eduonkhl
Superclocked Member
- Total Posts : 105
- Reward points : 0
- Joined: 2021/05/08 09:09:54
- Status: offline
- Ribbons : 0
I was wondering for a while if Multi-GPU systems may be on a comeback since chiplets rely on a hardware based sollution to split up the work between the chiplets and if latency still would be an issue. Many report that NVLINK, while much better than the "original" SLI, still doesn't feel as smooth and 1% and 0.1% lows tend to confirm that. You think we could finally have multiple chips be working so well in unified fashion that it's indistinguishable from a single GPU setup? You think precious silicon space could be used solely for metadata processing? I hope I'm correct with my understanding of all this, I'm no microchip engineer after all, keep the conversation civilised, have a nice day girls and guys :)
|
XrayMan
Insert Custom Title Here
- Total Posts : 73000
- Reward points : 0
- Joined: 2006/12/14 22:10:06
- Location: Santa Clarita, Ca.
- Status: offline
- Ribbons : 115
Re: Multi-GPU Comeback with Chiplets? Discuss your 2 cents
2021/05/13 21:36:49
(permalink)
My Affiliate Code: 8WEQVXMCJL Associate Code: VHKH33QN4W77V6A   
|
FloodControl
Superclocked Member
- Total Posts : 137
- Reward points : 0
- Joined: 2016/02/11 17:28:32
- Status: offline
- Ribbons : 1
Re: Multi-GPU Comeback with Chiplets? Discuss your 2 cents
2021/05/14 12:07:05
(permalink)
I don't think multi-chip GPUs are feasible for real-time rendering like games. The main bottleneck isn't bandwidth, but rather latency. SLI profiles in the past tried to get around this by dividing things up in neater packages for inter-GPU communication or by resorting to alternate frame rendering. To get fully-transparent multi-GPU support, we need the host system to be able to access all memory resources at the same speed. That means either fully implementing DX12 features (more strain on the developer side) or having fully coherent communication between GPUs and the rest of the system built directly into the communication protocol (NVLink CPU-to-GPU). Once that problem is solved, then we have to deal with the thermal dynamics of having two high-power chips packed into a small area. Dissipating 500W is hard enough when you have two chips on one board, let alone side-to-side on a single interposer.
|
kaledorm1
New Member
- Total Posts : 10
- Reward points : 0
- Joined: 2021/05/14 21:53:19
- Status: offline
- Ribbons : 0
Re: Multi-GPU Comeback with Chiplets? Discuss your 2 cents
2021/05/14 22:13:22
(permalink)
Not so very long ago, we had the GTX690, which had two GPUs on one card. Needless to say, the idea wasn't very successful.
The big problem is that multi-GPU is a niche that very few developers bother to even consider, nevermind optimize for.
|
Zixinus
Superclocked Member
- Total Posts : 165
- Reward points : 0
- Joined: 2020/10/10 04:18:52
- Status: offline
- Ribbons : 0
Re: Multi-GPU Comeback with Chiplets? Discuss your 2 cents
2021/05/15 14:40:40
(permalink)
I thought the idea of chiplets was to part of larger processor packages, rather than having to make one humongous chip? I don't see how that helps with multi-gpu support, at least with games.
What applications are multi-GPU support useful currently? That are not games.
|
eduonkhl
Superclocked Member
- Total Posts : 105
- Reward points : 0
- Joined: 2021/05/08 09:09:54
- Status: offline
- Ribbons : 0
Re: Multi-GPU Comeback with Chiplets? Discuss your 2 cents
2021/05/15 15:44:15
(permalink)
Guys. I think you got it a bit wrong on what I meant or just said your oppinion on multi gpus without even reading the text let alone understand it, maybe I should of been clearer on what I meant. The gtx 690 had 2 gpus on one board yes but those where not bound together chiplets but using the rather conventional SLI bridge. It was no different than if you'd just bought 2 680s. Only difference is that those are on one board and you'll need less slots and space to accompany 2 gpus. What I meant with multi-gpu is that since chiplets are a HARDWARE based sollution. It's not quiet the same on gpus as on the ryzen cpus. But both show up as a unified unit and so THERE IS NO PROGRAMMING to do to "optimize" it for games. My suggestion just meant that since this sollution can work (atleast in the labs for now) to bound together multiple chips and use them like its a singular unit in software on one single board (like the gtx 680 but with multiple chips showing up as one in software) I was wondering if it's possible to have the same behaviour by connecting multiple boards, like NVLINK but the connection is so good that it can act like it's a single unit in software (no optimization needed in this case). The Idea is since it's possible to do that on one board I thought if it could be possible with multiple ones aswell since the technology is already in place.
