mGPU - CFR vs AFR

[J7SC_Orion]

Admins please note: This thread is intended to settle somewhere between 'GPU' and 'Benchmarking'

 

...Ah, the joys of mGPU (multi-GPUs), such as NVidia SLI/NVLink and AMD Crossfire/Quadfire...and why CFR (checkerboard / tile-based) mGPU vs AFR (traditional Alternate Frame Rendering) matters...may be not now, but certainly in the not-too-distant future.

 

Yes, yes - there is the chorus that SLI/mGPUs "is dead"...while not entirely true, it certainly is the case that a single GPU will usually be far more painless to optimize for a given game, or really be the only option for other games...that said, I recently switched from four Quad-SLI / Quad-Fire systems to 'only' dual NVLink (2x 2080 TIs), and while I do not play as many games on this HEDT system that also does 'work' as others, I have relatively few problems with NVLink on my fav games I do play, such as various NFS and also Metro Exodus, never mind 3DM and other benches.

 

Yet this thread is NOT intended to convert anyone to mGPU. Instead, it is pointing out that there seems to be a movement aloft by various GPU producers (NVidia, Intel and likely AMD) to introduce 'mGPUs' in future generations of their GPUs. Think AMD Ryzen 7nm and soon smaller 'chiplets' vs Intel's difficulties with large 10nm 'all-in-one' giant and complex dies in the CPU realm. Likely, the next gens of GPUs will still be single die, but sooner or later, it will be mGPUs for the middle and upper class performance graphics cards - for which you need extremely-well performing mGPU drivers

 

As such, NVidia released, rather quietly, their CFR capability in their drivers as of late 2019 to do 'CFR' - for RTX only. CFR is actually not new, but was supplanted by the easier-for-developers 'AFR'. Yet with future GPUs, CFR seems to be the far more capable ticket for future mGPU generations than AFR...

 

 

 

 

Below are some early CFR vs AFR comps with the current gen of RTX GPUs. I already have run benchmarks of my own such as 8K Unigine Superposition with CFR vs AFR, but much more (tedious mod and setup) work is needed. I will update this post as I get more results of my own, time permitting...

 

In the meantime, I will note that CFR is not always faster than AFR in outright FPS, but it seems to have better frame times...and below are some YouTube vids someone else ran for DCX12 (Titan RTX) and Tom Clancy's Division 2 (DX11)...have fun and plan you next mGPU (oops )

 

 

 

 

[ENTERPRISE]

That is interesting, though I wonder why Nvidia has sat on it ? Possibly because A. It is not ready or B. It was not what they hoped for. I have been through the whole single gpu versus multi-gpu. Back in the day when I moved from single GPU to CF I was in awe of the additional performance and it was fantastic, the same when I moved over to NV with SLI. Even up to the 2080Ti's I remained with SLI, or in this case Nvlink. The issue for me was the 2080Ti multi-gpu was more of a "Maybe things will get better" type deal. We have all seen the decline in multi-gpu titles or the poor scaling if it was supported.

 

I remember during the 2080Ti launch Nvidia where stating that they would be working hard with devs and the likes to bring back the multi-gpu platform (for gaming) but I am yet to see it. I think that assuming the rumors are semi accurate about the 3080Ti, that will be the route I will go, but this time a single gpu build.

 

I think that the API fragmentation regarding DX, Vulkan, OpenGL has not helped matters either. It is great to have the competition for sure but multi-gpu has been even more hit and miss over these different API's. All in all multi-gpu for gaming is not dead, but it really is now a niche due to the lack of support and the fact you will see little return on your investment.

 

You never know Nvidia may pull something out of the hat, but whatever it is, it would need to make life so easy for devs to make it worthwhile, or to incentivize devs in some way.

[J7SC_Orion]

That is interesting, though I wonder why Nvidia has sat on it ? Possibly because A. It is not ready or B. It was not what they hoped for. I have been through the whole single gpu versus multi-gpu. Back in the day when I moved from single GPU to CF I was in awe of the additional performance and it was fantastic, the same when I moved over to NV with SLI. Even up to the 2080Ti's I remained with SLI, or in this case Nvlink. The issue for me was the 2080Ti multi-gpu was more of a "Maybe things will get better" type deal. We have all seen the decline in multi-gpu titles or the poor scaling if it was supported.

