• We’re currently investigating an issue related to the forum theme and styling that is impacting page layout and visual formatting. The problem has been identified, and we are actively working on a resolution. There is no impact to user data or functionality, this is strictly a front-end display issue. We’ll post an update once the fix has been deployed. Thanks for your patience while we get this sorted.

[VideoCardz]NVIDIA GP104 and first Polaris GPUs supposedly spotted on Zauba

Page 2 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
Silverforce11 said:
I don't have the link on me since its awhile back, but the company (UMT?) who did the interposer for Fury did a presentation to talk about the tech.

It is simply a matrix of interconnects that allow chips to talk to each other, chip <-> chip or HBM.
Click to expand...

I realize that, but from what I understand syncing all those cores isn't a feasible task.
 
Techhog said:
Just ignore it. This dumb idea is going to come up every single year going forward, no matter how flawed it is.
Click to expand...

Hahah zing.

Although if AMD has no true high end chip coming out in 2016, I think we'll see two mid-range chips put onto 1 PCB a la the HD 3870x2, with or without fancy interposer magic. I think it's actually a solid strategy for AMD to go back to using two mid-die chips on a single PCB to compete with higher end Nvidia chips instead of trying to match Nvidia with large monolithic dies out of the gate. Their R&D simply doesn't have the funding that Nvidia has right now. Better to put out two really good small dies and save up time and money for a really good big die later on, instead of botching yet another big die release.
 
Last edited:
Arachnotronic said:
The chatter that I've been hearing is that we shouldn't expect flagship AMD Polaris-based GPUs until late 2016/early 2017, but take it for what it is -- just chatter.
Click to expand...

I heard we should expect summer launch with the flagship coming out about the start of the new school year.

tviceman said:
Hahah zing.

Although if AMD has no true high end chip coming out in 2016, I think we'll see two mid-range chips put onto 1 PCB a la the HD 3870x2, with or without fancy interposer magic. I think it's actually a solid strategy for AMD to go back to using two mid-die chips on a single PCB to compete with higher end Nvidia chips instead of trying to match Nvidia with large monolithic dies out of the gate. Their R&D simply doesn't have the funding that Nvidia has right now. Better to put out two really good small dies and save up time and money for a really good big die later on, instead of botching yet another big die release.
Click to expand...
Wasn't the 290x the best gpu of the last generation? The furyx isn't even that far behind at 1440p+, and it was their first crack at hbm.
 
Last edited:
I've lost track how many times people have predicted AMD was incapable to make a "big die" GPU, let's just say the predictions were not exactly spot on.
 
tviceman said:
Hahah zing.

Although if AMD has no true high end chip coming out in 2016, I think we'll see two mid-range chips put onto 1 PCB a la the HD 3870x2, with or without fancy interposer magic. I think it's actually a solid strategy for AMD to go back to using two mid-die chips on a single PCB to compete with higher end Nvidia chips instead of trying to match Nvidia with large monolithic dies out of the gate. Their R&D simply doesn't have the funding that Nvidia has right now. Better to put out two really good small dies and save up time and money for a really good big die later on, instead of botching yet another big die release.
Click to expand...

AMD just delayed its Fiji dual card largely as a result of the terrible cfx/sli support landscape. It would be a really bad idea for them to rely on dual chip cards going forward. I think part of AMD's reasoning for even creating the monolithic Fiji die and the abandonment of the small die strategy over the course of the last several generations has been due to the realization that dual chip cards are becoming less and less viable options.
 
If the price point of the finfet gpus corresponds to performance of the current ones than the 1st chip should have around R9 390X performance and the 3rd one around 370-380 performance - I expect this was the gpu demonstrated recently.
 
bradsh said:
Hmmm... do we really think AMD will ship a high end part mid 2016? Should I be holding off on buying a 980ti at this point if I'm mostly looking to power VR and fall games?
Click to expand...

Fall this or next year? VR is still young, no experience available. I would wait for sure. As a stop gap buying used from ebay for cheap is often a good value proposition.

prtskg said:
If the price point of the finfet gpus corresponds to performance of the current ones than the 1st chip should have around R9 390X performance and the 3rd one around 370-380 performance - I expect this was the gpu demonstrated recently.
Click to expand...

Possibly. But will it be cheaper? Will performance/$ actually increase? 290x could be had a year ago for $300 easily. Then they raised price for 390x with no really performance gain. So it should cost $300 max. Doubt that will happen.
 
Silverforce11 said:
Because they already have some experience with the tech, interposer-linked multi-chips can behave like one big chip without the pitfalls of CF/SLI. So that's one way to get a high end SKU...
Click to expand...
So, do you think AMD will revive sideport spoted at RV770?
 
Silverforce11 said:
I don't have the link on me since its awhile back, but the company (UMT?) who did the interposer for Fury did a presentation to talk about the tech.

It is simply a matrix of interconnects that allow chips to talk to each other, chip <-> chip or HBM.
Click to expand...

Its not going to behave like 1 chip in that scenario.
 
beginner99 said:
Possibly. But will it be cheaper? Will performance/$ actually increase? 290x could be had a year ago for $300 easily. Then they raised price for 390x with no really performance gain. So it should cost $300 max. Doubt that will happen.
Click to expand...

That's a good question. But remember on a 290X/390X, the GPU cost is only a part of it. Even if it cost twice the cost to make the GPU itself for 50% or more performance, the card would "only" go up maybe 25% in cost.
 
DeathReborn said:
It would require a central hub with surrounding "compute" chips and significant boost in current technology to function at even current gen speeds.
Click to expand...
Do you mean amps?
What do you consider current gen speed?
Can you explain with some more detail?

The reason I'm asking was that I came to the same conclusion of a central shared hub having everything except the CU block, with the shaders, rops, rasterizer and geometry processor in another module. This latter module would be replicated to achieve the desired result.

Going by the Polaris slides for IP block names.
 
Last edited:
An interposer based very wide interface wouldn't be much different from what Nvidia is planning with their NVLink. Combine that with a shared memory space and it might reduce the amount of "awareness" games need to fully utilize multi-GPU. I don't see it being used for gaming GPUs until at least 2nd gen 14/16nm or even farther out, the technology will be tested and refined in HPC first.
 
Last edited:
ShintaiDK said:
Nvlink is just a faster PCIe. Serves no other purpose.
Click to expand...

Resource sharing puts stress on both bandwidth and latency. NVlink or an interposer based connection should be a big improvement over PCIe in both categories, directly sharing memory becomes more practical. In fact it's a core component in enabling unified memory, whether GPU-GPU or CPU-GPU. I don't think gaming GPUs will use it for a while yet considering it is just now getting close to being deployed in the very high end of HPC.
 
Last edited:
Techhog said:
Is there meant to be a joke here?
Click to expand...

"Because" is not an answer.

Vesku said:
An interposer based very wide interface wouldn't be much different from what Nvidia is planning with their NVLink. Combine that with a shared memory space and it might reduce the amount of "awareness" games need to fully utilize multi-GPU. I don't see it being used for gaming GPUs until at least 2nd gen 14/16nm or even farther out, the technology will be tested and refined in HPC first.
Click to expand...

HPC doesn't use crossfire. Each GPU can be given it's own work independent of the other GPU's.
 
You must log in or register to reply here.

TRENDING THREADS

Back
Top