|
Zixinus
Superclocked Member
- Total Posts : 165
- Reward points : 0
- Joined: 2020/10/10 04:18:52
- Status: offline
- Ribbons : 0
Re: Multi-GPU Comeback with Chiplets? Discuss your 2 cents
2021/05/16 13:10:33
(permalink)
eduonkhl
Guys. I think you got it a bit wrong on what I meant or just said your oppinion on multi gpus without even reading the text let alone understand it, maybe I should of been clearer on what I meant. The gtx 690 had 2 gpus on one board yes but those where not bound together chiplets but using the rather conventional SLI bridge. It was no different than if you'd just bought 2 680s. Only difference is that those are on one board and you'll need less slots and space to accompany 2 gpus. What I meant with multi-gpu is that since chiplets are a HARDWARE based sollution. It's not quiet the same on gpus as on the ryzen cpus. But both show up as a unified unit and so THERE IS NO PROGRAMMING to do to "optimize" it for games. My suggestion just meant that since this sollution can work (atleast in the labs for now) to bound together multiple chips and use them like its a singular unit in software on one single board (like the gtx 680 but with multiple chips showing up as one in software) I was wondering if it's possible to have the same behaviour by connecting multiple boards, like NVLINK but the connection is so good that it can act like it's a single unit in software (no optimization needed in this case). The Idea is since it's possible to do that on one board I thought if it could be possible with multiple ones aswell since the technology is already in place.
It would make sense to seperate certain functionalities into chiplets that already have their dedicated space on the processor chip. Like what they use for raytracing and make modules for regular processors (FP32 or what were they called?). I think the same could be done as AMD does for its Ryzen processors? However, what I think you are talking about is to make an improved NVLNK to have functionally single GPU out of multip GPUs? I don't think so, for the same reason you can't have multiple CPUs be a single CPU: the communication between processors would be slower than the communication within the processor housing.
|
Smiggie
Superclocked Member
- Total Posts : 109
- Reward points : 0
- Joined: 2021/04/21 06:50:04
- Status: offline
- Ribbons : 0
Re: Multi-GPU Comeback with Chiplets? Discuss your 2 cents
2021/05/16 14:09:29
(permalink)
Had an 780 SLI system in the past, moved on a higher single gpu never wanted to came back
|
hpak53
Superclocked Member
- Total Posts : 115
- Reward points : 0
- Joined: 2010/02/26 22:13:47
- Status: offline
- Ribbons : 0
Re: Multi-GPU Comeback with Chiplets? Discuss your 2 cents
2021/05/16 14:24:40
(permalink)
What would be awesome is if they could somehow develop a slave GPU system. You could buy a 3070 type of GPU and if the need arose, rather than buying a brand new higher tier GPU, you buy a slave card that only does additional graphics tasks.
Not sure how it's feasible with current technology to limit latency but it could add more powerful graphics processing presumably at a lower cost to consumers since you would only need to make the slave card with only the necessary parts to increase graphics processing.
You could even take it one further and include that ability to the CMP cards so they serve multiple needs rather than just mining so those cards can be of some use after they are of no use mining.
|
Sajin
EVGA Forum Moderator
- Total Posts : 49168
- Reward points : 0
- Joined: 2010/06/07 21:11:51
- Location: Texas, USA.
- Status: offline
- Ribbons : 199
Re: Multi-GPU Comeback with Chiplets? Discuss your 2 cents
2021/05/17 00:30:01
(permalink)
Don’t see it coming, but would be awesome if so.
|
FloodControl
Superclocked Member
- Total Posts : 137
- Reward points : 0
- Joined: 2016/02/11 17:28:32
- Status: offline
- Ribbons : 1
Re: Multi-GPU Comeback with Chiplets? Discuss your 2 cents
2021/05/17 05:19:08
(permalink)
eduonkhl
Guys. I think you got it a bit wrong on what I meant or just said your oppinion on multi gpus without even reading the text let alone understand it, maybe I should of been clearer on what I meant. The gtx 690 had 2 gpus on one board yes but those where not bound together chiplets but using the rather conventional SLI bridge. It was no different than if you'd just bought 2 680s. Only difference is that those are on one board and you'll need less slots and space to accompany 2 gpus. What I meant with multi-gpu is that since chiplets are a HARDWARE based sollution. It's not quiet the same on gpus as on the ryzen cpus. But both show up as a unified unit and so THERE IS NO PROGRAMMING to do to "optimize" it for games. My suggestion just meant that since this sollution can work (atleast in the labs for now) to bound together multiple chips and use them like its a singular unit in software on one single board (like the gtx 680 but with multiple chips showing up as one in software) I was wondering if it's possible to have the same behaviour by connecting multiple boards, like NVLINK but the connection is so good that it can act like it's a single unit in software (no optimization needed in this case). The Idea is since it's possible to do that on one board I thought if it could be possible with multiple ones aswell since the technology is already in place.
