"Multi-Core" graphics cards

[xFluing]

So, now that we are getting closer and closer to the end of the manufacturing process shrink, how will GPUs evolve from here on out? The only solution with current technology would be bigger dies, sure, but that will just drive up the costs because the bigger the dies, the less of them fit a wafer, and implicitly drive the costs higher and higher.

 

Unless we somehow find a way for multiple GPUs to work together like Ryzen CCXes, I'd say the GPU speed will just hit a plateau and stay there until GPUs become powered by quantum processors or some shit.

[bleedblue]

Multiple GPUs to work together? You mean SLi/Crossfire? That was a thing way back when, Asus Ares and Mars cards IIRC.

[xAcid9]

There was some talk about MCM(multi core modules) GPU from AMD side for Navi last year but recent rumor suspect Navi probably not base on MCM. 

[xFluing]

2 minutes ago, bleedblue said:

Multiple GPUs to work together? You mean SLi/Crossfire? That was a thing way back when, Asus Ares and Mars cards IIRC.

Nah, I'm talking about some kind of tech that's not shit, something that would make the GPU modules work as one singular GPU unit instead of taking turns rendering frames or each one rendering half a frame.

[Ross Siggers]

AMD already did this, admittedly a long time ago. The HD4870 X2 had two chips on a single pcb, I can't remember if they've done it since though.

*EDIT* The R295 X2 was another one, apparently. 

[xFluing]

Just now, Ross Siggers said:

AMD already did this, admittedly a long time ago. The HD4870X2 had two chips on a single pcb, I can't remember if they've done it since though.

295x2 also had two GPUs but they were running in crossfire, I'm talking about something that would have say 4 GPU modules, and each of those would act as a CPU core or something and working together as one big GPU unit made of many "cores", removing the issues of crossfire / sli.

[Tibon]

1 minute ago, xFluing said:

Nah, I'm talking about some kind of tech that's not shit, something that would make the GPU modules work as one singular GPU unit instead of taking turns rendering frames or each one rendering half a frame.

This is sadly still a dream in consumer grade gpu atm. I suspect the only chance they could achieve this (seamless gpu switching, hotswap, load balancing) is having a single standard in gpu development.

[GoldenLag]

9 minutes ago, xFluing said:

Nah, I'm talking about some kind of tech that's not shit, something that would make the GPU modules work as one singular GPU unit instead of taking turns rendering frames or each one rendering half a frame.

AMD is working on something like that for datacenters. For regular gaming and similar workload that involves drawing frames is very hard to achieve. So far AMD has stated they CAN'T do a a consumer MCM design based on current designs is impossible and making a design able to do so is something they havent been able to do so far. 

 

They cant apply infinity fabric tech to GPUs in other words.

 

Something they could possibly do is a modular GPU core and build accordingly.

 

Interestingly Intel is looking at modular die tech atm.  They are currently looking at it CPU side, though it is probably applicable on GPUs aswell

 

 

Edit: did a awful spelling misstake. Please re read

Edited September 12, 2018 by GoldenLag

[GoldenLag]

Updated my last post as it contained a spelling misstake. Refresh page to see the change

[Dabombinable]

Its possible that once again we could see multiple GPU on graphics cards. Not only 2 (which has been primarily AMD's thing for the lat 10+ years), but 4 ore more:

http://www.x86-secret.com/articles/divers/v5-6000/v56kgb-6.htm

[xFluing]

2 minutes ago, GoldenLag said:

AMD is working on something like that for datacenters. For regular gaming and similar workload that involves drawing frames is very hard to achieve. So far AMD has stated they can do a a consumer MCM design based on current designs is impossible and making a design able to do so is something they havent been able to do so far. 

 

They cant apply infinity fabric tech to GPUs in other words.

 

Something they could possibly do is a modular GPU core and build accordingly.

 

Interestingly Intel is looking at modular die tech atm.  They are currently looking at it CPU side, though it is probably applicable on GPUs aswell

Well Intel got the idea from AMD's CCXes but it's different enough, they're looking to make the cores on 7nm and the rest, less important stuff like memory controller on 14nm or 28nm which I have to say is pretty smart because those two processes are already mature enough for high yields and low prices, and because the cores will be made individually it'll allow for much smaller dies once again reducing the cost.

[xAcid9]

12 minutes ago, Ross Siggers said:

AMD already did this, admittedly a long time ago. The HD4870 X2 had two chips on a single pcb, I can't remember if they've done it since though.

*EDIT* The R295 X2 was another one, apparently. 

I think he meant something similar to Zen CPUs with multiple CCXs in a single die.

[xFluing]

1 minute ago, xAcid9 said:

I think he meant something similar to Zen CPUs with multiple CCXs in a single die.

