Test Setup and Synthetic Benchmarks

Test Setup

 Motherboard   ASUS ROG Maximus VI GENE (Intel Z87)
 CPU  Intel Core i7 4770K @ 3.5GHz
 Memory  Kingston HyperX Predator 8GB DDR3-2666 11-13-13-32 @ 1.65V
 G.Skill Sniper Series 8GB DDR3-2133 10-12-12-31 @ 1.60V
 (Baseline) Samsung OEM 8GB DDR3-1600 11-11-11-28 @ 1.50V
 GPU  NVIDIA GeForce GTX 770 (OEM), ForceWare 332.21
 Storage  OCZ Vertex 460 120GB SSD
 Cooling  Corsair H75 AIO Liquid Cooler, Noctua NT-H1 Thermal Paste
 Power  Corsair CX750 750W
 OS  Microsoft Windows 7 Professional 64-bit, SP1












Comparing Bandwidth

To test the HyperX memory against some more conventional examples, I began with some commodity Samsung RAM, of the common type found in any number of OEM systems. It's 1600MHz DDR3 memory operating at vanilla JEDEC standards, with automatic 11-11-11-28 timings at 1.50V. For an enthusiast example, I selected a pair from G.Skill’s Sniper series with X.M.P. support for 2133MHz, and very good 10-12-12-31 timings at 1.60V.

Notes on the Test Setup

I install only what is necessary on the test system, and Windows is configured very simply. In brief, Windows auto-updating, System Protection, and Hibernation are disabled, and the power profile is set to "High Performance". I use the “Windows Basic” theme with no translucency to further reduce system overhead.

For graphics benchmarking (when possible) I allow the GPU to reach full load temps before beginning the test run, preventing inconsistency in the results by attempting to normalize any thermal throttling that may occur. All benchmarks are performed from a clean state, with the system fully restarted before beginning each new test run. We'll start with synthetic benchmarks.

The HyperX kit shows unusually fast write speeds in AIDA64 – and it was a repeatable anomaly.

Bandwidth scales with memory speed, as expected. The tremendous spike in writes from the DDR3-2666 memory in AIDA64 is odd (approaching quad-channel bandwidth), but was constant after verifying the hardware configuration and re-testing (possibly a bug in this version of AIDA64).
SiSoft Sandra shows linear scaling and good performance increases, with the DDR3-2666 memory leading the baseline DDR3-1600 by over 25%, and by 10% over the DDR3-2133.

In performance mode 3DMark 11 exhibited perfect scaling as memory speed increased. The extreme preset was not as predictable, and the 2133MHz G.Skill set ended up on top there. The CINEBENCH OpenGL rendering test scaled impressively, with clear benefits to higher memory bandwidth in this application.

In the Valley benchmark we see some minor benefit from higher frequency memory at 1920×1080, but even so the 2666MHz HyperX kit’s 67% advantage in clock speed over the DDR-1600 only results in a whopping 1% gain in FPS. At 2560×1440 there is very little, if any, impact. Unlike the OpenGL performance scaling in CINEBENCH, the DirectX-based acceleration looks completely GPU bound. Now we will look at a couple of in-game benchmarks, and see how memory speed translates into 'real-world' gaming performance for this system.

« PreviousNext »