GTX 950 vs R7 370 - How is the GTX faster?

[reallyscrued]

I'm looking for a decent budget card for <$150. I root for the red team and have always assumed the underdog would give the most bang for buck considering they need to put the extra initiative and edge into staying competitive, especially in the midrange.

In which case, when looking at specs, I thought it would be pretty clearcut that the R7 370 would wipe the floor with the GTX 950 considering the memory bandwidth of the R7 is nearly double, but in every benchmark I've seen, the GTX 950 is near 10 - 20 % faster in the real world.

I don't understand how this is possible as the GTX 950 has a gimped 128-bit memory bus. Can anyone shed some light on this? I understand the R7 370 is basically just a 7870 but is the architecture so old that its brute numbers fall wayside?

Source: http://techreport.com/review/29061/nvidia-geforce-gtx-950-graphics-card-reviewed

 

[crisium]

370 isn't even a 7870, it's a 7850 with an overclock.

Memory bandwidth isn't everything, especially since Maxwell V2 has memory compression and GCN 1.0 does not. There's no shocker when you compare the two architectures' efficiency with other cards. Of all the GCN 1.0 GPUs, probably the 7950 Boost compares closest to the 960. Relative to each other, the 7850 is cut down much more than the 950 is.

 

[SPBHM]

Maxwell is a lot more efficient, and higher clocked

look at the gtx 970, it has a 224bit memory bus and 1664 SPs and it's similar to a Radeon R9 390 with 512bit memory bus and 2560 SPs

the 370 is GCN1.0 with just 1024 SPs (a 7850 with higher clocks), the 950 is Maxwell 2 with 768, the 370 should be slower.

 

[Piroko]

... but is the architecture so old that its brute numbers fall wayside?

A bit at least, yes. GCN in general hasn't been too great in turning theoretical numbers into fps, at least compared to Maxwell.

Additionally, Maxwell (GTX 900 series) makes more efficient use of its bandwith thanks to color compression. The newer GCN 1.2/GCN3 cards (only the Fury and HD380 series) have added this as well. That is why you can now buy the HD380x with 2048 shaders and 256 Bit memory controller.

And the last reason is that Pitcairn ( R7 370 series) has more bandwith than it needs. I believe one Anandtech poster has tested this in the past and didn't find much of a performance difference with lowered memory clocks.

 

[dogen1]

Look at the synthetics page. The 950 matches or beats the 370 in everything except memory bandwidth, with large advantages in pixel fill rate, tessellation and triangle rate. In addition to compression, Maxwell also added more l2 cache to further reduce how much bandwidth it needs.

 

[Silverforce11]

 

[Azix]

look for a 285 or 280

the r7 370 is not really well placed in price but I guess they had their reasons. They launched a 370x in china and probably are selling defective versions of those as 370.

Here is 370 vs 270x.

https://www.youtube.com/watch?v=jXaWUuK-Tzk

also note 380 vs 960.

There's years of difference in the chips but that the 370 is close is a good thing at least. Its a mess but there is always some AMD card you can buy that makes the 950 pointless in that price range.

 

[GodisanAtheist]

I was going to post this. 370 isn't slower than the 950 when averaged across a large suite of games. The 950 wins some, the 370 wins others.

AMD's GCN arch is really a thing of beauty with the legs that it has...

 

[Yuriman]

$150 is tough to swallow for a 370, considering that's about what I paid for my HD7850 - which is basically identical - just under 3 years ago, when it was close to a year old already.

 

[Silverforce11]

$150 is tough to swallow for a 370, considering that's about what I paid for my HD7850 - which is basically identical - just under 3 years ago, when it was close to a year old already.

Hah, think about the 7950, 7970 (Sub $200), R290/X (~$200 to $250) prices for a long time already and then compare to 280/X, 380/X, 390/X prices...

Stagnation on 28nm is very lame for gamers.

 

[3DVagabond]

Hah, think about the 7950, 7970 (Sub $200), R290/X (~$200 to $250) prices for a long time already and then compare to 280/X, 380/X, 390/X prices...

Stagnation on 28nm is very lame for gamers.

When people lined up in droves to pay $400 (970) and $550 (980) for basically the same performance a light bulb went off over AMD's head.

 

[master_shake_]

When people lined up in droves to pay $400 (970) and $550 (980) for basically the same performance a light bulb went off over AMD's head.

my bad...

sorry.

but it was a new card

and shiny too!

 

[3DVagabond]

my bad...

sorry.

but it was a new card

and shiny too!

Forgot about the shiny part. That makes sense.

