AMD silently nerfing the performance of older GCN cards

[raphidy]

AMD finewine tech going stale.

 

Personally, don't put features that you will remove later. It makes me remember of linux on ps3. Just don't do it in the first place because the internet will know and you'll pay.

[leadeater]

17 minutes ago, AluminiumTech said:

AMD's Dev tools easily differentiates them and their drivers enable specific features depending on what GCN architecture version they are using

Yea I am aware of that, hence I mentioned simplified generalization. Problem lies in actually how these potential bugs crop up, identifying them and if they are common reoccurring issues. Known bugs and the documented fixes are easy and get rolled in to game development tool sets or in to in-house tools/processes, but if developers are constantly having to implement their own fixes all the time which all stem from a common theme then it's time for a global disable of that feature.

 

If you hit multiple different issues across multiple different games on the same hardware architecture that utilize async compute in different ways then that is a problem.

 

Also I'm talking about the actual game developers not the game engine developers who actually have the resources and skill sets to identify these wide spread issues and communicate these down to their customers either through updates or just advisories.

 

Anyway from that linked twitter post it may not actually be a thing.

 

P.S. As a dual 290X user I don't really care if it got disabled on mine also, these cards are maxed out as it is and no amount of magic and DX12 is going to fix it. At some point more pure processing power is required, the ever marching progress of technology is a fact of life we all accept at some point.

[jimakos234]

6 hours ago, TheRandomness said:

 

Ok, I'll give you that, it was about 2am when I posted that  But I still disproved your last point, saying 'not many of these goodies will be availible for older cards'. It's pretty much everything that wasn't a part of the original 7000 series launch (which was 2012~, the bonaire chip was launched in 2013).

Dude i was specifically talking about these (the 280s) cards, if you take a closer look at my post

Anyway, we have to wait and see if tahiti will be partially supported, since there were not much info about it in the announcement.

[Notional]

 

 

[Notree]

As far as I'm aware its not even  AMD driver related but the benchmarks/games in question, which now only support gcn1.1 and above. Async is still on older cards .

[Kwee]

I just tested the last driver(16.12.1) with clean install, Async Compute still disabled

 

If you have a GCN 1.0 gpu, download this tool https://mega.nz/#!IJNQ0S4a!oLn5-FzqBqZ2potDqYRtY6nahF3bcvAPZQ26PUIMFvA and give me the results.

[Drak3]

17 hours ago, LAwLz said:

"Who needed async compute anyway? It barely made a difference."

For the GPUs in question, it really didn't make a difference. It'd be different if they did it with the R9 lines, where there can be a tangible difference. 

 

17 hours ago, LAwLz said:

"Buy AMD, they don't gimp their cards with future drivers"

The issue is that NVidia is seen for gimping their cards the second the next generation, or the second generation after that. However, We've seen the HD7000, HD8000, R9 200, R9 300, and RX lines before any gimping occurred,  and we're already half way through the first RX generation before it happened. We also don't know if it is intentional yet as it only occurred in a single driver.

[ivan134]

5 hours ago, Prysin said:

https://mobile.twitter.com/BitsAndChipsEng/status/806617175037513728

 

you sir have some explaining to do. And also some drivers to uninstall with DDU

Isn't bits and chips an AMD shill site though?

[Prysin]

4 minutes ago, ivan134 said:

Isn't bits and chips an AMD shill site though?

isnt any site not daily bashing AMD a AMD shill site?

 

[AresKrieger]

8 hours ago, Space Reptile said:

PhysX requires Cuda3 or higher , wich means 600 or 700 series cards or higher , async requires GCN2 or higher , wich means R9 290 or R9 300 series and up

I'm not familiar with how physx worked prior so did it ever support 500 series cards and earlier or did Nvidia simply only add the feature to its most recent cards and future cards when they acquired the tech.

 

Apparently it supports more cards than you said

[Space Reptile]

Just now, AresKrieger said:

I'm not familiar with how physx worked prior so did it ever support 500 series cards and earlier or did Nvidia simply only add the feature to its most recent cards and future cards when they acquired the tech.

they aquired ageia and physx in 2008  , that was before the GT 200 series was launched , those cards where the first ones to support it 

[Energycore]

Quote

allegedly

From the source

I think I will make the decision not to trust any article that begins with that word.

