The NVIDIA GeForce GTX 980 Review: Maxwell Mark 2
by Ryan Smith on September 18, 2014 10:30 PM ESTSynthetics
As always we’ll also take a quick look at synthetic performance. These tests mainly serve as a canary for finding important architectural changes, and with the exception of pixel throughput we are not expecting any major changes for GTX 980 and GM204.
GM204 is designed to have an ever-so-slightly higher triangle throughput rate than GK110 – 16 tris/clock versus 15 tris/clock, and sure enough the GTX 980 comes out on top in TessMark, slightly edging out the GTX 780 Ti. The difference is only very slight here, and though GM204 should be a bit more powerful than GK110 in practice it’s a dead heat.
Moving on, we have our 3DMark Vantage texture and pixel fillrate tests, which present our cards with massive amounts of texturing and color blending work. These aren’t results we suggest comparing across different vendors, but they’re good for tracking improvements and changes within a single product family.
Beginning with Maxwell NVIDIA reduced their texture-to-compute ratio from 12:1 to 16:1. As a result of this change Maxwell GPUs have fewer texture units than comparable Kepler GPUs. Compounding this effect is the fact that Maxwell CUDA cores are more efficient than Kepler CUDA cores, leading to NVIDIA placing fewer cores overall and further reducing the texture fill rate.
As a result the GTX 980 is not texture fillrate competitive with any of the GK110 cards. It is competitive with the GK104 cards, but only because these cards had the same number of texture units at 128. NVIDIA has told us that they believe this new ratio is a better fit for modern workloads, and judging from the performance we’re seeing elsewhere it would appear that NVIDIA is right.
On the other hand, thanks to NVIDIA’s newer 3rd generation delta color compression technology, our 3DMark pixel fillrate performance is through the roof. GTX 980 comes very close to doubling the throughput of our GK110 cards and more than doubles the throughput of the GK104 cards, reflecting the fact that it has 64 ROPs and more importantly has the available bandwidth to put them to good use.
This benchmark in a nutshell is why NVIDIA can deliver chart-topping performance despite having only 2/3rds the memory bandwidth of GTX 780 Ti. By improving their color compression to this point, NVIDIA can significantly reduce their memory bandwidth requirements Maxwell 2, allowing them to do more with less. In real games the result won’t be anywhere near this remarkable since this is a pure pixel fillrate test, but it goes to show that NVIDIA has been able to expand their effective memory bandwidth in concert with their ROP and shader performance improvements.
Facebook
Twitter
RSS
274 Comments
View All Comments
garadante - Thursday, September 25, 2014 - link
Yeah. To be honest nobody except ardent Nvidia fanboys would've believed Nvidia would release cards as performance and price competitive as they did, especially the 970. The 980 is honestly a little overpriced compared to a few generations ago as they'll slap a $200 premium on it for Big Maxwell but $330 MSRP for the 970 (if I remember correctly) wasn't bad at all, for generally what, 290/780/290X performance?tuxRoller - Friday, September 26, 2014 - link
It's not too surprising as we saw what the 750ti was like.What is disappointing, though, is that I thought nvidia had made some fundamental breakthrough in their designs where, instead, it looks as though they "simply" enabled a better governor.
garadante - Friday, September 26, 2014 - link
It'll be interesting to see how the efficiency suffers once nvidia releases a proper compute die with area dedicated to double precision FP. I have to keep in mind that when factoring in the stripped down die compared to AMD's 290/290X cards, the results aren't as competition. Lowing as they first seem. But if AMD can't counter these cards with their own stripped down gaming only cards then nvidia took the win this generation.tuxRoller - Friday, September 26, 2014 - link
That's an excellent point. I take it you already read the tomshardware review? They're compute performance/W is still good, but not so unbelievable as their gaming performance, but I'm not sure it's b/c this is a gaming only card. Regardless, though, amd needs to offer something better than what's currently available. Unfortunately, I don't think they will be able to do it. There was a lot of driver work than went into making these maxwell cards humgaradante - Friday, September 26, 2014 - link
One thing that really bothers me though is how Anandtech keeps testing the 290/290X with reference cards. Those cards run at 95 C due to the fan control profile in the BIOS and I remember seeing that when people ran those cards with decent nonreference cooling in the 70 C range that power consumption was 15-20+ watts lower. So an AMD die that sacrifices FP64 performance to focus on FP32(gaming, some compute) performance as well as decreasing die size due to the lack of FP64 resources seems like it could be a lot more competitive with Maxwell than people are making it out to be. I have this feeling that the people saying how badly Maxwell trounces AMD's efficiency and that AMD can't possibly hope to catch up are too biased in their thinking.tuxRoller - Saturday, September 27, 2014 - link
Do you have a link to those reviews that show non-reference fans make gpus more efficient? I don't know how that could be possible. Given the temps we're looking at the effects on the conductors should be very, very small.Regarding the reduction in fp performance and gaming efficiency, that's a good point. That may indeed be part of the reason why nvidia has the gaming/compute split (aside from the prices they can charge).
garadante - Sunday, September 28, 2014 - link
Here's an example of a card with liquid cooling. Factor in the overclock that the nonreference card has and that it draws something like 20 watts less in Furmark and the same in 3Dmark. I could be mistaken on the improved power usage but I do recall seeing shortly after the 290X launch that nonreference coolers helped immensely, and power usage dropped as well. Sadly I don't believe Anandtech ever reviewed a nonreference 290X... which is mind boggling to consider, considering how much nonreference cooling helped that card, even outside of any potential power usage decreases.garadante - Sunday, September 28, 2014 - link
http://www.tomshardware.com/reviews/lcs-axr9-290x-... Whoops, forgot the link.jman9295 - Friday, September 26, 2014 - link
I wonder why they still give these cards these boring numbered names like GTX 980. Except for the Titan, these names kinda suck. Why not at least name it the Maxwell 980 or for AMD's R( 290 series the Hawaii 290. That sounds a lot cooler than GTX or R9. Also, for the last several generations, AMD and Nvidia's numbering system seems to be similar up until AMD ended that with the R9/R7 200 series. Before that, they had the GTX 700 and HD 7000 series, the GTX 600 and HD 6000 series and so on. Then, as soon as AMD changed it up, Nvidia decides to skip the GTX 800's for retail desktop GPUs and jump right up to the 900 series. Maybe they will come up with a fancier name for their next gen cards besides the GTX 1000's.AnnonymousCoward - Saturday, September 27, 2014 - link
Naw, names are much harder to keep track of than numbers that inherently describe relative performance.