Testbed Configuration and Testing Methodology
Since the new AMD FX-8350 is priced lower than the competitor’s top LGA 1155 processors, we will compare it against the junior quad-core Intel processors, besides the traditional Core i5-3570K and Core i7-3770K. They are Core i5-3470 and Core i5-3330. Of course, we also included the top Socket AM3+ processor from the previous generation – AMD FX-8150.
As a result, our testbeds were built with the following hardware and software components:
- AMD FX-8350 (Vishera, 8 cores, 4.0-4.2 GHz, 4 x 2 MB L2, 8 MB L3);
- AMD FX-8150 (Zambezi, 8 cores, 3.6-4.2 GHz, 4 x 2 MB L2, 8 MB L3);
- Intel Core i7-3770K (Ivy Bridge, 4 cores + HT, 3.5-3.9 GHz, 4 x 256 KB L2, 8 MB L3);
- Intel Core i5-3570K (Ivy Bridge, 4 cores, 3.4-3.8 GHz, 4 x 256 KB L2, 6 MB L3).
- Intel Core i5-3470 (Ivy Bridge, 4 cores, 3.2-3.6 GHz, 4 x 256 KB L2, 6 MB L3);
- Intel Core i5-3330 (Ivy Bridge, 4 cores, 3.0-3.2 GHz, 4 x 256 KB L2, 6 MB L3).
- Processor cooler: NZXT Havik 140;
- ASUS Crosshair V Formula (Socket AM3+, AMD 990FX + SB950);
- ASUS P8Z77-V Deluxe (LGA1155, Intel Z77 Express).
- Memory: 2 x 4 GB, DDR3-1866 SDRAM, 9-11-9-27 (Kingston KHX1866C9D3K2/8GX).
- Graphics card: NVIDIA GeForce GTX 680 (2 GB/256-bit GDDR5, 1006/6008 MHz).
- Disk drive: Intel SSD 520 240 GB (SSDSC2CW240A3K5).
- Power supply unit: Corsair AX1200i (80 Plus Platinum, 1200 W).
- Operating system: Microsoft Windows 7 SP1 Ultimate x64.
- AMD Chipset Driver 12.8;
- Intel Chipset Driver 18.104.22.1689;
- Intel Graphics Media Accelerator Driver 22.214.171.12461;
- Intel Management Engine Driver 126.96.36.1998;
- Intel Rapid Storage Technology 188.8.131.526;
- NVIDIA GeForce 301.42 Driver.
For our tests of the Socket AM3+ platform we installed KB2645594 and KB2646060 OS patches, which adapt the scheduler operation for Bulldozer and Piledriver microarchitectures.
Piledriver vs. Bulldozer: How Serious Are Microarchitectural Advantages?
When we described the innovations AMD introduced in their new processor design, we kept stressing that they weren’t too significant. However, many of you are most likely wondering about the specific performance boost resulting from the microarchitectural improvements and from the clock frequency increase. Therefore, before we precede with the actual performance tests in applications, we decided to pay special attention to a comparison between Vishera and Zambezi working at the same clock frequencies. For this round of tests we selected two eight-core processors from different AMD FX generations and manually set their frequency to 4.0 GHz. Turbo Core technology was disabled. Since other characteristics of Vishera and Zambezi processors are the same, the obtained results demonstrate the effect from replacing the Bulldozer microarchitecture with Piledriver.
Artificial comparisons like that are best performed in synthetic benchmarks. A popular SiSoftware Sandra SP5c suite offers a very good set of tests for that. Let’s use them:
Overall, I have to admit that the microarchitectural improvements introduced in new processors did not boost their performance that much after all. Although in some tests we see an almost 20% acceleration in the floating-point operations, it is important to understand that it results primarily from the FMA3 instructions support introduced in the Piledriver processors, which is missing in the original Bulldozer microarchitecture. General-purpose applications can hardly boast the same efficiency with new instructions, especially since Intel processors will only acquire similar functionality in the upcoming Haswell CPUs. Therefore, new AMD microarchitecture can only boost the performance in everyday tasks by a few percent. We owe most of the performance improvement to the increased clock frequency.
However, Vishera boasts another important advantage: increased cache-memory efficiency. For example, Cachemem test from the Aida64 2.60 suite shows a significant increase in the cache memory bandwidth almost with any operations. Our only concern is the read speed from L2 cache, which has dropped down by half, but in this case it seems to be more of the benchmark peculiarity rather than a microarchitectural issue.
Vishera 8C, 4.0 GHz
Zambezi 8C, 4.0 GHz
However, engineers decided not to provide Vishera processors with support for additional multipliers in the memory controller. Just like with Zambezi, the maximum supported memory mode is DDR3-2400. If you want to equip your Socket AM3+ system with even faster memory, which is already widely available in the market, then you would need to increase the base generator clock.
Let’s sum up a few things. The performance of the new AMD processors improved mostly due to the higher clock frequencies, more aggressive Turbo Core technology and faster cache-memory. Hypothetically, the support of new instruction sets implemented in the Piledriver processors could also help improve the performance in floating-point calculations, but the software developers aren’t rushing to integrate them just yet.
Will these improvements help to resolve the low specific efficiency of the AMD cores? Let’s check out the results of another benchmark, which makes all the pros and cons of the new AMD microarchitecture very obvious. The graph below shows the performance of AMD and Intel processors in Fritz Chess Benchmark, which creates a different number of simultaneous computing threads.
The two major bottlenecks of Vishera processor design are right there in the open: they obviously didn’t go anywhere and have simply migrated from the Bulldozer. First, the performance of individual FX-8350 cores is significantly lower than that of the cores in contemporary Intel processors. Second, when more than four cores of the FX-8350 processor are utilized, the performance gain from adding new cores to the equation is lower because of the “paired” design. As a result, the eight-core AMD FX-8350 can only compete successfully against the quad-core Core i5-3570K when the operational load has more than six threads. As for the Core i7-3770K, the flagship AMD processor can’t get even close to it: the performance of the Intel CPU in case of more than four threads continues to increase due to Hyper-Threading technology.