facebook rss twitter

Review: Asus GeForce GTX 950 Strix Direct CU II OC

by Tarinder Sandhu on 20 August 2015, 14:00

Tags: ASUSTeK (TPE:2357), NVIDIA (NASDAQ:NVDA)

Quick Link: HEXUS.net/qactrr

Add to My Vault: x

Introduction

Nvidia has gained approximately 80 per cent of the discrete graphics card market according the latest research. Undoubtedly strong in the premium segment with products such as the GTX 980 Ti, GTX 980 and GTX 970, the green team's mainstream stack doesn't, well, stack up.

You see, there's the $199 (£150) GTX 960 GPU that, Nvidia says, is the killer card for full-on eye candy at the ubiquitous 1080p resolution. Those without that kind of budget need to look a step further down the line and there you will find the GTX 750 Ti that's based on first-generation Maxwell architecture and costs $129 (£100) or so. The yawning gap is obvious and painful for Nvidia because it's an area that rival AMD has exploited with a succession of reimagined GPUs.

Nvidia needs another GeForce to tempt those gamers with a $150-$170 budget, but constructing such a fit-for-purpose GPU isn't easy. The erstwhile mainstream champ, GTX 750 Ti, packs in the full complement of cores and shaders into the GM107 die so there's no scope there. Redesigning the second-generation Maxwell die for a particular price point is irksome, so Nvidia opts for the lesser of two evils by castrating a GPU already available to the masses.

GM206 to the rescue

That GPU is the GM206 as found on the GeForce GTX 960. Harvesting such silicon isn't the most economical method of adding a cheaper offering. GM206 is reasonably big, filling out to 2.9bn transistors and using a considerable 227mm² die area. As a comparison to three years ago, the range-topping GTX 680, fabbed on the same process and costing over £350 on arrival, measured in at under 300mm².

The full-fat GM206 die is harvested for GTX 950

Nvidia segregates the GTX 950 from the incumbent GTX 960 by removing two of the eight SMM units present in the full-fat GM206 die. It is therefore reasonable to assume that GTX 950 loses about one-quarter of everything, from cores, texture units, setup engines and so on and so forth. The backend continues to carry 32 ROPs connecting to system memory via a narrow 128-bit bus.

GeForce
GTX 950 (2GB)
Launch date
August 2015
January 2015
February 2014
October 2012
Codename
GM206
GM206
GM107
GK106
Architecture
28nm Maxwell
28nm Maxwell
28nm Maxwell
28nm Kepler
Transistors (billion)
2.9
2.9
1.9
2.5
Die Size (mm²)
227
227
148
221
SM Units
6
8
5
4
Cores
768
1,024
640
768
Texture Units
48
64
40
64
ROP Units
32
32
16
16
Base Clock (MHz)
1,024
1,126
1,020
928
Boost Clock (MHz)
1,188
1,178
1,085
NA
Peak GFLOPS
1,825
2,413
1,389
1,425
Memory Clock (MHz)
6,600
7,012
5,400
5,400
Memory Bus (bits)
128
128
128
128
Bandwidth (GB/s)
106
112
86
86
Power Connectors
6-pin
6-pin
None
6-pin
TDP (watts)
90
120
60
110
GFLOPS per watt
20.2
19.2
23.1
13.0
US Launch Price
$159
$199
$149
$129

The numbers speak for themselves. It's sensible to accompany the architecture cuts with an appropriate price snip, here evidenced by the most basic GTX 950 cards coming to retail at around $159 (£130). Such a retail presence puts GTX 950 squarely up against the Radeon R7 370 from AMD - a card that has enjoyed success against the GTX 750 Ti.

Comparing GTX 950 to previous generations shows there's considerably more shading power on tap though the backend isn't overly endowed with potential, because AMD plies the same segment with cards boasting a 256-bit bus and therefore more juicy bandwidth.

Pay little heed to the base and boost clocks of the GTX 950 GPU. Nvidia's partners rarely toe the company line, so much so that the reference frequencies bear little resemblance to a well-overclocked card from, say, Palit or Asus. Shaving performance has the positive ancillary benefit of reducing overall board power, now down to 90W and, should a partner be bold enough, low enough for passively-cooled solutions.

Yet numbers on a page aren't always the best proxy for determining in-game performance. Let's take a look at a partner card and then run the rule over the GTX 950.