 

I remember during the 2080Ti launch Nvidia where stating that they would be working hard with devs and the likes to bring back the multi-gpu platform (for gaming) but I am yet to see it. I think that assuming the rumors are semi accurate about the 3080Ti, that will be the route I will go, but this time a single gpu build.

 

I think that the API fragmentation regarding DX, Vulkan, OpenGL has not helped matters either. It is great to have the competition for sure but multi-gpu has been even more hit and miss over these different API's. All in all multi-gpu for gaming is not dead, but it really is now a niche due to the lack of support and the fact you will see little return on your investment.

 

You never know Nvidia may pull something out of the hat, but whatever it is, it would need to make life so easy for devs to make it worthwhile, or to incentivize devs in some way.

 

 

...good point "that the API fragmentation regarding DX, Vulkan, OpenGL has not helped matters either. It is great to have the competition for sure but multi-gpu has been even more hit and miss over these different API's."

 

What I'm thinking, based on some rather preliminary and speculative pieces, is that future mGPU will not so much be separate cards (like current NVLink/SLI), but more like multiple (ie 5nm) chiplets on a single card, thus the renewed need for some mGPU drivers...and when properly implemented, CFR has some advantages over AFR. But as with anything else 'forecasting is difficult, especially when applied to the future'...

 

[ENTERPRISE]

Well they could go that route with a chiplet like design, however I am wondering if that makes it any easier for the devs. Unless Nvidia or AMD make it so that it is game agnostic and the GPU/Drivers decide how to best process the game ? I think the only way Multi-GPU will return or another guise of it is if the workload is taken away from the dev.

[J7SC_Orion]

Well they could go that route with a chiplet like design, however I am wondering if that makes it any easier for the devs. Unless Nvidia or AMD make it so that it is game agnostic and the GPU/Drivers decide how to best process the game ? I think the only way Multi-GPU will return or another guise of it is if the workload is taken away from the dev.

 

 

...along the lines of what happened with DLSS 1.0 > DLSS 2.0 ?

Per Nvidia's site, " The original DLSS required training the AI network for each new game. DLSS 2.0 trains using non-game-specific content, delivering a generalized network that works across games. This means faster game integrations, and ultimately more DLSS games."

 

[ENTERPRISE]

 

 

...along the lines of what happened with DLSS 1.0 > DLSS 2.0 ?

Per Nvidia's site, " The original DLSS required training the AI network for each new game. DLSS 2.0 trains using non-game-specific content, delivering a generalized network that works across games. This means faster game integrations, and ultimately more DLSS games."

 

I forgot about DLSS ! Yeah a technology akin to that would be the best way to go even with non chiplet design multi-gpu.

 

Just imagine if they achieved that, multi-gpu would then come back to the enthusiast level rather than a small niche.

[schuck6566]

 

I forgot about DLSS ! Yeah a technology akin to that would be the best way to go even with non chiplet design multi-gpu.

 

Just imagine if they achieved that, multi-gpu would then come back to the enthusiast level rather than a small niche.

 

my personal opinion is that if they expect to see multi gpu usage in rigs by peeps like me,they need to lower the price of the cards that can be run together. When I bought my RTX 2070,it was $600 after taxes and didn't support sli(the super the added support back in) while the RTX 2080 from EVGA is on sale for $800($50 off). My total build cost was roughly $1600 after taxes & a new OS. A pair of RTX 2080's are the same price as my whole rig,and God forbid I were to go with anything faster in sli. lol, Maybe a pair of Titans in sli @ $2490 a piece? Unless people are using the comp for work also or can easily afford it,I don't see sli making gains. My non super 2070 meets the rec specs for the game releases in 2020 that I'm interested in,it even meets half life alyx requirements.:thumbs_thumbsupup:

 

[ENTERPRISE]

Cost is a massive factor. SLI used to be good if you wanted to add in another card later down the line for a performance boost, obviously by then the second GPU will have come down in price. However still a moot point until SLI comes out of a niche it's stuck in.