 

[Flapdrol1337]

With the 7970 an 7950 having 384 bit memory I guess it made sense for the 7870 and 7850 to have 256 bit, but the 370 probably doesn't require all of it.

maxwell is very efficient in using the memory bandwidth, so the 950 can perform similar with higher clocked memory on a 128bit bus.

 

[3DVagabond]

With the 7970 an 7950 having 384 bit memory I guess it made sense for the 7870 and 7850 to have 256 bit, but the 370 probably doesn't require all of it.

maxwell is very efficient in using the memory bandwidth, so the 950 can perform similar with higher clocked memory on a 128bit bus.

I'm sure with the lossless compression they've improved performance in 4 years.

 

[96Firebird]

When people lined up in droves to pay $400 (970) and $550 (980) for basically the same performance a light bulb went off over AMD's head.

970 launched at $330.

 

[MiddleOfTheRoad]

$150 is tough to swallow for a 370, considering that's about what I paid for my HD7850 - which is basically identical - just under 3 years ago, when it was close to a year old already.

Agreed, the 380 can be purchased for around that price.

 

[3DVagabond]

970 launched at $330.

Correct. I remembered wrong.

 

[reallyscrued]

Sounds like memory bandwidth is helped immensely by this "color compression" technique on the GTX. I know one poster said memory bandwidth isn't everything but the GTX is the *only* line of cards I've ever encountered where a narrower bus is not slower in any games but actually faster in some.

I suppose my next question is why can't this compression technique be applied in software as drivers or a firmware flash?

 

[LTC8K6]

Sounds like memory bandwidth is helped immensely by this "color compression" technique on the GTX. I know one poster said memory bandwidth isn't everything but the GTX is the *only* line of cards I've ever encountered where a narrower bus is not slower in any games but actually faster in some.

I suppose my next question is why can't this compression technique be applied in software as drivers or a firmware flash?

AMD has the same color compression as of GCN1.2

It was why the R9-285 was not too bad at intro even with only a 256 bit memory bus.

 

[dogen1]

Sounds like memory bandwidth is helped immensely by this "color compression" technique on the GTX. I know one poster said memory bandwidth isn't everything but the GTX is the *only* line of cards I've ever encountered where a narrower bus is not slower in any games but actually faster in some.

The 8x increase(4x in the case of the 960 and 950) in L2 cache also makes a big difference.

 

[tential]

970 launched at $330.

Given the state of the R9 290 drivers the GTX 970 looked good. Now, we see with even MORE mature drivers, the R9 290 STILL had room to grow and is now definitely a faster card.

The problem with AMD isn't the hardware, it's the drivers/timing. The R9 290 with fully mature drivers would be great.

Nvidia gets a LOT out of their hardware day 1. So it looks great then and doesn't "grow over time"(Because they get most of the performance out when it matters.... the day a product comes out).

AMD takes forever to mature the drivers. If we throw the R9 290 drivers from when the GTX 970 came out vs the GTX 970 drivers then, now, the GTX 970 would demolish the R9 290.

 

[Headfoot]

Sounds like memory bandwidth is helped immensely by this "color compression" technique on the GTX. I know one poster said memory bandwidth isn't everything but the GTX is the *only* line of cards I've ever encountered where a narrower bus is not slower in any games but actually faster in some.

I suppose my next question is why can't this compression technique be applied in software as drivers or a firmware flash?

it almost certainly requires hardware in the chip to do the compression/decompression fast enough that it doesn't become a new bottleneck on the memory bandwidth. So it's got to be able to compress/decompress faster than the bandwidth of the memory, which is quite a lot. (this is applicable to a transparent "always on" gpu implementation that requires no knowledge of the feature by the end software e.g. games -- other memory compression techniques are certainly available). Also, any technique that would utilize shaders or other existing on chip resources on older chips not having fixed function hardware configured for the task would add overhead in both performance and energy, cancelling out the benefit.

On a different note, bus bit width isn't necessarily a good indicator of performance. In fact, specs are only indicative of performance within a single GPU family. Specs arent comparable between architectures, only raw measured performance figures like FPS.

 

[Azix]

can't assume the memory compression helps much just because the 370 has more bandwidth.

it could just not need much bandwidth. Its barely a 1080p card

 

[SPBHM]

can't assume the memory compression helps much just because the 370 has more bandwidth.

it could just not need much bandwidth. Its barely a 1080p card

it's a combination of a few factors, the 950 is more efficient with bandwidth (compression, cache) and it has 1GHz higher effective memory clock, so while the r7 370 memory advantage is there, it's not as significant as it might look at first just by 128 vs 256bit and the 950 is a faster GPU outside of that, that's why the 950 is clearly faster on some games (combined with Gameworks optimizations and so on)