Is the 290(x) Going to Outlive a GTX 970?

[Silverforce11]

It's already happening.

Newer titles have the R290X/390/X have a clear lead over the 970. Just look at recent reviews of FC4, ACU, SoM, Civ BE, Batman: AK, Ryse, Witcher 3.

In older titles, the 970 is on parity with the R290X/390 but it seems as modern games are developed for consoles (GCN) in mind, it makes it easier for AMD to optimize those games (even NV GameWorks titles, eventually) to run fast on their GCN dGPUs.

Coming soon to DX12 (later this year), Battlefront, Hitman, Deus Ex, all AMD GE titles, designed for consoles/GCN ground up. There's going to be a big swing towards AMD's GCN in most benchmarks that include these newer games and drop the older titles from their lists.

Then you have to question whether Pascal will receive the major focus from NV on optimization, like Maxwell 2 did to Kepler. With GCN, the core uarch remains unchanged and the current console will live for awhile yet, we can be more confident GCN will remain optimized.

 

[Ryan Smith]

This is a little bit off topic, but we're talking about aging of GPUs and Kepler came up. So I'm wondering what core clock speeds AT's GTX 780 maintains? I just fired up Shadow of Mordor @ Ultra settings with the texture pack installed and i'm seeing some odd results.

Anandtech's GTX 780 ???/??? - 47.9 FPS
My GTX 780 down-clocked to @ 1006/1502 - 53.49 FPS
My GTX 780 Stock out of the box @ 1124/1502 - 56.89 FPS
My GTX 780 Gaming overclock @ 1306/1656 - 64.65 FPS

Now I know we don't have the same system setup, but if anything Ryan has a much better system than mine. If I had to guess looking at these numbers, AT's GTX 780 doesn't even boost into the 1Ghz range on the core. I guess it's something to keep in mind when reading through the benchmarks. I guess it's good to see my overclocked 780 is keeping right up with a stock GTX 980 @ 1440p In this title at least...

Reference GTX 780, so 863MHz base, 900MHz boost, 6GHz memory clock. This specific card has a maximum boost clock of 1006MHz, so under any sustained load it tends to end up in the mid-to-high 900s.

And we don't have the ultra texture pack installed for SoM.

 

[Majcric]

"Outlive" is the key word here. Not which card is a few frames faster. Perhaps if a benchmark was to be done with a large number of games between the 7970 vs. 680. The 7970 would creep out a win by a few frames. But to me, it still didn't outlast the 680. Fact is both cards are 3+ years old. A lot of folks have moved on to newer and better things. And for those that haven't I don't think a 680 owner would want a 7970 or vice-versa. In the end both cards are still in the same tier of graphics performance.

Back to 970 vs 290 this will be a wash 2 to 3 years from now performance wise.

 

[TeknoBug]

The newer the Nvidia driver, the worse their older cards get. AMD sometimes make their older card better with newer Catalyst drivers. Of course Nvidia doesn't age well.

 

[Silverforce11]

"Outlive" is the key word here. Not which card is a few frames faster. Perhaps if a benchmark was to be done with a large number of games between the 7970 vs. 680. The 7970 would creep out a win by a few frames. But to me, it still didn't outlast the 680. Fact is both cards are 3+ years old. A lot of folks have moved on to newer and better things. And for those that haven't I don't think a 680 owner would want a 7970 or vice-versa. In the end both cards are still in the same tier of graphics performance.

Back to 970 vs 290 this will be a wash 2 to 3 years from now performance wise.

Outlive as in playable in newer games at decent settings.

7970 is a heck of a lot faster than 680 in Witcher 3 and other newer titles, at playable settings to boot.

Not to mention in some games, the 680 can't run Ultra Textures by the 7970 can. Such as the NV sponsored Watch Dogs.

 

[guskline]

I am lucky to own both a GTX980TI in my rig below and 2 Sapphire Tri-X R9290s in CF in my 3770k rig at 4.5Ghz.

The 3770k rig had a Gigabyte G1 GTX970 gamer in it but I sent it back due to the snafu of the 4GB Vram issue. I previously had 2 GTx670 FTWs in SLI in it but wanted a 4GVram card. I replaced the GTX970 with a Sapphire Tri-X R9 290. At the time (late January, early February of 2015) the performance favored the 970. Since that time the 290/290X has improved at a greater pace than the 970. Obviously the 970 has the clear edge in power usage.

With AMD's release of the 300 series, it's likely the 290 will have decent driver support.

With my recent purchase of the GTX980TI SC to run as a single card in the 5960X rig the dual Sapphire Tri-X 290s were freed up. I put 1 of the two in my 3770k rig to run in CF and will shortly be selling the remaining one along with the 290 waterblocks.

All in all if you have a 290 that you got for a decent price it appears to be a solid buy. The GTX970 was a better performer at release and still is a smooth card.