[J7SC_Orion]

A bit more on mGPU CFR, AFR/2...

 

 

 

 

- Rumour has it (rumour = add salt shaker!) that NVidia, very quietly, added CFR for developers working on 'Hopper' (architecture after Ampere). With GPU dies still being huge even after node shrink to 7nm, mGPU is expected to gain some wind in its sails as vendors *might* have to focus on multi-die single socket GPUs (a la AMD Ryzen CPU pattern)

 

- Below are 2 x 2 benches (Valley, Neon Noir) comparing CFR and AFR NVLink modes. For now, just at 1080p, no GPU OC etc which is also held back a bit by the TR 2950X at that resolution (unlike 4K, but that monitor is in use on another test-bench). Then again, the same handicap applied to all tests, and other factors such as ambient temps were the same. Still, I look forward to add full-OC, 4K results.

 

- CFR seems to work in Valley, Heaven, Superposition and Cryengine Neon Noir..and work very well. I noticed not only slightly higher scores but also a bit lower frame times

 

Valley 1080 AFR

 

 

 

Valley 1080 CFR

 

 

 

Neon Noire 1080 (ultra) AFR2 (better than AFR for this bench)

 

 

Neon Noire 1080 (ultra) CFR

 

 

 

 

...and for 1080p AFR 'all stock' vs 4k full-OC AFR (incl. GPUs + CPU) with the same quality settings (but slightly older driver)...

 

 

 

 

 

 

Edited May 18, 2020 by J7SC_Orion

[J7SC_Orion]

Well, some select 4K results now for Crytek's Neon Noir...

 

First, in case you're wondering why NVlink_SLI 'CFR' is also called checkerboard ...not everything works as it should if you 'force' CFR...here is some Firefox browser fun

 

 

 

...but Crytek's Neon Noir bench works very well with CFR ! Not only better scores but apparently also lower and more consistent frame times.

 

4K results. Please note that there was no overclock (GPU or VRAM) on the GPUs in order to keep it all compatible between runs and also between resolutions. Typically, full-tilt OC with this bench and setup adds about 10% -15% to the scores. Also, this might still be the older Crytek Neon Noir engine (I thought I updated it, but every time I log in, it wants to update again - still, the same engine was used for all the runs below).

 

Settings: 4K, Ultra, Fullscreen, stock clocks

 

7073 - Forcing CFR -

6326 - Forcing AFR2 -

4282 - Forcing AFR - (AFR is default dual card setting)

4733 - Single/Auto-select GPU -

 

Edited May 24, 2020 by J7SC_Orion

[ENTERPRISE]

Well, some select 4K results now for Crytek's Neon Noir...

 

First, in case you're wondering why NVlink_SLI 'CFR' is also called checkerboard ...not everything works as it should if you 'force' CFR...here is some Firefox browser fun

 

 

 

...but Crytek's Neon Noir bench works very well with CFR ! Not only better scores but apparently also lower and more consistent frame times.

 

4K results. Please note that there was no overclock (GPU or VRAM) on the GPUs in order to keep it all compatible between runs and also between resolutions. Typically, full-tilt OC with this bench and setup adds about 10% -15% to the scores. Also, this might still be the older Crytek Neon Noir engine (I thought I updated it, but every time I log in, it wants to update again - still, the same engine was used for all the runs below).

 

Settings: 4K, Ultra, Fullscreen, stock clocks

 

7073 - Forcing CFR -

6326 - Forcing AFR2 -

4282 - Forcing AFR - (AFR is default dual card setting)

4733 - Single/Auto-select GPU -

 

 

Looks like CFR will be the future, will be interesting if Nvidia keep working on it or if they more publicly announce the enhancements.