I touched on this with the issue of latency. Ryzen CPUs work so well because CPU tasks are less sensitive to latency than GPU counterparts; additionally, the chiplets are treated in almost a NUMA fashion where tasks that must communicate between threads are prioritized to an individual chiplet by the OS. If you look at an inter-core latency graph, you'll see a severe (2.5-10x) latency penalty for moving off-chip. Remember that this is all happening through an interposer utilizing Infinity Fabric communication protocol like the theoretical GPU you're proposing. When you have something as latency-sensitive as a GPU, incurring a penalty that can reach into the 80 ms range can really mess with frame delivery. 60 Hz translates to 16.7 ms, so not even touching on high refresh rate monitors you're already talking about almost 5 frames rendered before the whole GPU can respond to a change. My argument is still that we need a solution that can talk to multiple GPUs with full coherency, rather than trying to get the massive parallelized engines to talk to each other faster. It's the difference between telling two people to complete Task A and Task B vs. handing a task list to the both of them and telling them to figure out how to get it all done together.
|
hftvhftv
New Member
- Total Posts : 59
- Reward points : 0
- Joined: 2014/09/19 13:59:29
- Location: Canada
- Status: offline
- Ribbons : 0
Re: Multi-GPU Comeback with Chiplets? Discuss your 2 cents
2021/05/23 10:08:33
(permalink)
I really doubt chiplets will come in the near future, but you never really know.
|
ysesq
Superclocked Member
- Total Posts : 130
- Reward points : 0
- Joined: 2021/03/18 15:48:44
- Status: offline
- Ribbons : 0
Re: Multi-GPU Comeback with Chiplets? Discuss your 2 cents
2021/05/24 13:29:24
(permalink)
amd may do chiplets but doubt its economically sound.
|
excesspickles
New Member
- Total Posts : 1
- Reward points : 0
- Joined: 2020/10/27 17:59:25
- Status: offline
- Ribbons : 0
Re: Multi-GPU Comeback with Chiplets? Discuss your 2 cents
2021/05/24 16:21:25
(permalink)
Chiplets don't address the heat problem, as you can't power all of the silicon in a chip for very long before you run into thermal limits (~200w for cpu/gpu sized chip seems to be the current limit). If you do mount the chiplet far enough away for the heat to not be a problem, as other people mentioned, you run into latency/bandwidth/cost problems.
Chiplets are a good idea for things that are not heat limited, such as mobile phone CPU/GPUs, where many different chips are mounted in the same package to save space, and chiplets let the vendor mix&match GPU/CPUS/etc to match the phones price target without needing new silicon.
|
OhmicFutility
Superclocked Member
- Total Posts : 101
- Reward points : 0
- Joined: 2021/05/24 19:20:34
- Status: offline
- Ribbons : 0
Re: Multi-GPU Comeback with Chiplets? Discuss your 2 cents
2021/05/27 01:53:19
(permalink)
Chiplets are a wonderful solution to the manufacturing yield issues associated with massive monolithic dies. Since yield is a product of defect density, the smaller the chip-let, the number of good dies per wafer increases exponentially, decreasing overall expenses (with some offset due to increased packaging costs). At least it sounds good on paper.
But that is not the discussion here, nor is it the issue currently facing multi-gpu solutions.
I believe the biggest issue for multi-gpu support is software rather than hardware. Even though the technology as existed in some form or another for several generations, the number of modern games that have supported this has dwindled exponentially. And there are numerous reasons why that has happened purely from the perspective of SW advances.
Linus did a video where he explored the performance of two 3090's in SLI. I recommend watching that video.
At least for me, I am reasonably convinced that mutli-gpu solutions are probably dead for the near future, at least from a gaming perspective.
For heavily parallelized workloads other than gaming... I suppose I could still see the application. But that is why NVidia also makes server components.
|
cbutsko
New Member
- Total Posts : 100
- Reward points : 0
- Joined: 2021/04/08 10:37:14
- Status: offline
- Ribbons : 0
Re: Multi-GPU Comeback with Chiplets? Discuss your 2 cents
2021/06/07 19:37:51
(permalink)
Yes, I try to think of it similar to how AMD's processors have different chiplets. The chiplets communicate through some sort of middle-man to act as one
|
leelawrencecs
Superclocked Member
- Total Posts : 104
- Reward points : 0
- Joined: 2020/10/05 20:56:41
- Status: offline
- Ribbons : 0
Re: Multi-GPU Comeback with Chiplets? Discuss your 2 cents
2021/06/08 13:26:16
(permalink)
hopefully apple silicon GPUs can pull this off and they make discrete graphics cards
|
103JL
New Member
- Total Posts : 100
- Reward points : 0
- Joined: 2021/06/14 14:34:49
- Status: offline
- Ribbons : 0
Re: Multi-GPU Comeback with Chiplets? Discuss your 2 cents
2021/06/14 17:17:07
(permalink)
i hope so. in a perfect world, double gpu = double performance. but i guess nvidia won't sell as many GPUs if you did this so they have no incentive to support this
|