AMD Vega 10, Vega 11, Vega 12 and Vega 20 Rumors and Discussion

In fact some games give a counter-intuitive result and HBCC performance gains are weaker at 4K compared to 1080P:
That actually makes sense as the CPU bottleneck would be higher at 1080P and HBCC should reduce management work. If the app thinks all resources are loaded there is nothing to do. Should be similar to testing cards with more memory.

I think HBCC is a good concept and every little bit of performance helps, but I do wonder for real-world gains and practicality whether a PC will then need a minimum of 24GB as there is no such perfect system environment when it comes to
Shouldn't take that much more memory. Just needs devs to adjust how they cache textures and meshes. Storing them in memory and vram(system memory with HBCC) would waste a lot of space. Perhaps RR can solve this as system and GPU memory are the same. Might end up with the CPU reading data from the GPU which could be interesting.

Really just need to load as much of a map or level as possible in advance. Perhaps a low priority DMA copy queue that loads unused memory when the bus is idle. Regardless the tech is cool.

On other news, it looks like Wolfenstien 2 has an option called GPU Culling, recommended for AMD GPUs only, could this be the first appearance of Primitive Shaders in games?
PCGH had RPM on the culling. So probably cluster culling with FP16 data on the vertices. Definitely not geometry or compute holding things back. Curious with that 80% async comment that was made.
 
That actually makes sense as the CPU bottleneck would be higher at 1080P and HBCC should reduce management work. If the app thinks all resources are loaded there is nothing to do. Should be similar to testing cards with more memory.


Shouldn't take that much more memory. Just needs devs to adjust how they cache textures and meshes. Storing them in memory and vram(system memory with HBCC) would waste a lot of space. Perhaps RR can solve this as system and GPU memory are the same. Might end up with the CPU reading data from the GPU which could be interesting.

Really just need to load as much of a map or level as possible in advance. Perhaps a low priority DMA copy queue that loads unused memory when the bus is idle. Regardless the tech is cool.
.
Although ironically HBCC was mentioned origally by Raja in context of higher texture/4K along with better VRAM management in general, yet we have 1080p with greater gains from HBCC relative to 4K in some games but importantly not all and that was my point regarding inconsistencies and not a stable trend; it is not a given one will get 5% or better gains but maybe this will change as it matures albeit possibly still with a limited scope.
Anyway It seems to need 8GB of system RAM to really work well (context work well is 6-10% gains in games).
Like I said that means to be safe one needs 32GB of fast and tight timing DDR4 as no real world consumer gaming PC is an ideal system environment.
 
Last edited:
Although ironically HBCC was mentioned origally by Raja in context of higher texture/4K along with better VRAM management in general, yet we have 1080p with greater gains from HBCC relative to 4K in some games but importantly not all and that was my point regarding inconsistencies and not a stable trend.
Keep in mind 4k texture and 4k resolution are different things. To draw any conclusion the level of detail would need to change and that is somewhat independent of actual screen resolution. Not all games would take resolution into account when adjusting for LoD. Setting textures from Ultra@4K to High@1080P would be the best example of that. Then use a midrange card with only 4GB memory.

Like I said that means to be safe one needs 32GB of fast and tight timing DDR4 as no real world consumer gaming PC is an ideal system environment.
That depends on the VRAM pool though. Using 4k textures at 1080P on a budget machine should show the intended gains. HBCC really is a HPC focused tech, that also happens to provide an interesting opportunity for gaming. For high end gaming the likely outcome would be a notice from your ISP upon installing it.

If there are any tools that can still bias LoD that might provide some interesting testing opportunities.
 
Wolfenstein 2 ist strange. It has GPU culling. What is that? And there is no word about primitive shader.

Why implementing a new culling feature when Vega has primitive shader for culling?

Also strange, the behaviour of this culling feature. In the Computerbase test the worste case szene which have wide view, lot of buildings over a big area with a lot of polygons, Vega loose 25% performance also when this feature is on. Where Vega should benefits most.
 
Last edited:
Wolfenstein 2 ist strange. It has GPU culling. What is that? And there is no word about primitive shader.

Why implementing a new culling feature when Vega has primitive shader for culling?

Also strange, the behaviour of this culling feature. In the Computerbase test the worste case szene which have wide view, lot of buildings over a big area with a lot of polygons, Vega loose 25% performance also when this feature is on. Where Vega should benefits most.