Yeah

[GoldenLag]

2 minutes ago, xFluing said:

Well Intel got the idea from AMD's CCXes but it's different enough, they're looking to make the cores on 7nm and the rest, less important stuff like memory controller on 14nm or 28nm which I have to say is pretty smart because those two processes are already mature enough for high yields and low prices, and because the cores will be made individually it'll allow for much smaller dies once again reducing the cost.

The issue lies in connecting 2 dies that can function just fine on their own. Which is what the zen approach is. Due to differences im memmory distrubution and cache such an approach is not possible. 

 

They would have to make "CU" dies to be fed by a "cache" die that will feed the different "CU" dies. Something that intel can apply with their different node tech. 

 

This approach would also probably involve an active substrate. Something AMD is currently working on

[Cereal5]

Did someone say... RTX 2090?

 

edit: I saw someone earlier mentioning the HD4870 X2 and the R9 295 x2. There's also the GTX 590 and 690 which are NVIDIA examples of PCBs with two dies

[GoldenLag]

2 minutes ago, Cereal5 said:

Did someone say... RTX 2090?

 

edit: I saw someone earlier mentioning the HD4870 X2 and the R9 295 x2. There's also the GTX 590 and 690 which are NVIDIA examples of PCBs with two dies

Those cards are just 2 GPUs sharing 16PCIe lanes through a PLX chips abd bonded using either crossfire or a on board SLI bridge

[Cereal5]

Just now, GoldenLag said:

Those cards are just 2 GPUs sharing 16PCIe lanes through a PLX chips abd bonded using either crossfire or a on board SLI bridge

Still 2 GPUs. 8x never slowed any other GPU down that I've heard of

[xFluing]

9 minutes ago, Cereal5 said:

Still 2 GPUs. 8x never slowed any other GPU down that I've heard of

But they work through crossfire / sli which has its faults and needs to be supported by the game devs, what I mean is something to circumvent that.

[Cereal5]

28 minutes ago, xFluing said:

295x2 also had two GPUs but they were running in crossfire, I'm talking about something that would have say 4 GPU modules, and each of those would act as a CPU core or something and working together as one big GPU unit made of many "cores", removing the issues of crossfire / sli.

The issue I see with this is that in each die, just like a CPU, a GPU has many cores in it. It's not just one big core, there's hundreds or thousands of cores. It would be like using a server mobo that has multiple CPU sockets. I'm not totally sure how those types of boards work, but a GPU would have to do that all in about 1/10th of the amount of PCB. I think it would be easier to just print a bigger die and charge more. However, if the Titan V is anything to look to, it seems like the technology is around to still stuff way way more cores into a die. No matter what happens, the price will increase. Whether or not the technology does. Inflation's a bitch, lol

[GoldenLag]

10 minutes ago, Cereal5 said:

Still 2 GPUs. 8x never slowed any other GPU down that I've heard of

Scaling is bad and only there if the developer codes it in. They still operate as 2 different GPUs. 

[xFluing]

Just now, Cereal5 said:

The issue I see with this is that in each die, just like a CPU, a GPU has many cores in it. It's not just one big core, there's hundreds or thousands of cores. It would be like using a server mobo that has multiple CPU sockets. I'm not totally sure how those types of boards work, but a GPU would have to do that all in about 1/10th of the amount of PCB. I think it would be easier to just print a bigger die and charge more. However, if the Titan V is anything to look to, it seems like the technology is around to still stuff way way more cores into a die. No matter what happens, the price will increase. Whether or not the technology does. Inflation's a bitch, lol

Inflation has nothing to do with the costs I'm talking about. Wafers cost a lot so fitting as many dies onto it as possible greatly increases the yield and implicitly reduces the cost.

 

The RTX 2080 price does not follow any inflation curve either, so still no inflation.

[GoldenLag]

1 minute ago, Cereal5 said:

-snip-

No GPUs have cores. They have execution units. Calling them "cores" is nice but it is missleading.

 

The only GPU with cores is "project larebee". A intel GPU using atom cores and 4threaded cores. We see them as Xeon Phi today

[Cereal5]

9 minutes ago, xFluing said:

Inflation has nothing to do with the costs I'm talking about. Wafers cost a lot so fitting as many dies onto it as possible greatly increases the yield and implicitly reduces the cost.

 

The RTX 2080 price does not follow any inflation curve either, so still no inflation.

It reduces the price per die, due to not using as many wafers, but it increases the price of the wafers, due to having more dies.

[xAcid9]

3 minutes ago, Cereal5 said:

It reduces the price per die, due to not using as many wafers, but it increases the price of the wafers, due to having more dies.

Sorry what? That's not how this work i think. 

[xFluing]

3 minutes ago, Cereal5 said:

It reduces the price per die, due to not using as many wafers, but it increases the price of the wafers, due to having more dies.

Cost is overall lower, otherwise Ryzen prices would have been through the roof.