The mid-range of the Graphics card market is arguably the most important segment for both AMD and NVIDIA for the competing parts – it’s where the majority of gamers look when they want to purchase new hardware. Today I’m going to take a look at each company’s competing solutions around the $330-350 price point. That puts the GeForce GTX 970 from NVIDIA, and the Radeon R9 390 from AMD, firmly in my sights. It’ll primarily be a performance comparison with games that have built in benchmarks.
GeForce GTX 970
Serving at the frontline of the mid-range for NVIDIA, we have the GeForce GTX 970. Launched in September of 2014 for $329, this middle-weight contender features a cut-down version of the same GM204 graphics processor used in the more expensive GTX 980. The card I’m testing is EVGA’s custom variety; with the ACX 2.0 cooler.
EVGA’s GTX 970 ACX 2.0
Taking a closer look at the GM204-200 chip powering the GTX 970.
- 13 SMMs
- 1664 CUDA cores
- 104 Texture Mapping Units
- 56 ROPs
- 256-bit memory interface (224-bit aggregated)
- 1.75MB L2 Cache
- 4GB GDDR5 (3.5GB High-speed)
- 7Gbps memory clock
- 1050 MHz core clock
- 1178 MHz boost clock
Out of GM204’s 16 SMM (Streaming Multiprocessor Maxwell), three have been disabled in the GTX 970, this brings the total CUDA core count down to 1664 from 2048, and the texturing unit count to 104, from 128. The memory interface is 256-bits wide, but what’s not shown on this graph, is that the GM204-200 actually has some disabled cache and the ROP (Render Output Pipelines) associated with it. This brings the total ROP count down from 64, to 56. This also has implications for the GTX 970’s 4GB of GDDR5; whereby only 3.5GB of that is aggregated on an effective 224-bit interface. The other 512MB is only addressable via a narrow 32-bit wide data path. The TechReport’s article on the subject is a good read for those interested in just how the 970’s strange memory subsystem functions.
The GM204-200 chip powering the GTX 970.
Compared to the R9 390, the GTX 970 has pretty much less of every on-chip resource available. Nvidia’s Maxwell architecture relies on efficient use of resources and higher clock speeds to achieve competitive performance. The GM204 chip will often boost higher than its rated clock speed during load dependent on heat and power consumption, and this sample is no different: clocks in my tests ranged anywhere between 1228 MHz and 1254 MHz, with the average being around 1240 MHz.