a performance boost from the turbo technology that these CPUs have got. Which according to the page at cpu-world.com for the A10-7850K, means that 2 cores will operate at 3900MHz, which equates to 3900/3700 = 1.054, so roughly 5% more performance.
      • In other words, the 82s with 2 cores should have been 86s, resulting in a speed up of 86/78 = 1.10 when using the 4 cores versus 2 cores at 3700MHz. Which is within the usual HT technology performance boost that can be observed in many situations.
  2. The RAM was very likely limiting the capabilities on the 1055T processor, therefore further tests should be conducted, although I don't have the hardware to do it myself.
  3. Using primarily x86_64 instructions, there are two ratios that are well accounted for between the estimates and the runs:
    1. A10-7850K will compile roughly 50% faster when using the 4 cores instead of just 2. This is on-par with the estimate made based on the performance indexes from cpubenchmark.net and the estimated cumulative frequency.
    2. The A10-7850K should be theoretically almost 50% faster, not just roughly 10% faster: 14800/10452 ~= 1.42, which was reflected in the performance ratio between them is 118/78 = 1.51 when running `icoFoam` in parallel.
      • The discrepancy can be accounted for with the low performing memory installed in the machine with the 1055T CPU.
  4. Actually, there is another factor that should be accounted for: the number of memory channels. It's possible that only having 2 memory channels in both CPUs is why they are bottlenecking so much after the 2 core count.
    • A test needs to be made where only memory copying is done, in order to assess the speed at which it is able to copy memory from one 1GB array to another 1GB array, since this will give us the ratio of performance that affects these CPUs.

Addendum

I've done some tests with AVX versus the standard FPU. The results are available here: https://github.com/wyldckat/avxtest/wiki/A10-7850K-with-DDR3-at-2133-MHz

⚠️ **GitHub.com Fallback** ⚠️