• 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.

Something big is brewing with Radeon Software

flopper said:
fixing cpu overhead, finally?
Biggest thing they could do.
Click to expand...

I really wish people would stop repeating this as if it's fact. What people misinterpret as CPU/driver overhead is actually legacy API design vs modern GPU architecture. DX11 all the way back to the beginning are at their core single threaded. AMD's GCN GPUs are built for multi-thread, highly parallel operation. Legacy APIs simply cannot feed GCN fast enough using a single main thread. Consider this. AMD's DX12 driver doesn't exhibit any "overhead" issues. Given that they know how to make that driver, do you really think it's not within their power to make a better DX11 driver? No, it's just legacy API vs GCN architecture. It's why they pursued Mantle. It's why we have DX12 and Vulkan. They needed a new APIs to allow GCN to be used to its full potential.
 
Last edited:
I hope they unveil a shadowplay equivalent. AMD has good video encoding hardware, which is going to waste on the majority's GPU's. Shadowplay feels like it's been around for so long, and is such an obvious benefit to the NVIDIA playerbase.
 
Despoiler said:
I really wish people would stop repeating this as if it's fact. What people misinterpret as CPU/driver overhead is actually legacy API design vs modern GPU architecture. DX11 all the way back to the beginning are at their core single threaded. AMD's GCN GPUs are built for multi-thread, highly parallel operation. Legacy APIs simply cannot feed GCN fast enough using a single main thread. Consider this. AMD's DX12 driver doesn't exhibit any "overhead" issues. Given that they know how to make that driver, do you really think it's not within their power to make a better DX11 driver? No, it's just legacy API vs GCN architecture. It's why they pursued Mantle. It's why we have DX12 and Vulkan. They needed a new APIs to allow GCN to be used to its full potential.
Click to expand...
I'm sure they could spend the time and effort to refactor the base architecture of their DX11 drivers to make them more performant. I'm also sure it would cost immense amounts of time and money, for very little payback since DX11 is already on its way out.
 
Headfoot said:
I'm sure they could spend the time and effort to refactor the base architecture of their DX11 drivers to make them more performant. I'm also sure it would cost immense amounts of time and money, for very little payback since DX11 is already on its way out.
Click to expand...

DX11 will be the most popular and widely used API for many years to come.
 
DX11 is still used, but base API development has been EOL years ago. At this point might as well just brute force through it with better hardware. I'd rather not see AMD wasting resources on continuing to hack DX11 when they need to continue developing next generation APIs with Microsoft, Sony, and AAA developers along with supporting it for smaller developers.
 
Dave2150 said:
DX11 will be the most popular and widely used API for many years to come.
Click to expand...
"many" is a stretch, but it will certainly be the majority API for a while, just like every single other DX API upgrade before it will take time to move to the new one. The point is DX11 is getting reduced adoption, not increased. It's on its way out, not on its way in.

It's very, very expensive to refactor a massive codebase like that. It makes tons more sense to do a new codebase for a new API better than to try and refactor a widely used, widely tested codebase just so the tail end of a single generation has one more small checkbox for fanboys to use in internet fights.

Basically, it's a waste of time and resources. If they had done this in 2012 - different story. But in Q4 2016, not a good idea.

Incremental increases in performance is all that is reasonable to do - so I hope they continue providing small but consistent performance improvements, as they have done since GCN released.
 
AMD could use the tech they acquired when they bought with HiAlgo to optimize their low level driver code instead of using code injection. Essentially baking it right into the drivers and not risking triggering anything.
 
Despoiler said:
I really wish people would stop repeating this as if it's fact. What people misinterpret as CPU/driver overhead is actually legacy API design vs modern GPU architecture. DX11 all the way back to the beginning are at their core single threaded. AMD's GCN GPUs are built for multi-thread, highly parallel operation. Legacy APIs simply cannot feed GCN fast enough using a single main thread. Consider this. AMD's DX12 driver doesn't exhibit any "overhead" issues. Given that they know how to make that driver, do you really think it's not within their power to make a better DX11 driver? No, it's just legacy API vs GCN architecture. It's why they pursued Mantle. It's why we have DX12 and Vulkan. They needed a new APIs to allow GCN to be used to its full potential.
Click to expand...

well your fact logic wont work here.
the 390 driver suddenly improved dx11 for the 290 along the way due to amd found some oh overhead they could remove.
if the legacy whatever would be true that wouldnt have happen.
so I am right your not.
I won the prize.
 
flopper said:
well your fact logic wont work here.
the 390 driver suddenly improved dx11 for the 290 along the way due to amd found some oh overhead they could remove.
if the legacy whatever would be true that wouldnt have happen.
so I am right your not.
I won the prize.
Click to expand...