 

[Sabrewings]

*Caution, total speculation below*

Owners of R2xx cards, and GCN cards in general, are for sure in a good spot for the future simply due to the similarity to the R3xx series. Nvidia historically hasn't adopted such a reuse scheme to its architectures and as such would have to put work into further optimizations for older cards, which no one is buying anymore. While this has a negative impact on users who squat on cards for a long period of time, this allows them to continue refining their architectures for efficiency. The jump from Kepler to Maxwell, while not beneficial to Kepler owners, provided huge gains to Maxwell itself through architectural reorganization of the SMXs into SMMs. Nvidia obviously put a huge amount of R&D effort into eeking every last performance out of their design transistor for transistor. This shows when compared to AMD's latest offerings. Watt by watt, transistor by transistor, Nvidia's architecture is more efficient and powerful.

Here's where my speculation begins. If I were Nvidia, and I had dumped that much R&D into Maxwell due to being stuck on 28nm another couple years, I wouldn't be so keen to throw it away for a new architecture. I'd be willing to bet that Pascal will be very similar to Maxwell, simply with a node shrink, updated architectural feature sets, and different memory controllers for HBM. They achieved a home run in terms of architectural efficiency with their current SMM design and I'm not sure how much more they could really find. In that regard, holding on to 28nm yielded them bonus time to really engineer the architecture to its best. Improvements can always be made, and 14/16nm is a different process entirely so some of the transistor level gains won't be there as they were on the mature 28nm, but it stands to reason that keeping all of that hard work with some tweaks going into Pascal is sound business strategy.

TL;DR - I suspect Pascal's SMMs will be very similar to Maxwell's, which will be a boost to Maxwell based cards in the future for the same reason AMD's GCN cards continue to improve.

Disclaimer: Yep, I have a 980 Ti and would love for the above to come to fruition. If not, it doesn't bother me as I'm strictly a performance biased consumer (with a few side concerns for implementation in my system). If my 980 Ti is not keeping up for any of my uses it doesn't bother me to buy the next halo card, probably Nvidia because low heat output is important to me. Though, I'm not opposed to buying AMD's products in any way if they're competitive and meet my needs.

 

[poofyhairguy]

You bring up a great point and I wondered that too. Nvidia needed Maxwell to basically emulate the benefits of Kelper on a smaller node since the node change is taking forever. When you look at how big of leap Pascal will be (new node, new memory, etc) you would think that they would want to change as little else as possible for time to market reason. Basically the Intel tick tock.

In reality though I don't think that is a real safety net to Maxwell users, as I think the Pascal generation will be such a leap that something like a GTX 960 becomes a lowend part overnight. According to this thread the real safety net is the Apple effect for Nvidia, aka the concept that a 970 owner could sell for more than a better performing 290x is worth if in a few years the 290 series becomes the clear retrospective a winner.

 

[Hi-Fi Man]

The 7970 was a competitor to the GTX 680 when the 7970 launched.

From your own charts, in the first 4 bench's it lost to the 680. The next 4 it wins, then loses 1, then wins 1.

That looks like a tie to me, and it actually beat the 680 in most games at launch. So based on that, it's actually gone backwards.

http://www.tomshardware.com/reviews/radeon-hd-7970-ghz-edition-review-benchmark,3232-19.html

Advantage longevity for 7970 vs 680 not found.

Found, http://www.anandtech.com/show/5699/nvidia-geforce-gtx-680-review/13

680 was on average faster at release and was the reason why AMD released the GHz edition. In my previous post I stated they were now tied. Keep in mind I'm not arguing for the 7970s longevity, I'm stating that it just seems like it has more longevity to some people because it has 3GiB of RAM and AMD's drivers weren't up to snuff at launch like nVIDIA's were.

 

[Sabrewings]

In reality though I don't think that is a real safety net to Maxwell users, as I think the Pascal generation will be such a leap that something like a GTX 960 becomes a lowend part overnight.

In that regard, it's going to hit both camps equally as hard. The next halo card could have 50-100% as many transistors as the 980 TI. It doesn't matter which side of aisle you're sitting, compared to the latest and greatest it'll be fractional performance.

Though, I would consider this. I don't think there are many developers out there aside from CIG that really have a game in development that has the promise to really be able to push such a leap in hardware. I wouldn't be surprised if we see a "GTX 1080" next summer, followed by a "GTX 1080 Ti" by holiday 2016. The first real games to push these new cards to their limits (and make them worth the purchase) will be around then. However, VR has the potential to totally disrupt this as hardware like StarVR comes out (5120x1440 resolution which is nearly 4k brutal in terms of pixel count) with the need to push 90FPS or better.

 

[richaron]

I find recent driver improvements (with AMD cards 2-3-4 years old constantly gaining performance) and nVidia's software strategy odd; since nVidias brand loyalty is mostly based upon the "experts"/pros/loudmouths who loved nV 10-15 years ago.

