AMD crossfire users - check this out (SFR instead of AFR)

[wand3r3r]

Check out the blog by Firaxis, it sounds pretty revolutionary for us crossfire users, presuming it works. The low level GPU control sounds like it will reap some serious benefits and it looks like mantle is getting some serious commitment. The mantle benefits are looking pretty amazing as long as it's not buggy. Have you guys tried it out?

Firaxis focuses on Civilization naturally, however this should be relevant with other games going forward. Bring on the smooth crossfire.
I look forward to bf5 if they use mantle and if it's not buggy like bf4.

http://www.firaxis.com/?/S=813e3681...antle-multi-gpu-for-civilization-beyond-earth

Current multi-GPU solutions are implemented in the driver, without knowledge of, or help from, the game rendering engine. With the limited information available drivers are almost forced to implement AFR, or Alternate Frame Rendering, which is an approach where individual frames are rendered entirely on a single GPU. By alternating the GPU used each frame, rendering for a given frame can be overlapped with rendering of previous frames, resulting in higher overall frame rates. The cost, however, is an extra frame of latency for each GPU past the first one. This means that AFR multi-GPU solutions have worse response time than a single GPU capable of similar frame rates.
...

we implemented a split-screen (SFR) multi-GPU solution for the Mantle version of the game. Unlike AFR, SFR breaks a single frame into multiple parts, one per GPU, and processes the parts in parallel, gathering them into the final image at the end of the frame. As you might expect, SFR has very different characteristics than AFR, and our choice was heavily motivated by our design of the Civilization rendering engine, which fits the more demanding requirements of SFR well. Playing the game with SFR enabled will provide exactly the same quality of experience as playing with a single, more powerful GPU.

We do not claim to have the perfect multi-GPU solution with Civilization: Beyond Earth, and our current implementation has does have some limitations. Our split-screen rendering probably won’t get the same frame rates as AFR techniques, and we are currently limited to 2 GPUs. Some machines (especially those with slower CPUs) may not see high performance gains, and the highest gains will be seen at higher resolutions. However, we believe the time is ripe for multi-GPU machines to provide a user experience just as good as single-GPU systems, and with the explicit multi-GPU control provided by AMD Mantle we have started working to make this a reality.

What does this result in?

From an Anandtech addendum.
http://www.anandtech.com/show/8643/...crossfire-with-mantle-sfr-not-actually-broken

Extremely high minimums

Unbelievably smooth XDMA mantle crossfire gameplay

So it appears to sacrifice the higher FPS with a frame of latency for smooth high quality FPS which performs identically to a single card with higher performance.

To me it appears to be a great alternative to dx and really highlights some potential that Mantle brings. It's interesting to see the resurrection of a technology and it appears to be a massive success.

As for the somewhat lesser FPS, I am very satisfied that I can choose to run a game in crossfire for more power, but it will not sacrifice gameplay and is revolutionary in how well it works. I hope they don't drop dx so you can choose to run with the benefits of AFR or SFR as you want, but I can see why they chose to optimize mantle the way they did. I haven't seen anything this substantial in dual card setups ... ever?

Keep the thread on topic please.

 

[Jodell88]

Anandtech had a write up earlier today.
http://www.anandtech.com/show/8643/...crossfire-with-mantle-sfr-not-actually-broken

 

[Erenhardt]

min FPS and frame times look very good with mantle. But average FPS scaling got hit hard in the process. Is it their (fireaxis) design decision. I guess there will be split in the opinions what matters more (min or avg FPS).

I guess they went the right direction, since that's where the market is steered by other parties.

Its good to see amd have fixed crossfire and improved a lot upon that.

It's cool to see the dev deciding how they want to use the hardware in the way that benefits gameplay the most.
I hope the amount of development around mantle will not decrease and we will see many more neat features.

 

[Wild Thing]

Interesting research he is doing with the Mantle API.
Thanks for posting wand3r3r :thumbsup:

 

[ShintaiDK]

The problem with SFR vs AFR is the rendering requirement per frame. Anyone can also use SFR with Mantle, OpenGL, DX etc. But practically you are more or less limited to 2 cards of relatively the same type and speed. Its just a lot less flexible. And thats why AFR got its popularity. SFR dates back to the 3Dfx era.

While SFR looks much better (at least in this implementation) the question begs, if people are willing to give up so much for so little. And it also raises the question about CF/SLI vs faster singlecard.

 

[flash-gordon]

This implementation should be amazing for FPS games.

 

[caswow]

The problem with SFR vs AFR is the rendering requirement per frame. Anyone can also use SFR with Mantle, OpenGL, DX etc. But practically you are more or less limited to 2 cards of relatively the same type and speed. Its just a lot less flexible. And thats why AFR got its popularity. SFR dates back to the 3Dfx era.

While SFR looks much better (at least in this implementation) the question begs, if people are willing to give up so much for so little. And it also raises the question about CF/SLI vs faster singlecard.

you are not looking at the whole picture. its not only about highend cfx. hybrid cfx works too with apus and low/mid range cards.

 

[Silverforce11]

The problem with SFR vs AFR is the rendering requirement per frame. Anyone can also use SFR with Mantle, OpenGL, DX etc. But practically you are more or less limited to 2 cards of relatively the same type and speed. Its just a lot less flexible. And thats why AFR got its popularity. SFR dates back to the 3Dfx era.

While SFR looks much better (at least in this implementation) the question begs, if people are willing to give up so much for so little. And it also raises the question about CF/SLI vs faster singlecard.

Even now, AFR in CF/SLI is applied with the same cards, but in CF there is some flexibility: ie. a 7970 can CF with a 7950.