[J7SC_Orion]

...for now they'll probably keep it more in the shadows, with support still spotty depending on app...but per above posts, this is probably more for SW developers for 'some future-gen GPU architecture (such as Hopper)' when mGPUs / multi-chiplet designs come out. Still, some apps work great already - and yes, "it can play Crysis"

 

 

EDIT: Not so much about CFR vs AFR/2, but about SLI/NVLink in general vs a single card...:

Here's an interesting YouTube vid (by BFG / Benchmarks For Gamers) about 2x 2080 Ti in some modern game titles. Looking at 1% low FPS, when gaming with 4K especially (rather than 1080p), the results are pretty impressive

 

Edited May 25, 2020 by J7SC_Orion

[ENTERPRISE]

SLI certainly still has a place as per the video, I just wish it was as bigger place. Still going single GPU this coming gen as the ROI on a second card is really poor still ....oh but while it can run Crysis, will it be able to run the new remaster of Crysis !?

[J7SC_Orion]

SLI certainly still has a place as per the video, I just wish it was as bigger place. Still going single GPU this coming gen as the ROI on a second card is really poor still ....oh but while it can run Crysis, will it be able to run the new remaster of Crysis !?

 

 

...yeah, definitely still very much a niche market for SLI/NVLink, as is 4k (according to Steam user data). Perhaps with Hopper onward, it *might* change for the aforementioned reasons.

 

As to running remastered Crysis, possibly, if it uses the Cryengine updates in Neon Noir

[ENTERPRISE]

 

 

...yeah, definitely still very much a niche market for SLI/NVLink, as is 4k (according to Steam user data). Perhaps with Hopper onward, it *might* change for the aforementioned reasons.

 

As to running remastered Crysis, possibly, if it uses the Cryengine updates in Neon Noir

 

Yeah I was only messing with the Crysis thing. With the advancement in game engine technology you can get games with the fidelity of Crysis without the performance hit that once was, plus the efficiency of drivers and API's have come such a long way. That is not to say Crytek cannot make something that will give our machines a good run for their money. Hopper may spur on SLI/Nvlink but as per our prior conversation, Nvidia really will need to take the work away from the developer as I do not see anything changing, fingers crossed.

[J7SC_Orion]

 

Yeah I was only messing with the Crysis thing. With the advancement in game engine technology you can get games with the fidelity of Crysis without the performance hit that once was, plus the efficiency of drivers and API's have come such a long way. That is not to say Crytek cannot make something that will give our machines a good run for their money. Hopper may spur on SLI/Nvlink but as per our prior conversation, Nvidia really will need to take the work away from the developer as I do not see anything changing, fingers crossed.

 

...exactly, IF (!) Hopper turns out to be a chiplet/mGPU based design, then NVidia will do the development in-house, a la DLSS 2.0 (3.0)

 

 

[J7SC_Orion]

 

A quick update / November 2020...'for your viewing pleasure', the Microsoft Flight Simulator 2020 thread focuses on NVlink / SLI / CFR results >  here

 

Also, per earlier posts in this thread, some stirrings in the multi GPU / single card field...Intel Xe / server apps (for now)...

  

   

 

 

[ENTERPRISE]

So what are the stirrings specifically ?

[J7SC_Orion]

1 hour ago, ENTERPRISE said:

So what are the stirrings specifically ?

   

  

...it's the first commercial release (rather than rumour), albeit for the lucrative enterprise market...specifically for streaming. Each card has 4 Intel Xe Lp GPU 'tiles' but with 8 GB VRAM each (32 GB total) and an interesting interconnect*. Apparently, 4-way card configurations (= 16 Xe Lp) are possible. Ecosystem and driver development (including for Linux) seem thorough.

 

This is not your home gamer competitor to 6900 Xt and/or 3090 yet (which is currently known as Intel Xe HPG), nevertheless a step forward for mGPU, IMO. There is talk of an Intel discrete 4-tile 'gamer' card competitor (Xe HPG) in the works, but when / how much, who knows.

 

Still, the first commercial mGPU products are hitting the enterprise market, and that seems like a necessary step on the way for HPG consumer variants, not least as it seems to suggest that the Embedded Multi-die Interconnect Bridge (EMIB) high density interconnect is working - the tech that makes it appear as a single GPU and avoid traditional SLI / Crossfire issues. 