 

If true; dammit AMD, we like you because you don't do these things. But people on those cards can just not update their drivers like us Nvidia users have been doing since 361.91

[AresKrieger]

29 minutes ago, Prysin said:

isnt any site not daily bashing AMD a AMD shill site?

 

No, but their are certain sites that try to give a neutral view and actually do rigorous testing, as for Bits and Chips I actually have never heard of them prior to this thread thus what they say means very little (though this info came from a site called "legitreviews" so their is a possibility it is not legit )

[LAwLz]

28 minutes ago, ivan134 said:

Isn't bits and chips an AMD shill site though?

It is. You just have to look at their twitter feed to realize it.

 

Here are some of their latest tweets:

Async is still working.

Nvidia are enemies with everyone else and is ruining the world.

AMD is the best place to work for LGBT equality.

Saying that a huge Intel client will shift from Xeon to Zen but they can't say the name of the client.

Nvidia will exit the automotive market in 2018-19.

The A10-9700 beasts the GTX 1080 and RX 480 at FP64.

 

 

That's pretty much their tweets from the last week.

 

 

 

5 minutes ago, AresKrieger said:

I'm not familiar with how physx worked prior so did it ever support 500 series cards and earlier or did Nvidia simply only add the feature to its most recent cards and future cards when they acquired the tech.

PhysX works on any Nvidia GPU that has 32 CUDA cores, 256MB of VRAM and is from the 8-series or later. I have no idea why he says you need a card that's at least 600 series or higher, because that is false.

For those wondering how old the 8 series is, here is a timeline:

8 series - 2006

9 series - 2008

200 series - 2009

400 series - 2010

500 series - late 2010

600 series - 2012

[Mira Yurizaki]

17 minutes ago, LAwLz said:

The A10-9700 beasts the GTX 1080 and RX 480 at FP64.

In their defense, it doesn't take very much to beat an NVIDIA GPU at FP64. The GTX 1080 is only good for 257 GFLOPS of FP64 performance, the RX 480 gets 323 GFLOPS.

 

But yeah, an APU beating an RX 480 at anything is suspect as hell.

[Prysin]

7 minutes ago, M.Yurizaki said:

In their defense, it doesn't take very much to beat an NVIDIA GPU at FP64. The GTX 1080 is only good for 257 GFLOPS of FP64 performance, the RX 480 gets 323 GFLOPS.

 

But yeah, an APU beating an RX 480 at anything is suspect as hell.

not really.

Fiji and Polaris is made for 3D workloads, not raw compute. Basically, their FP64 is something akin to 1/32 in the case of Fiji, same with Maxwell and pascal. Thus making them completely worthless in FP64 compute.

 

if the APU has 1/2 or 1/3rd FP64, with a base FP32 of 1.3-1.5 TFLOP, then it will destroy the much bigger and more powerful GTX 1080 and RX 480 in FP64. It is all about shader configuration, L2GPU

[Mira Yurizaki]

3 minutes ago, Prysin said:

not really.

Fiji and Polaris is made for 3D workloads, not raw compute. Basically, their FP64 is something akin to 1/32 in the case of Fiji, same with Maxwell and pascal. Thus making them completely worthless in FP64 compute.

 

if the APU has 1/2 or 1/3rd FP64, with a base FP32 of 1.3-1.5 TFLOP, then it will destroy the much bigger and more powerful GTX 1080 and RX 480 in FP64. It is all about shader configuration, L2GPU

https://techaltar.com/amd-rx-480-gpu-review/2/

 

This to me suggests the RX 480 isn't 1/32 FP32 FP64 performance, considering it handily beats a Titan XM.

[79wjd]

1 hour ago, Drak3 said:

For the GPUs in question, it really didn't make a difference. It'd be different if they did it with the R9 lines, where there can be a tangible difference. 