I saw that driver update and the forum threads spawned from it claiming some giant leap in performance. A 5% increase is normal driver optimization. Nvidia keeps optimizing too. Note that the 3dmark benchmark is for API overhead, which is exactly my point. They can't address the rest of the giant gap that Nvidia leads with because they can't. There at no point will be a driver update that magically gives AMD a 20% perf boost in DX11 across the board. DX11 simply cannot deliver to GCN.
 
Despoiler said:
I really wish people would stop repeating this as if it's fact. What people misinterpret as CPU/driver overhead is actually legacy API design vs modern GPU architecture. DX11 all the way back to the beginning are at their core single threaded. AMD's GCN GPUs are built for multi-thread, highly parallel operation. Legacy APIs simply cannot feed GCN fast enough using a single main thread. Consider this. AMD's DX12 driver doesn't exhibit any "overhead" issues. Given that they know how to make that driver, do you really think it's not within their power to make a better DX11 driver? No, it's just legacy API vs GCN architecture. It's why they pursued Mantle. It's why we have DX12 and Vulkan. They needed a new APIs to allow GCN to be used to its full potential.
Click to expand...

Very well written post.

Also, I believe it's very naive to even think that AMD wouldn't have already solved overhead issue , if it was fixable with drivers directly. They have a lot of talented engineers working for them. Surely not all of them can be incompetant.

Besides, people too easily claim overhead to be the problem , without even checking what thread is holding the back. Surely often in AMD's case it is that AMD driver thread that is stressing one core heavily, but so often it's also game's main logic thread. For the users performance perspective they perform similiarly, but are pretty easily identified.
 
Dygaza said:
Also, I believe it's very naive to even think that AMD wouldn't have already solved overhead issue , if it was fixable with drivers directly. They have a lot of talented engineers working for them. Surely not all of them can be incompetant
Click to expand...
This is the naive thing to think. Software development is very, very expensive. Especially for something like this. It would be millions to tens of millions of dollars of man-hours you have to divert from other projects in order to refactor a massive, well tested codebase. It could be technical debt - it could be a fundamental mismatch between architecture and API, it could be tons of things we don't have any idea about. AMD might not even know, because discovering the places to optimize is very difficult. Way easier said than done. The fact AMD and nVidia keep delivering black-box performance optimizations for dozens of games every year is nothing short of incredible. It doesn't matter how competent your people are if they dont have enough time to work on each project.
 
Last edited:
Despoiler said:
What people misinterpret as CPU/driver overhead is actually legacy API design vs modern GPU architecture
Click to expand...

Yep, you can see that when some engines run amazing on both AMD and Nvidia, while other engines (mostly single thread heavy) will do much worse.

Unity vs Frostbite vs Unreal vs others like ones Square Enix uses.
 
Despoiler said:
DX11 all the way back to the beginning are at their core single threaded. AMD's GCN GPUs are built for multi-thread, highly parallel operation. Legacy APIs simply cannot feed GCN fast enough using a single main thread.
Click to expand...

If AMD needs this higher throughput from the driver so badly, then why have they not implemented it as their competitor has? Is there a specific reason or technical roadblock that prevents it?
 
dogen1 said:
If AMD needs this higher throughput from the driver so badly, then why have they not implemented it as their competitor has? Is there a specific reason or technical roadblock that prevents it?
Click to expand...
Difference between Static Scheduling and Dynamic Scheduling. AMD does not rely on CPU to schedule tasks for the GPU. The fact that it was core design of DX11 was problem for AMD GPUs.

AMD GPUs can manage themselves.
 
If AMD could have just rewritten their drivers to better work with DX11 they would have. The fact they spent so much time and money on mantle shows that GCN was simply not optimal on DX11 and it needed something else to let it shine.
 
lifeblood said:
If AMD could have just rewritten their drivers to better work with DX11 they would have. The fact they spent so much time and money on mantle shows that GCN was simply not optimal on DX11 and it needed something else to let it shine.
Click to expand...
No. Software development is extremely expensive. It's not that easy.
 
You must log in or register to reply here.

TRENDING THREADS

Back
Top