Back then I'd buy an nV card because I knew it would keep getting better as time goes by.

Now with AMD's recent driver history it's pretty obvious the tables have turned. I can only assume nVidia is using drivers to push their fans to upgrade more often (100% against the reason I loved nV). Vise versa, like we have seen for the last few years, AMD cards will continue to perform better with time.

 

[Sabrewings]

I can only assume nVidia is using drivers to push their fans to upgrade more often (100% against the reason I loved nV). Vise versa, like we have seen for the last few years, AMD cards will continue to perform better with time.

That's not the only thing you can assume. See my post above.

The gist is that AMD has been using revisions of the same basic architecture for years. Nvidia has been making large jumps in architecture design to refine it for performance and efficiency and as such each architecture requires its own driver optimizations since compatibility between them is not maintained.

Nvidia isn't deliberately hurting older cards; they're simply a casualty of their current design strategy. As a particular architecture gets older, their motivation to spend time specifically on that architecture for optimization decreases. On the AMD side, optimizations for their current gen also affects several previous generations due to running on the same basic design. Hopefully one can hope they'll soon reach a design they're fairly happy with and choose to iterate off of it instead of completely reorganize. The jump from Kepler to Maxwell was pretty large in terms of the internal changes to SMXs (into SMMs) and it paid quite large dividends. I can't personally fault Nvidia for the strategy as it is paying off. They do as much or more as AMD with less (less power, less cores, less transistors, less bandwidth) and that will put them in a great position with a die shrink for the next node alone without any architectural changes (which Pascal will surely offer if only in feature sets and HBM2).

 

[richaron]

*quote*

You could well be correct. You're saying the jump from Kepler to Maxwell was some sort of quantum leap as apposed to the incremental updates to GCN. So perhaps the software couldn't be incrementally improved?

Fair enough. nV drivers are also well known to be mature at release (opposed to 7970 performance release vs 6 months later), but how long has Kepler's software had to be improved upon? Less motivation to support an older prpduct no doubt, but why do you think there are so many fans spewwing the "better drivers" rhetoric? A huge part of it is based on ongoing support.

AFAIK the design of a new architecture shouldn't stop the ongoing support & optimization of past products. At the very least it should be obvious that the myth of nVidia's driver support is totally wrong these days.

EDIT: I'm sorry if that came off as a huge anti nVidia rant. My post is against the perceived superiority of nV's ongoing support and their business morals in general. No one can argue they have some kick ass cards.

 

[Sabrewings]

You could well be correct. You're saying the jump from Kepler to Maxwell was some sort of quantum leap as apposed to the incremental updates to GCN. So perhaps the software couldn't be incrementally improved?

It's widely known that, CUDA core for CUDA core, Maxwell is more powerful. That indicates fairly serious changes under the hood at the low level. From a high level of organization and how the driver interacts with the GPU, the SMX was seriously reorganized into what became the SMM. Ryan Smith goes into this with his Maxwell review article when the 980 launched. There's also some good diagrams which show how seriously Nvidia changed Maxwell's structure compared to Kepler.

http://www.anandtech.com/show/8526/nvidia-geforce-gtx-980-review/2

Such large changes would mean that they can't use the same driver software branch to support both.

AFAIK the design of a new architecture shouldn't stop the ongoing support & optimization of past products. At the very least it should be obvious that the myth of nVidia's driver support is totally wrong these days.

It doesn't "stop" it, but it means you have to break away engineers from your current products to support past products which aren't even on sale anymore. Yes, they have face to save, but after a few years they're going to pretty much leave them as is. This didn't bother me with my GTX 275 which I was still using up until a couple months ago.

AMD would be in the same position if they hadn't settled on an architecture and stuck with it for several generations. The point which they leave GCN (or move to a totally seperate iteration of it, i.e. GCN2.0 or whatever they call it) will leave AMD owners in the same state. The proof is in the past prior to GCN (which haven't improved like GCN has, due to the trickle down to the GCN enabled generations).

I have faith Nvidia will get to a similar position sooner or later with a successive generation architecture they iterate off of. Whether Maxwell was it, it's hard to say for certain. It's a good candidate given how much work they put into it and the results they were able to achieve. It's really up to the Nvidia engineers where they feel they want to stop. If they think they can squeeze even more performance from another architecture rather than iterate off Maxwell, then they'll continue to move forward. It is more economic for them to stop and iterate, though. The R&D spent on large redesigns is much more than iterations like AMD has been doing.