 

The issue is that NVidia is seen for gimping their cards the second the next generation, or the second generation after that. However, We've seen the HD7000, HD8000, R9 200, R9 300, and RX lines before any gimping occurred,  and we're already half way through the first RX generation before it happened. We also don't know if it is intentional yet as it only occurred in a single driver.

The 300 series and 200 series are essentially the same generation, and the 7000 series is mostly the same generation. So technically, the RX series is the first (and a half-ish) generation since the 7000 series.

[Drak3]

31 minutes ago, djdwosk97 said:

The 300 series and 200 series are essentially the same generation, and the 7000 series is mostly the same generation. So technically, the RX series is the first (and a half-ish) generation since the 7000 series.

Not really, while it doesn't seem like there are differences on the surface, there are subtle differences in the hardware and firmware that separates the generations.

It also doesn't change the fact that it took 5ish years for AMD to apparently "gimp" their cards (even though the difference is marginal at best).

[Mr.Meerkat]

8 minutes ago, Drak3 said:

It also doesn't change the fact that it took 5ish years for AMD to apparently "gimp" their cards (even though the difference is marginal at best).

Which is why I feel the argument of Nvidia not optimising their cards anymore after the first or second year is still valid. 

 

A 7970 is about as fast if not faster than a 680 now with the 290X being as fast if not faster than a 780ti (at 1080p and 1440p) and is almost able to compete against a 980. Even without Async compute, GCN 1 owners can't complain. Ever since they bought it, AMD has done their driver work and allowed them to perform considerably better than what they were able to even when compared to the competition. 

[RagnarokDel]

19 hours ago, Kwee said:

Somes games received optimization on last drivers. If you play news games and a game that required old drivers, it's not a easy way to switch every time you want to play.

In this case this is more between 5 and 7% more. 56,1-> 59,8 ; 47,9 -> 50,8 ; 39,1 -> 43,3

Congratulations, you agree with me? Guess what 5% of 16 is.

[Drak3]

6 minutes ago, Mr.Meerkat said:

Which is why I feel the argument of Nvidia not optimising their cards anymore after the first or second year is still valid. 

 

A 7970 is about as fast if not faster than a 680 now with the 290X being as fast if not faster than a 780ti (at 1080p and 1440p) and is almost able to compete against a 980. Even without Async compute, GCN 1 owners can't complain. Ever since they bought it, AMD has done their driver work and allowed them to perform considerably better than what they were able to even when compared to the competition. 

To be fair, launch day drivers on AMD's side are stable, but in terms of unlocking the performance potential, pretty horrendous.

[79wjd]

23 minutes ago, Drak3 said:

Not really, while it doesn't seem like there are differences on the surface, there are subtle differences in the hardware and firmware that separates the generations.

It also doesn't change the fact that it took 5ish years for AMD to apparently "gimp" their cards (even though the difference is marginal at best).

Subtle difference that make it a refresh, not a new generation. AMD launched the 200 and then 300 series as new generations, but since the cores were basically the same as their 7000 series counterparts there wasn't much they could do other than artificially restricting performance on older cards. It's less a case of Nvidia gimping older cards and more of a case of them not prioritizing driver support for older cards -- whereas AMD's older generation was using essentially the exact same set of drivers as the current generation due to being almost identical on a hardware level.

[Drak3]

1 minute ago, djdwosk97 said:

Subtle difference that make it a refresh, not a new generation. AMD launched the 200 and then 300 series as new generations, but since the cores were basically the same as their 7000 series counterparts there wasn't much they could do other than artificially restricting performance on older cards. It's less a case of Nvidia gimping older cards and more of a case of them not prioritizing driver support for older cards -- whereas AMD's older generation was using essentially the exact same set of drivers as the current generation due to being almost identical on a hardware level.

Generations are defined by the manufacturers and developers of the technology, not the users.

Otherwise, Kaby Lake wouldn't be regarded as a generation (It's literally Skylake with a higher clock and onboard TB3 controller). Haswell-E wouldn't be part of the 5th generation lineup, nor Broadwell-E as the 6th either.

 

Also, they're not almost identical. There is enough variance in the generations that each generation requires driver optimization.

[Kwee]

Double post