AMD Radeon R9 300 Series MegaThread: FAQ and Resources

[Mousemonkey]

Tradisonally folding has always been better with nvidia hardware. But since they update the client with OpenCL some probably would think that AMD now can do better than nvidia since when it comes to opencl it seems AMD are much more proactive in supporting the latest spec of opencl.

But in my opinion just because they were much faster than nvidia supporting the latest spec it doesn't mean they also better at it than nvidia. I always believe that nvidia is very capable to support opencl properly but they purposely holding back to push their CUDA ecosystem more.

Anyway it seems Vulcan will also going to use OpenCL quite extensively. So i think this will give nvidia a reason to fully support OpenCL spec ASAP.

They may be "holding back" I have no info on that but the R9 380/285/7950 gets its arse handed to it by the GTX960 when it come to F@H! :lol:

Specific benchmarks favor specific hardware.

F@H has never been a strong point for GCN. Still generally GCN offers more throughput for Open CL than nVidia in many applications.

The *best* compute processor in the GCN lineup is Hawaii I believe as it offers AMD's highest DP rate...

I also think from memory that Tahiti is stronger than Tonga at compute tasks (so the HD7950 / 7970 are probably a better option than the 380 / 380X in ppd).

Other folders have confirmed that the PPD for a 7950 and 380 are pretty much about the same, memory has no bearing on the outcome it seems and F@H switched from a CUDA based client to an OpenCL based client to even the playing field IIRC so GCN being pants speaks volumes for their claim to support open standards and software IMHO.

doesn't the latest version of openCL used by the program matter just as much as what version is supported? amd supports 4.3 but if F@H only uses openCL 3.0 then NVidia isn't behind... only if the software utilizes the latest version and NVidia has to downgrade to an older version to be compatible will you see a loss based on version.

regardless I get back to my original question for you mouse. every aspect of every card gets benchmarked today. there is no way you couldn't have found exactly what the 270x was capable of before making the purchasing decision. do not complain about how a piece of hardware isn't what you expected when it was given hundreds of reviews as it was comparing it to every other piece of hardware created in the last 5 years. You should have known before you bought it that it was slower than a 960 at f@h and therefor you should have bought the NVidia card as you knew you were only going to use f@h and 960 was faster for the money spent.

I am an amd fan I will not lie, but I would never claim that a NVidia card I bought was a bad card because a similarly priced amd card performed faster an one specific task. I may say I got a bad deal, may say I should have gotten the amd card as I mostly only play this game, but never say the NVidia card is bad. This is low key fanboy bashing.

Were here to discuss future cards and gpu rumors essentially and this IMO is out of the spectrum not to mention its the R9 thread...

Really?

 

[cdrkf]

@mouseonkey, then have you looked at your dp performance? I'll bet the 380 is faster at dp, given how massively nvidia cripples dp on Maxwell.

I'm also curious to know what real world dp looks like Tahiti vs Tonga. I have a 280, I'd be open to running a session of f@h on it both single and double to get some numbers if you'd be interested?

I'm genuinely interested as technically gcn should be the faster architecture based on the hardware. My guess would be it has a decided advantage in dp, whilst nvidia is stronger at sp.

What I'm unsure of is which is more useful to the f@h project? I'm guessing dp is needed or they wouldn't offer it? Do they take sp results and refine them or is the dp used for something else entirely?

 

[Mousemonkey]

@mouseonkey, then have you looked at your dp performance? I'll bet the 380 is faster at dp, given how massively nvidia cripples dp on Maxwell.

I'm also curious to know what real world dp looks like Tahiti vs Tonga. I have a 280, I'd be open to running a session of f@h on it both single and double to get some numbers if you'd be interested?

I'm genuinely interested as technically gcn should be the faster architecture based on the hardware. My guess would be it has a decided advantage in dp, whilst nvidia is stronger at sp.

What I'm unsure of is which is more useful to the f@h project? I'm guessing dp is needed or they wouldn't offer it? Do they take sp results and refine them or is the dp used for something else entirely?

For true F@H results just download the client and let it run for a couple of weeks, those "benchmarks" don't mean squat as you don't have any option to select DP or SP in reality.

 

[renz496]

AMD seems impressive with DP but still hard to see AMD firepro being use in super computer. If you look at the top500 list the listing were flooded by system using nvidia tesla and xeon phi. Many thought that xeon phi should be meh comparing to nvidia tesla and amd firepro but the reality is not like that. In fact the fastest supercomputer right now were using intel xeon phi. And right now when it comes to compute accelerator the one that really compete probably intel and nvidia.