 

[tg2708]

The only decisive win Nvidia has over AMD and probably will for possibly a long time is power efficiency or anything centered around power management. AMD might take the performance crown later on as drivers mature for the top end and could possibly win by a good margin as the performance per dollar crown. But in all honestly in efficiency that's where AMD will most likely have to trek up a sizeable mountain. I still have certain bias towards Nvidia but let's hope history is not repeated where their top end is now a mid range sooner rather than later and can no longer compete with AMDs high end due to neglect or possibly bad architecture revamp timing.
No matter how it's looked at if a 290 outlives another Nvidia card or perform better with time power efficiency will mean little to nothing if it can't maintain a good fps in future games Nvidia is at a disadvantage even though most pc gamers will not see it.

 

[Sabrewings]

No matter how it's looked at if a 290 outlives another Nvidia card or perform better with time power efficiency will mean little to nothing if it can't maintain a good fps in future games Nvidia is at a disadvantage even though most pc gamers will not see it.

Until it starts to affect their market share, it really doesn't matter. What matters is the performance of the current generation while it is the current generation. Not next year once Nvidia has something else out and AMD has had time to massage drivers. Most consumers don't care and will continue to upgrade as they see fit.

If, as I suspect, Nvidia settles down for a few generations on the architecture side in a year or two, that will be a nice bonus for their customers. However, unless they're really being made to suffer for it they won't until they're ready on the engineering side. Nvidia typically doesn't market to the last generation or two anyway. They go back a few generations as they understand the typical upgrade cycle.

 

[richaron]

*snip*

All you've said are that nV's hardware improvements are such that they don't need/want/afford to keep working on their software for past products?

AMD has made obvious & consistant improvement in software support over the past few years.

I'm hesitant to be completely negative but your main focus is a belief nV will improve their support next node shrink...

 

[Seba]

There is a point where no realistic driver improvements can be made. What if someone reaches that point sooner than the competition?

 

[thesmokingman]

There is a point where no realistic driver improvements can be made. What if someone reaches that point sooner than the competition?

Then the card you purchase is only good for two years? :thumbsdown:

 

[Seba]

Then the card you purchase is only good for two years? :thumbsdown:

Do you have one example where new drivers made such a difference that the card had some spectacular performance improvements after 2 years?

 

[CriticalOne]

Then the card you purchase is only good for two years? :thumbsdown:

No, it just means that there isn't really much optimization you can do. People seem to think that optimization is some sort of unlimited magic that gradually makes a GPU faster.

AMD has been on GCN for 4+ years. AMD doesn't have a lot of money, so their driver improvement is more gradual than that of NVIDIA. We all know that the release drivers on the 7970 weren't great seeing how it took months for it to eventually be faster than the 680. It also took them a while to add proper frame pacing (however it remains absent for D3D9 still), and AMD drivers still suffer from high D3D11 overhead.

 

[Face2Face]

As a 780 owner, I'm going to have to call you out on the fact that 780's don't have enough ram to run Mordor smoothly with ultra textures at 1440p. there is clear hitching. The framerate looks fine, until you flick the camera, then there's some obvious stutter. It can do it at 1080p alright as far as i can tell, but not 1440p. High textures have to be used in order to avoid it.

I mean sure, the game is [poorly optimized], but still.

That's exactly what I thought or expected. I'll have to do more testing, but for the 10-15 minutes I played with the Ultra texture pack along with ultra settings, I didn't notice any stutter or hitching. When I get some more time I'll record the frame times to verify, but I certainly was expecting hitching or stutter of some sort.

Reference GTX 780, so 863MHz base, 900MHz boost, 6GHz memory clock. This specific card has a maximum boost clock of 1006MHz, so under any sustained load it tends to end up in the mid-to-high 900s.

And we don't have the ultra texture pack installed for SoM.

Thanks for the clarification Ryan. I should of looked back at your GTX 780 review to confirm the clock speeds. Also, it's good to know that you don't use the high resolution texture pack for testing purposes.

 

[iiiankiii]

There is a point where no realistic driver improvements can be made. What if someone reaches that point sooner than the competition?

That highly unlikely because there are always new games coming out. It's a matter of continuing support. There are ways to extract more out of the hardware for that game. It's up to the hardware vendor to continue to provide software support for as long as possible.

 

[poofyhairguy]

It's already happening.

Newer titles have the R290X/390/X have a clear lead over the 970. Just look at recent reviews of FC4, ACU, SoM, Civ BE, Batman: AK, Ryse, Witcher 3.

That shift seems very resolution dependant. Where GCN seems to really shine is resolutions greater than 1080p, that is where we see the real separation in the new games.

That is good for AMD though, that is where the market is going. Until VR is a real thing the best reason to drop big money into a gaming PC over a console is 4k. That is the part of the market where a lot of the margin is made, people buying multi-GPUs rigs to run 4k.

 

[moonbogg]

The 290X will probably out live everything. That card won't die. As for me, as soon as Pascal drops, I wouldn't have permission to complain if my 980TI's suddenly couldn't manage to run minesweeper.