 

[SPBHM]

it should be interesting to test input lag with some high speed camera

 

[ShintaiDK]

you are not looking at the whole picture. its not only about highend cfx. hybrid cfx works too with apus and low/mid range cards.

Hybrid is pretty much impossible with SFR. But hybrid have never been a success even with AFR.

If I had to use multicards, I would prefer SFR. But again, you lose flexibility, speed and so on. However in my mind the pros outweight the cons.

But again, its not going to set any benchmark records.

 

[monstercameron]

Hybrid is pretty much impossible with SFR. But hybrid have never been a success even with AFR.

If I had to use multicards, I would prefer SFR. But again, you lose flexibility, speed and so on. However in my mind the pros outweight the cons.

But again, its not going to set any benchmark records.

what is your source on the sfr technique capabilities of AMD cards?

 

[ShintaiDK]

what is your source on the sfr technique capabilities of AMD cards?

Split frame is nothing new. And the basic requirement (per card) that is also the reason for the performance loss (Full scene geometry, lightning etc for each card). So mixing it with a much slower card simply destroys performance below singlecard performance since the faster card will spend time waiting. Even with 2 equal cards the perfomance impact is quite high, since both cards needs to do part of the rendering as if it was an entire frame.

There is a reason why AFR got popular in the first place with basicly up to 100% scaling.

As you can see, in terms of the 290X. The SFR benefit for 100% more GPU power turns out at 20%. While the regular AFR method is far from potential. It does offer 42%.

 

[gorobei]

Split frame is nothing new. And the basic requirement (per card) that is also the reason for the performance loss (Full scene geometry, lightning etc for each card). So mixing it with a much slower card simply destroys performance below singlecard performance since the faster card will spend time waiting. Even with 2 equal cards the perfomance impact is quite high, since both cards needs to do part of the rendering as if it was an entire frame.

the articles indicate its proportional slices of the frame. so if you have a 290 with 2560sp and a 270x with 896sp, presumably if it is proportional the 270 would get 35% of the frame and the 290 would get the other 65% so that the buffer isnt waiting for the slower card to finish.

As you can see, in terms of the 290X. The SFR benefit for 100% more GPU power turns out at 20%. While the regular AFR method is far from potential. It does offer 42%.

why are you measuring avg fps when the point of sfr is to improve latency by improving min fps?

290x cf @4k dx= 20.92 minfps
290x cf @4k mantle = 37.46 minfps
sfr = 179% of afr

290x cf @1440p dx = 41.52 minfps
290x cf @1440p mantle = 63.23 minfps
sfr = 152% of afr

290x cf @1080p dx = 44.34 minfps
290x cf @1080p mantle = 64.42 minfps
sfr = 149% of afr

50-80% gains seems significant enough for devs to target if the game genre means that gamers want to avoid any lag spikes that might result in a death instead of a frag. while this is in cpu bound cases, most competitive players turn down graphics anyways.

 

[ShintaiDK]

the articles indicate its proportional slices of the frame. so if you have a 290 with 2560sp and a 270x with 896sp, presumably if it is proportional the 270 would get 35% of the frame and the 290 would get the other 65% so that the buffer isnt waiting for the slower card to finish.


why are you measuring avg fps when the point of sfr is to improve latency by improving min fps?

290x cf @4k dx= 20.92 minfps
290x cf @4k mantle = 37.46 minfps
sfr = 179% of afr

290x cf @1440p dx = 41.52 minfps
290x cf @1440p mantle = 63.23 minfps
sfr = 152% of afr

290x cf @1080p dx = 44.34 minfps
290x cf @1080p mantle = 64.42 minfps
sfr = 149% of afr

50-80% gains seems significant enough for devs to target if the game genre means that gamers want to avoid any lag spikes that might result in a death instead of a frag. while this is in cpu bound cases, most competitive players turn down graphics anyways.

Never questioned minimum FPS. But using minimum FPS is not always a correct measurement as shown before. If 1 frame out of a million dips, you get a low score nomatter what the 999999 others did.

In terms of slide per propertion. Its not quite that simple.

Remember each GPU still needs to handle geometry, lightning and so on as if they had the full scene. This is also why SFR was abandoned in the first place when rendering got more advanced. So it doesnt matter if a slower card got a lower slide, if it still cant keep up with the full scene parts and holds the faster GPU back.

 

[KingFatty]

Generally the trend seems to be that you lower the super-high max FPS, but increase the minimum FPS value. Seems like a good tradeoff, especially if you are stuck with a 60 Hz monitor?

 

[wand3r3r]

Thanks for the AT link. Updated OP with my thoughts anyway...

The beauty of this all is that we can choose whether to use SFR or AFR and have superior gameplay where desired or simply the high FPS great scaling crossfire.

 

[Silverforce11]

Ive been playing a fair bit of Civ BE today and definitely the minimum fps isn't a rare occurrence, it happens with late game, lots of units on screen, and if you zoom out a bit & pan across the map.

With Mantle, its butter smooth all the time, quite an amazing feat.

 

[BFG10K]

What's old is new again, like Mantle being "Glide 2".

Still, in this case it's a good idea, even though it has absolutely nothing to do with Mantle. Actually AMD should bring back the old super-tiling mode which provides automatic load balancing and doesn't need game profiles to scale. The FPS graphs don't look as pretty as AFR, but it's far more robust and simpler for the developer, for the user, and for AMD.

There's also a way to provide free super-sampling that automatically scales too, but neither vendor seems interested in going down that path.

 

[SirPauly]

Indeed! Past methods with CrossFire:

http://techreport.com/review/8826/ati-crossfire-dual-graphics-solution/3

Personally enjoyed the x10 setting with super-AA years and years ago-- virtually free performance for the x2 RGSSAA component of the hybrid mode!