 

[-Fran-]

This might be a bit old, but it is relevant to the discussion: https://foldingforum.org/viewtopic.php?f=67&t=22444 (and page 2).

Basically, AMD drivers use the CPU too much and they lose performance. To make a simple test, measure how much of the GPU is being used when you fold, MM. If it's not pegged at 100%, then you need a better CPU for it until AMD fixes that behavior.

Plus, F@H does not do it's thing using DP: https://foldingforum.org/viewtopic.php?f=61&t=25069

So, AMD will always be so so compared to nVidia in F@H not because of favoritism. It's a design choice, since they saw no benefit from using DP over SP.

Cheers!

 

[renz496]

If they rely on DP then those with gtx titan and titan black will dominate the score. Even AMD did criple radeon DP performance starting with hawaii. That's why if you look bench that actually measuring DP performance even 280X come out on top of 290X

 

[fudgecakes99]

too many games are used to factor into the overall performance graph on techpowerup. who cares if world of warcraft gets 200fps on a 980ti and 150fps on a fury x. find another irrelevant game and the tables turn. averaged out it throws off the relative performance graph. and they have no min/max or 99th% frames on top of that. i guess i shouldn't say i dont trust them, their individual game benches are ok i guess, although lacking more precise measurements.

but then how would we know whats better i mean. Which company has the bigger e-pe***. The world must know these things ya now.

 

[Mousemonkey]

@Yuka, I have already stated that AMD cards/drivers do NOT use the CPU when folding so having a powerful CPU is kind of pointless if all the machine is doing is folding.

 

[-Fran-]

@Yuka, I have already stated that AMD cards/drivers do NOT use the CPU when folding so having a powerful CPU is kind of pointless if all the machine is doing is folding.

The implication is not around "heavy CPU use" per sé, but "CPU dependency of drivers". Now, that information is old, so I will take your word for it.

Cheers!

 

[de5_Roy]

AMDGPU PowerPlay Is Working Great So Far; Here's An Ubuntu PowerPlay Kernel
http://www.phoronix.com/scan.php?page=news_item&px=AMD-GPU-PowerPlay-Init
AMD Publishes AMDGPU PowerPlay Support For Re-Clocking / Power Management
http://www.phoronix.com/scan.php?page=news_item&px=AMDGPU-PowerPlay-Patches


sorry if re-post:
AMD Raja Koduri Promises Two Brand New GCN GPUs in 2016 Based on FinFET Process - Focusing Discrete GPU Market Share Gains in 2016 - 2017
http://wccftech.com/amd-raja-koduri-promises-brand-gcn-gpus-2016-based-finfet-process-focusing-discrete-gpu-market-share-gains-2016-2017/

 

[cdrkf]

AMDGPU PowerPlay Is Working Great So Far; Here's An Ubuntu PowerPlay Kernel
http://www.phoronix.com/scan.php?page=news_item&px=AMD-GPU-PowerPlay-Init
AMD Publishes AMDGPU PowerPlay Support For Re-Clocking / Power Management
http://www.phoronix.com/scan.php?page=news_item&px=AMDGPU-PowerPlay-Patches


sorry if re-post:
AMD Raja Koduri Promises Two Brand New GCN GPUs in 2016 Based on FinFET Process - Focusing Discrete GPU Market Share Gains in 2016 - 2017
http://wccftech.com/amd-raja-koduri-promises-brand-gcn-gpus-2016-based-finfet-process-focusing-discrete-gpu-market-share-gains-2016-2017/

That last link worries me a bit... two gpus won't account for a full lineup, which suggests they are going to rework and rebrand at least one of the current GCN gpu's *again*...

The real question is which one? My guess would be one of the lower end models (oh god *please* don't respin Pitcarin yet again, it was a great gpu when you launched it but it's GCN 1.0)... maybe they'll demote Tonga to an entry position and have 2 gpu's above it? Or maybe one above to replace Fury and then one below?

 

[-Fran-]

That last link worries me a bit... two gpus won't account for a full lineup, which suggests they are going to rework and rebrand at least one of the current GCN gpu's *again*...

The real question is which one? My guess would be one of the lower end models (oh god *please* don't respin Pitcarin yet again, it was a great gpu when you launched it but it's GCN 1.0)... maybe they'll demote Tonga to an entry position and have 2 gpu's above it? Or maybe one above to replace Fury and then one below?

Well, it depends on what he considers a "new GPU".

If we give him the benefit of the doubt and making major tweaks to GCN can be considered a "new GPU", then I'd love to see that. Remember AMD and before (ATI) used to have 2 main GPU designs. If you take a look at nVidia, they also have 2 main. So, that would mean they would basically put forward 2 hitters: heavy and mid.