  

[UltraMega]

I've done sli a few times, a long time ago. I think back then sli support was at it's peak and even then the results were disappointing far too often. It was amazing when it worked well but even at its peak that was far from guaranteed. I think it was a smart move for Nvidia to stop marketing it to gamers, even if it's still technically an option in some games, simply because it doesn't work well enough to be a selling point especially with gpu prices being so much higher now compared to when sli was more common. 

[J7SC_Orion]

On 11/16/2020 at 8:43 AM, UltraMega said:

I've done sli a few times, a long time ago. I think back then sli support was at it's peak and even then the results were disappointing far too often. It was amazing when it worked well but even at its peak that was far from guaranteed. I think it was a smart move for Nvidia to stop marketing it to gamers, even if it's still technically an option in some games, simply because it doesn't work well enough to be a selling point especially with gpu prices being so much higher now compared to when sli was more common. 

 

...yeah, I know those frustrations (I built at least five past quad-SLI / quad-fire system and lost count of dual and triple GPU setups). But this is about post-SLI/Crossfire (or what is traditionally known as SLI, such as AFR/2) and looking at upcoming mGPU technologies to 'replace' such traditional SLI with tech that has multiple GPU 'chiplets' or 'tiles' on one card that can 'appear as' a single GPU

 

For now, I am having a grand old time via 2x 2080 Ti in CFR in Microsoft Flight Simulator 2020 (EHW thread on that > here ) - got CFR working with that shortly after Flight Sim 2020 release...4K Ultra  

 

 

 

 

 

 

Edited January 3, 2021 by J7SC_Orion

[ENTERPRISE]

20 hours ago, J7SC_Orion said:

   

  

...it's the first commercial release (rather than rumour), albeit for the lucrative enterprise market...specifically for streaming. Each card has 4 Intel Xe Lp GPU 'tiles' but with 8 GB VRAM each (32 GB total) and an interesting interconnect*. Apparently, 4-way card configurations (= 16 Xe Lp) are possible. Ecosystem and driver development (including for Linux) seem thorough.

 

This is not your home gamer competitor to 6900 Xt and/or 3090 yet (which is currently known as Intel Xe HPG), nevertheless a step forward for mGPU, IMO. There is talk of an Intel discrete 4-tile 'gamer' card competitor (Xe HPG) in the works, but when / how much, who knows.

 

Still, the first commercial mGPU products are hitting the enterprise market, and that seems like a necessary step on the way for HPG consumer variants, not least as it seems to suggest that the Embedded Multi-die Interconnect Bridge (EMIB) high density interconnect is working - the tech that makes it appear as a single GPU and avoid traditional SLI / Crossfire issues. 

  

Right thanks for the the explanation bud. Well as it is hitting enterprise market, there is a good chance that it will trickle down to the consumer market. However, again assuming that game developers are kept out of the equation when it comes to the software support and its inherently done via drivers then it could work out well. Lets hope this does not stay in the enterprise/propriety's cupboard.

[J7SC_Orion]

I recently got Cyberpunk 2077, and initial tests with 2x 2080 Ti / CFR were not encouraging. I then went back to Microsoft Flight Simulator 2020 (CFR) w/ an updated driver...forgot about that update and loaded Cyberpunk 2077 (4K native) - it actually worked for a bit in CFR, but then I tried to change graphics settings and boom, crash   Still, this was an encouraging experience as I have been using CFR since last spring on other apps...hopefully, I can figure Cyberpunk 2077 CFR out. I can play that game w/just one (high-clocked) card at 4K / RTX max / (DLSS quality unless combat, then <> DLSS balanced). But it would be so much nicer to have CFR... even better if CD PROJEKT RED would do the CFR profile for us !

[J7SC_Orion]

As @ENTERPRISE already posted here , AMD has also confirmed it is working on multi-chiplet- GPUs via a recent patent filing...looks interesting as a serial+semi parallel approach. Here's another source / interesting write-up about it:

  

 

 

 

[ENTERPRISE]

Actually looking forward to how a Ryzen style GPU shapes up !