|2015-11 Appraisal of Discrete Desktop GPU Cards|
Appraisal of Nvidia GTX and AMD Radeon GPU Cards as at 2015-11
• Nvidia GTX (standard options for ages) Maxwell
o 960 1024c (08), 1.2GHz, 2GB, 7GHz, 128b, 150W P45
o 970 1664c (13), 1.1GHz, 4GB, 7GHz, 256b, 256W P65
o 980 2048c (16), 1.1GHz, 4GB, 7GHz, 256b, 300W P90
• AMD Radeon (new options) GCN 1.1 (corrected 2016-04)
o R7-360 0768c (12), 1GHz, 2GB, 6.5GHz, 128b, 150W P25
o R7-370 1024c (16), 1GHz, 4GB, 5.6GHz, 256b, 150W P40
o R9-380 1792c (28), 1GHz, 4GB, 6.0GHz, 256b, 150W P45
o R9-390 2560c (40), 1GHz, 8GB, 6.0GHz, 512b, 300W P65
a) All models listed above are PCIe v3 x16 cards. Most of them takes up 2 slots each, consumes a substantial amount of electricity, and requires a lot of ventilation cooling. They all support a range of display screen connectors. Most importantly, Nvidia and AMD are the only 2 vendors releasing graphic processing unit (GPU) chips in the world with discrete PCIe card interface.
b) This appraisal is mainly on the superficial specification released by the PCIe card vendor using GPU chips from either Nvidia or AMD. Their specification is slightly different from the GPU vendor in order to fit in with their implementation. On this occasion, the card vendor is Asustek of Taiwan. The cards chosen from Nvidia is GTX series based on their latest Maxwell technology as at 2015-10, and the cards chosen from AMD is Radeon7 and Radeon9 based on their GCN 1.1 technology. The chosen cards are not the most powerful from each vendor. They are chosen for their mainstream appeals and are positioned at the medium to high end of the range.
c) Key hardware features selected for comparison are:
• Number of compute cores. They are not compute units or streaming processors. Compute cores in the same compute unit work on the same task. The number of compute unit in each card is shown in brackets. More compute units are better for single or multiple task computing. Similarly more cores per unit are better for the performance on a single task. Radeon based on GCN1.1 has 64 cores in one compute unit in AMD terminology. GTX based on Maxwell has 128 cores in one stream processor in Nvidia terminology.
• Amount of global memory and transfer bit width. Both ranges of cards use GDDR5 memory which is based on DDR3 but with more controllers for higher bandwidth. Some cards have more controllers than others to increase the memory transfer bit width. The amount of memory is vital to the size of the problem to be solved, and the memory transfer bit width is vital to the speed of overall task processing.
• Frequency and wattage are related. High frequency will incur higher wattage. The GPU cards are normally designed with a defined max operating temperature which would be equivalent to the max wattage consumed by the components on the card. Some cards consume as much as 300W. They are like furnaces within the PC box!
• Lastly the P figure is an indication of the relative selling price of the cards to each other. For example, P90 is the most expensive and P20 is the cheapest and is about 2/9 of the most expensive card. The figures are rough and not accurate or reliable.
d) Once the above notes have been read, each reader will form an opinion of the value of each card based on vendor’s superficial hardware features. The real value of the card would depend on the actual performance of software applications and the range of software applications available. Some software applications are free and some are very expensive. Enough bombs for today.
e) Examples of more bombs are the size of cache in each GPU (Maxwell has 2MB L2 cache. GCN1.1 has 1MB L2 cache) and the peak processing capability (GTX980 for 4.616FGLOPS SP and R9-390 for 5.120GFLOPS SP).