Monday, May 9, 2016

You Don't need a new video card Part 2 The GTX 1080


So the big secret is out.  Nvidia has once again broken through the 3 digit naming convention for it's video cards and announced the GTX 1080 and 1070.


Nvidia is claiming leaps and bounds of the new Pascal based GPU over the previous Maxwell generation of cards.  That may be true in a limited context but the specs may be telling a different story.  Speaking of specs, take a look at the GTX 1080 compared to a GTX 980




What I'm seeing here isn't that earth shaking if you've purchased an Nvidia card in the past 3 years.  While the clock speeds, memory fill rates and power efficiency are marginally better the lack of HBM2 memory is probably the most glaring omission from Pascal's wish list.  

Nvidia opted for  GDDR5x memory which reportedly can "double" the bandwidth of GDDR5 by utilizing a 64 bit wide memory bus.  Other than that, however, There's not much more to say. 

Consider GDDR5x to be a Hot Rod version of GDDR5 that is pin compatible with the previous generation.  Sadly, that leaves out the even higher bandwidth and power savings potential offered by HBM2 which would explain the only moderate power efficiency gains of the 1080 vs the 980.  

The bottom line is this.  The GTX 1080 is currently a flagship card destined for the mid-range.  It's a stopgap that implements a new GPU until Nvidia can get a real flagship card online.  If you look at Nvidia's performance charts you'll see marginal improvements between the two cards with the only major difference being the handling of VR applications.  

This is more of a VR card than anything else.  As such it will likely replace the GTX 970 as the entry level Nvidia Graphics platform for VR applications.

Of course the Hype machine is running full bore hoping to relieve you of $600 to $700 (reference card) to be in on the latest and greatest.  The price point tells you this is a high-mid-range card with the next step down the GTX 1070 coming in at roughly half  that price.  

With all that I find myself dubious at Nvidia's performance claims mostly because of the fuzzy math

Such as the claim that the card is "75% faster"  

75% faster doing what? Just clocks?  Great, but that's like saying  that running your case fan at 4000 RPM is faster than running it at 2000 RPM.  That's great but if you're not getting any more airflow it's a meaningless number.

Keep in mind that Graphics performance isn't solely determined by the hardware.  Drivers, game coding, testing setup and even room temperature can make conclusions subjective at best.  Take it all with a grain of salt.

Then there's the claim that 1 GTX 1080 is faster than 2 980's in SLI.  

Again, meaningless without context.  What were they doing?  That kind of claim  is pretty much  meaningless these days since most recent benchmarks haven't shown a great performance improvement (which I classify as 40% or more) over a single card.

My rule is that for every extra card I run should see a performance improvement of at least 50% in graphically intensive operations that should benefit from better graphics hardware.  

Most of the industry pundits would say that's an unrealistic expectation and they would be right.  In fact adding anything more than 1 card in SLI automatically engages the law of diminishing returns exponentially.

Regardless, the industry is moving away from Multi GPU's anyway so this is kind of a weak metric for Nvidia to tout.

Multi-card setups only came about as a crutch for comparatively mediocre GPU's to keep up with more demanding games and 3D applications.  It was sold as performance but it was really just a stopgap till the technology could catch up.  

So...

If you can accept Nvidia's claim at beating 2 980's in SLI with the GTX 1080 then you also have to accept that the average SLI setup nets maybe a 20% improvement at higher resolutions with a modern GPU.  

In fact the only real reason to invest in SLI these days is running games with extreme graphics settings at extremely high resolutions ( like 4K or more. )

Meaning 90% of the time besting SLI numbers is no big deal with a new generation of graphics processor regardless of the vendor.

Conclusion:  This claim is bullshit...

So what do I think?  

Simple.  If you've bought into anything above a GTX 960 I'd wait for the next refresh of Pascal before upgrading.  Perhaps by then they'll have HBM2 with its higher bandwidth potential and power savings. The first crop of Pascal gaming cards looks more like an interim refresh than a sea change over Maxwell from a performance standpoint.

If, however, you're running a Pre-Maxwell vintage card this may be an ideal time to upgrade.  The marked improvements in memory performance and power efficiency alone are worth the upgrade especially if you expect to be doing ANYTHING in 4K.  Besides, it's been about 3 years which is just about right for an upgrade.  

So what have we learned?  

Nvidia is full of bullshit.  So is AMD.

It's perfectly OK to skip an entire generation of video cards.

The GTX 1080 will be a flagship card for about 6 months before it gets relegated to the mid-range.

So if you're still rocking a pair of GTX 670's in SLI it may be time to check out the new GTX 1080.

The video below from the Hardware Canucks gives a decent overview of the new GTX 1080 although they do tend to buy into the hype machine. 



Then there's my take...



No comments: