Nvidia putting the squeeze on AIB partners

[Stingy McDuck]

I'm willing to invest $500 on a GTX 2070. What do you think about getting a $450~ish for a GTX 1080?

 

[meric]

I'm willing to invest $500 on a GTX 2070. What do you think about getting a $450~ish for a GTX 1080?

I'd say let's first see the performance of the new series

 

[Tom01]

AIB partners should start making SLI on board or X2 version of every last-gen GPU to deplete the inventory quickly.

There are nearly no Games, which are supporting Crossfire or SLI.

 

[Tom01]

I'm willing to invest $500 on a GTX 2070. What do you think about getting a $450~ish for a GTX 1080?

I'm willing to invest $500 on a GTX 2070. What do you think about getting a $450~ish for a GTX 1080?

Get a 1080 Ti, when Prices are dropping.

 

[enemys]

If they had enough bandwidth and low latency on a custom design then they would be fine sharing memory. Nvidia themselves are rolling out Nvlink in the new higher end RTX series, which if it works the same as the Quadro cards you'll be able to have the GPUs share memory as one pool. SLI was always severely bandwidth limited in the connection between cards but this would change that.

No, they wouldn't. Sharing memory means synchronizing access to it and implies a different computation/rendering pipeline. Software that was designed with shared memory in mind, for example "professional" rendering/computation programs that are aware of the option to share memory, can make use of that. DX12 can make use of that, but it needs to be baked in to every title separately by its developers. It can't happen automatically. And even when shared memory pool is supported, it isn't simple "all memory is equally available" - only GPU's own memory is fast and always available, other chips' memory is only accessable through a much slower interconnector. Even in case of NVLink it only gives like 10% of ~500GB/s available to RTX 20xx GPUs.

On top of that, if both chips were to be connected to the same memory chips, they would also need new, low level hardware that is not trivial to design and implement. And if they were to be directly connected only to half of the memory, and go through some interconnector to the other half, it would also require a memory controller redesign and would essentialy halve memory and bandwidth in games/programs that do not support shared memory pool and increase latency when accessing the other half.

 

[Puiu]

You must be mistaken!! Ask some of our loyal AMD posters - they’ll tell you that all AMD video cards are future proof... they might not be as good as their Nvidia equivalents at launch, but in a few years, they’ll be able to give be Titan performance!

Haha, nice joke. Like it or not, during the Maxwell era (and earlier) that was completely true. Maxwell still doesn't have the promised DX12 features like async compute that Nvidia said that they'll make available with a driver update. They did do a driver update that helped with DX12, for Pascal though.

Nobody has ever told anyone that buying Vega is going to future proof anything beyond maybe having access to variable refresh rates outside of expensive high end gaming monitors.

 

[Vulcanproject]

No, they wouldn't. Sharing memory means synchronizing access to it and implies a different computation/rendering pipeline. Software that was designed with shared memory in mind, for example "professional" rendering/computation programs that are aware of the option to share memory, can make use of that. DX12 can make use of that, but it needs to be baked in to every title separately by its developers. It can't happen automatically. And even when shared memory pool is supported, it isn't simple "all memory is equally available" - only GPU's own memory is fast and always available, other chips' memory is only accessable through a much slower interconnector. Even in case of NVLink it only gives like 10% of ~500GB/s available to RTX 20xx GPUs.

On top of that, if both chips were to be connected to the same memory chips, they would also need new, low level hardware that is not trivial to design and implement. And if they were to be directly connected only to half of the memory, and go through some interconnector to the other half, it would also require a memory controller redesign and would essentialy halve memory and bandwidth in games/programs that do not support shared memory pool and increase latency when accessing the other half.

You cited the obvious problems, I only said it was possible with a custom design, and it is. You also said that it is not trivial to design and implement, it isn't, but one company has already taken the biggest steps towards it.

Nvlink obviously still does not have enough bandwidth or good enough latency for one GPU core to fully access the performance of all the memory on another card. It was an example of sharing memory however I am aware it is still an external link with limitations. What is required in a modern custom design with a twin GPU PCB is something more along the lines of AMD's Infinity fabric.

In fact the near future of GPU design is almost certainly this very type of interconnect, because the hard buffers of silicon lithography are fast approaching. AMD have made no secret that they are looking very seriously at this for their road maps.

It's not going to happen next year or anything. However once you are scraping the bottom of the barrel beyond 5nm in another 3-4 years then this will probably be the first port of call to offer a next generation.

 

[pcnthuziast]

Nivida has always been pretty shady in their business practices imho.

 

[Boilerhog146]

You mean sli on Board as in built in sli with the mobo?

I meant cards like GTX 295, 590, and 690 but made with non-flag ship chips. AMD had something like 4850 X2 as well as 4870 X2 back in the day.

To be honest this makes the most sense, both for moving inventory, and keeping cost down for board partners in terms of cost per chip on the market. Since they can basically halve the number of memory chips they need per chip (SLI and Xfire cards don't pool memory iirc, they just share the 'master' chip's mem), that means the still relatively high cost of graphics memory would hurt quite a bit less per board. I really don't understand why these cards stopped being produced in the recent series.

No, dual cards need dual memory - each chip needs it's own memory budget.

You mean sli on Board as in built in sli with the mobo?

I meant cards like GTX 295, 590, and 690 but made with non-flag ship chips. AMD had something like 4850 X2 as well as 4870 X2 back in the day.

To be honest this makes the most sense, both for moving inventory, and keeping cost down for board partners in terms of cost per chip on the market. Since they can basically halve the number of memory chips they need per chip (SLI and Xfire cards don't pool memory iirc, they just share the 'master' chip's mem), that means the still relatively high cost of graphics memory would hurt quite a bit less per board. I really don't understand why these cards stopped being produced in the recent series.

No, dual cards need dual memory - each chip needs it's own memory budget.

with sli they need their own memory buffer.but isn't nvlink different ? using what ever memory that's available.in tandem . so 2 or 3 gpu should be able to use the one bank of memory?

 

[Reehahs]

There are nearly no Games, which are supporting Crossfire or SLI.

That is true since the developers noticed that multi-GPU users were in single digit percentages. However, the default profiles do a pretty good job of GPU utilisation moving forward. Previously you had games that without multi-GPU profile the performance tanked worse than it would on a single GPU,