If we don't give him the benefit of the doubt, then it will be another Fiji type of upgrade over Tonga. Maybe change the memory controller to support HBM, improve the compression mechanism, attack a few bottlenecks and add a few more transistors to it. They won't be full re-works of GCN and close to lipstick on a pig IMO. And to be honest, I'm afraid this is the most probable one 🙁

Cheers!

 

[renz496]

So new gpu in mie 2016 or late 2016?

 

[fudgecakes99]

late 2016 early 2017

 

[renz496]

If we going by AMD cadence which is new series for every two years then i would expect AMD R 400 series at least in 2017 time frame. Would nit be surprise if we see actual shrink of current gpu instead of new series or GCN 2.0. In the end Raja only say about new gpu using FinFet tech not new series or GCN 2.0 specifically. And when it comes to launching new product AMD have the tendency to let the time line to slip further than originaly planned.

 

[Guest]

From 2 new gpu family AMD can have 6-7 new gpus. Not bad for 1 year.

 

[Guest]

Improvements in Tonga and fiji's performance in linux for amdgpu drivers.
http://www.phoronix.com/scan.php?page=article&item=amdgpu-powerplay-test&num=2

 

[renz496]

Hard to say. To me it sounds like AMD would only come with two new gpu next year while the rest is rebrands as usual. But one thing for sure AMD probably need something in sub 75w segment.

 

[cdrkf]

@renz496 what frustrates me so much is AMD replaced the capable sub 75w hd7750 (512 cores) with the far weaker oland gpu (384 cores). I mean why?!

I used to use the 7750 for low power multi screen setups for training rigs however they are really hard to come by now.

 

[renz496]

Well not the first time. Current 370 also based on 265 config instead of 270/270X. Many said that AMD no longer need sub 75w card because that segment was taken by APU instead. But i don't think so.

 

[-Fran-]

http://www.anandtech.com/show/9792/amd-sc15-boltzmann-initiative-announced-c-and-cuda-compilers-for-amd-gpus

Uuuh... Interesting readings/interpretations can be obtained from that.

I wonder how that might change adoption for FirePros.

Cheers!

 

[Vogner16]

Well not the first time. Current 370 also based on 265 config instead of 270/270X. Many said that AMD no longer need sub 75w card because that segment was taken by APU instead. But i don't think so.

most say if you looking for a gpu that is low performance/ low wattage you should just get older tech used. price to performance simply isn't there.

7970 on ebay is same cost as R7 370. and faster.

only reason to have the new card is if you system simply couldn't power a larger gpu. in that case Id say upgrade the power supply or like renz said get an APU.

There are just better cheaper solutions than sub 75W gpu's out there so its impractical to use them in 90% scenarios.

 

[renz496]

Well what i mean is AMD did need to fill that sub 75w segment despite AMD thinking they did not need it. True while APU can address the segment but did all budget OEM machine use AMD APU? That is the market where sub 75w card can address. Nvidia was confident to sell 750ti for 150 because they know AMD has nothing comparable to compete with. They drop 270 price down to 150 to compete with 750ti. But despite that 750ti still quite popular. Because you don't need to replace your OEM power supply with 750ti. True 270 is significantly better at the same price as 750ti but you still need to replace your PSU which will add to the cost.

 

[renz496]

@Yuka

That's mean AMD admit that nvidia as a significant player in HPC space. Nvidia for sure will do all their might to defend their position. It is a market they try to dominate ever since they come up with CUDA back in 2007-2008. Back in 2013 nvidia acquire PGI to bolster their HPC solution more. Will be interesting to see how things will pan out in the future.

But when it comes to HPC intel also becoming a threat to both AMD and Nvidia. In fact if you look at top500 listing you'll be surprise to see the usage between xeon phi vs tesla vs firepro. If you look at the list it is pretty much two horse race between nvidia and intel.

 

[Vogner16]

But when it comes to HPC intel also becoming a threat to both AMD and Nvidia. In fact if you look at top500 listing you'll be surprise to see the usage between xeon phi vs tesla vs firepro. If you look at the list it is pretty much two horse race between nvidia and intel.

top 500 doesn't apply to majority of servers, just the really expensive ones. to get a real picture look not only at the top 500 servers but all running servers. firepro does not have a lot of high end servers but they may have a lot of gpu's in the grand scheme.

for example if intel xeon phi had 200 computers with their gpus for a total of 50,000 gpu's and most to all of them were in the top 500 and amd had firepros in 800 computers for a total of 40,000 gpu's and most of them were in the slower than top 500 it shows amd has nearly as many gpu's on market as intel, despite having almost none in the top 500.

**I make up my numbers its just to prove a point**