We really, really want to like NVIDIA's latest range of cards... but the raw facts just keep getting in the way.
We've finally got our hands on the GTX470 card from GIGABYTE, and though it doesn't offer anything special like aftermarket heatsinks or games, it's still gonna be a port of call for a lot of people. It's the not-quite-top-end of NVIDIA's new wave of DX11 cards, so let's have a look at the specs.
In terms of processing power the GTX470 offers only 14 of the total 16 Streaming Multiprocessors that the GF100 die physically has, giving it a total of 448 CUDA Cores, otherwise known as shader units. These Cores run at a much slower stock speed of 607MHz, and unlike previous NVIDIA cards, the shader clocks are linked permanently to the core clock at a 1:2 ratio; as you overclock the core, you'll inadvertently overclock the shaders too. Memory capacity is an impressive 1280MB of GDDR5 running on a wide 320-bit memory bus, though is disappointingly clocked at an effective 3448MHz.
Thanks to this slowdown on both Cores and clocks, the power consumption of the card is significantly lower when compared to the GTX480 - sitting 35W lower at 215W. This may not sound like much, but it's just enough to remove the need for the gigantic (and quite frankly, ridiculous) radiating plate of the higher-end model, and give the card a relatively normal appearance. Cooling is performed by an aluminium-finned five-heatpiped heatsink kept within the black shroud, and even though it has a lower TDP, temps remain disappointingly high. The GTX470 idles at 47 degrees while generating 55.4dBA of noise; load becomes a little silly, with temps of 89 degrees and a volume of 68.6dBA. This is nowhere near quiet, and has a similar 'space heater' effect as the hotter model. And hey, winter is coming soon, so this might be attractive to you! The card also comes with two DVI ports and a mini-HDMI port, and curiously contains a 1.5m mini-HDMI to HDMI cable within the box. Also thrown into the mix are NVIDIA's PhysX tech demos, which are fun if slightly pointless.
Overclocking wasn't possible at time of writing thanks to existing overclocking tools not understanding what to do with Fermi: they could sometimes change memory clocks, but core clocks wouldn't stick, or even read correctly. GPUz 0.4.2 read the clocks correctly, but ultimately just told us what we already knew. This will be fixed in the coming weeks as tools are updated, so expect overclocking results to be included in the future.
Performance in games is pretty nice, performing faster than both the 5850 and the 5830 across every benchmark result - though not beating the former by much. Vantage scores seem impressive until PhysX acceleration is taken into account, and Unigine's scores seem nice as well. That is, until you look at the price. For this amount of money - or less! - you can grab an ATI 5870 that'll outperform the GTX470, be cooler, quieter and use less power. For those reasons, the verdict is disappointingly clear.
Memory MattersAs you may notice in our Disassembly Guide of the GTX470, the Samsung GDDR5 chips used in the reference card are rated for much higher clocks than they're actually running in either GTX4xx card. The reason behind this is supposedly NVIDIA's memory controller - while it's more than good enough to handle GDDR3, it hasn't matured enough to support GDDR5's insane datarate with the large memory bus they've used. As it stands the bandwidth is higher than most ATI cards.
Issue: 133 | February, 2012