Because Vega (and possibly polaris too) doesn't really need it. It's for older GCN versions that more easily get geometry bottlenecked.

Here's dice's presentation on what I believe is a very similar technique.
https://frostbite-wp-prd.s3.amazonaws.com/wp-content/uploads/2016/03/29204330/GDC_2016_Compute.pdf
 
It's kind of weird. You'd think the graphics options would have tooltips to give recommendations on whether you should turn a feature on/off: "May help with x generation video cards", "May benefit x configuration of RAM", "May benefit PCs with x CPU cores" etc
 
It's kind of weird. You'd think the graphics options would have tooltips to give recommendations on whether you should turn a feature on/off: "May help with x generation video cards", "May benefit x configuration of RAM", "May benefit PCs with x CPU cores" etc

I think the tooltip recommends it on for AMD and off for Nvidia.
 
No Vega10 love in here?

it has DiLDO's everywhere!

I wonder how much more advanced then Vega64 it is in terms of power management? my Vega 56 can sit at very low clocks but does it gate off the majority of CU's/Tu/ROP i dont know.
 
That depends on the VRAM pool though. Using 4k textures at 1080P on a budget machine should show the intended gains. HBCC really is a HPC focused tech, that also happens to provide an interesting opportunity for gaming. For high end gaming the likely outcome would be a notice from your ISP upon installing it.

If there are any tools that can still bias LoD that might provide some interesting testing opportunities.
Would be interesting but that needs to be proved regarding 4K textures at 1080P and its use with gains, but that is rather a niche use.

The context not just from myself but many others IS gaming, AMD even went out to promote it with gaming generally.
Regarding HPC, well we discussed that in the past and there it is up against Unified Memory/scaled up and out resource management, but that is a different topic and we are still to see how a practical HBCC solution with the diverse libraries and framework works in the real world with HPC applications and architectures.

So coming back to gaming where this has been for last X pages is HBCC worth the cost considering it seems to need more than 4GB to perform with games?
Whether that be reducing available system memory in general down to 8GB (where 16GB seems to be the primary consumer purchase these days for gaming and would mean 8GB of that for HBCC) or the value proposition of upgrading to 32GB to use 8-16 GB of that for HBCC while still ensuring plenty of general system memory resources.
Part of that equation would be the price and type-spec and timings of memory the user/gamer requires.

Edit:
The question may too limited as HBCC is pretty easy to enable/disable.
 
Last edited:
When you set the HBCC size in the drivers, it's "up to", no ? It's won't eat up all the allocated memory if not needed... ?

Here I've a Vega FE (so 16gb), an d 32gb of ram. HBCC at 24gb (slider say 24gb, so I guess it's 16+8). But in the Windows task manager, it's showing 40gb available for the gpu (16 shared+24 dedicated). It makes no sense, I guess windows doesn't really understand what the driver is doing. I hope game and apps do :eek:
 
When you set the HBCC size in the drivers, it's "up to", no ? It's won't eat up all the allocated memory if not needed... ?

Here I've a Vega FE (so 16gb), an d 32gb of ram. HBCC at 24gb (slider say 24gb, so I guess it's 16+8). But in the Windows task manager, it's showing 40gb available for the gpu (16 shared+24 dedicated). It makes no sense, I guess windows doesn't really understand what the driver is doing. I hope game and apps do :eek:
Does raise the question why 4GB gave worst performance than 8GB of HBCC or even HBCC off in one of the games at all resolutions tested by Computerbase.de

That said one benefit of HBCC is that it is pretty easy to enable/disable so my last post and question of worth is a bit too absolute.

Edit:
Windows 10 does not support any tech like Unified Memory, so what Vega is doing is partially hidden from the OS.
Raises interesting point if one can create a memory resource conflict if it becomes too limited; I do not see that being an issue in general with higher memory availability but more of a curiosity.
 
Unfortunately , I'm betting no laptops will launch with 3000+ GT/s memory . It would be interesting to test the competitivity with Intel at this lower power brakets
 
Unfortunately , I'm betting no laptops will launch with 3000+ GT/s memory . It would be interesting to test the competitivity with Intel at this lower power brakets
How much 16 GB of very high clocked memory consumes power? Do you really want to combine it with a 15W SOC?
 
Back
Top