Radeon Linux 4.11 + Mesa 17.1-dev vs. NVIDIA 378.13 Graphics Performance

[linuxgeex]

OK, this is not working.

I am trying to explain that the old vs new gap has been generally related to specific transitions, eg VLIW to GCN, radeon to amdgpu, closed-to-open for gaming focus etc, and that those transitions are largely behind us now (and so should not be generalized to the present day).

Your response was to go back in time and cherry-pick examples from the two biggest transitions in our history - your first benchmarks were around the VLIW to GCN transition, and your second benchmark focused on R9 285 at the point where the initial amdgpu code had been pushed but before power management had been enabled.

Please try to separate the transitions from the current state. It will help you give better advice, seriously.

Please try to separate your optimism for what is to come, from the reality of what consumers have experienced for the last several years.

I didn't "cherry pick" - there doesn't exist a period from 2008 to 2015 where you can't find a newer AMD GPU performing slower than a previous generation GPU in the same class. Period.

I have already granted you that with Polaris you have taken a huge step forward. Why are you still arguing that? Is there something I should know? Am I wrong?

[linuxgeex]

Just to be clear, we are talking about the Mesa / RadeonSI / AMDGPU driver stack here:

Unfortunately this is not true. Launch-day performance of new AMD hardware using open drivers has typically been poorer than 2-year old hardware. That is not adequate performance. I agree that AMD's launch-day performance has improved over the years, but if you are upgrading cards quickly you are going to be painfully disappointed by AMD. If you buy a new card every 2 years and you are buying a year-old card when you do so, then, possibly, your statement becomes true.

4) if you are not an enthusiast and just want bang for the buck and you care about openness, then you can buy year-old AMD hardware inexpensively and Linux will have strong open support for it, generally good enough that you won't be tempted to dual boot, sometimes it will even exceed the performance on Windows. :-)

I don't think this has been the case for the last few years. It did take us a while to bring open source driver performance on GCN in line with earlier generations (since the shader compiler toolchain was all-new) but for at least the last couple of years new and older hardware has had roughly the same level of performance relative to HW capabilities.

We have also caught up with new HW introduction, which is a big change from when we launched GCN. When we restarted open source driver efforts we were 3-4 generations behind what was shipping and 5-6 generations behind what was in development, so in addition to building up the open source driver we had to catch up on a lot of HW generations. That is all done now.

Hmm... this does not bear out what you are expecting us to believe. AMD's Open driver performance has not been at even 50% on launch day, at least not until the Polaris launch in June 2016. So you can't begin to claim that AMD has a "years" long track record of strong Open driver support within the first year of launch.

I know you WANT to make that claim, and I know that you INTEND to make it happen, but you don't yet have a single year of track record under your belt to make that claim.

I WANT to believe you, and I INTEND to buy AMD hardware again when there is more truth to your statement that AMD has a track record of strong driver support within the first year of a product launch. You're almost there. Let's not sully it with false claims, please.

[snip]

I really do want to believe that AMD is on the road to being great again, so let's take the praise I've given AMD for what it is, and not force me to cite more examples of why people should, given AMD's track record, steer clear of the company altogether. Because [years of] history is not your friend. Stick to the present. You're doing well now.

You are missing my point, or I'm not making it sufficiently clearly. I am not making any specific claims about *what* the numbers are in terms of performance at launch vs HW potential, but I *am* saying that whatever those numbers are they are now pretty much the same for newly launched HW and for 1-2 year old HW at the same point in time.

Rather than comparing Fury & Tonga a year ago vs today, you need to look at that ratio against the same relative performance for earlier HW across the same time period. I'm not saying that is the only comparison you are allowed to make, but it is the one which illustrates the point I am trying to make.

That will obviously vary a bit depending on whether each new HW generation has new HW features that require substantial driver development, but in general I think you may be taking specific situations (3+1 to VLIW transition,VLIW to GCN transition) and generalizing them across all old vs new HW... and that's what I am disagreeing with.

If the same situation happened again (significantly new programming model requiring new shader compiler and significant core driver changes, AND not being able to start driver work until fairly close to HW launch) then I agree the same result would probably happen again, however we are now able to start driver development much earlier in the HW lifecycle than before so in general you will see much less old-vs-new delta than in the past.

Here's a list of benchmarks from this very site showing previous-gen cards outperforming next-gen "same day" with updated drivers 8 months to years after launch:

Mid-range AMD GPUs:
Radeon HD 6870, 22 October 2010, USD $239
This is our starting point

Radeon HD 7850, 19 March 2012, USD $249
Jan 14 2014, almost 2 years later, still being beat by cheaper previous gen hardware:

http://www.phoronix.com/scan.php?pag...eri_gpus&num=2

Radeon R9 285, September 2 2014, USD $249
No launch day open driver support, one year later being absolutely slaughtered by 3-year-old Radeon HD 7850:
Sept 2015 https://openbenchmarking.org/result/...HA-AMDGPULIN89

Radeon R9 380, Nov 19 2015, USD $229
Unfortunately I couldn't find many benchmarks of this hardware, but results are poor.
July 2016 https://openbenchmarking.org/result/...DREH-150827130

Radeon RX 480, 29 June 2016, USD $239
This hardware had good support and performance at launch, and keeps getting better. Good start AMD.

To recap:

I've suggested that given AMD's track record, consumers who care about open drivers should buy older hardware to help ensure they get good support and good value for the money. I provided examples to support my suggestion.

bridgman has said that you should ignore AMD's track record, and just buy the new hardware. Just trust him. He knows, after all he works for AMD, and has a completely unbiased opinion on this topic. He believes that all the problems are behind us now.

Unless they aren't:

If the same situation happened again (significantly new programming model requiring new shader compiler and significant core driver changes, AND not being able to start driver work until fairly close to HW launch) then I agree the same result would probably happen again, however we are now able to start driver development much earlier in the HW lifecycle than before so in general you will see much less old-vs-new delta than in the past.

Facts are important to me. If you were the poor sucker who purchased a Radeon HD 7850, then 2 full years later you were still looking at performance that was weaker than the previous generation AMD hardware of the same class and price point. If you were the poor sucker who purchased an R9 285 Tonga, then a year later it was still being beaten by that same 5-year old Radeon HD 6870. Neither of these pieces of hardware was a wise purchasing decision within one year of release. It should be obvious to the reader that if some AMD hardware is not a wise purchasing decision within one year of launch, and if some AMD hardware is still outperforming newer hardware 5 years later, that the wisest time to purchase AMD hardware, in a general case, falls somewhere between 1 and 5 years of release.

Company track record also means a lot to me. I wonder what would happen if Honda adopted the same attitude towards their track record as bridgman has adopted for AMD? How many cars do you think they'd sell if their motto was "we think we've solved all the problems now - ignore everything you've seen in Consumer Reports and JD Power for the last 10 years and just buy one, buy one now!"

Having said all of that, the Polaris hardware was well supported on launch day and the drivers continue to improve all the time. In this one special case, you'll probably do well to purchase the Polaris RX400-series [RX 480 8G @ $229.99 linked] hardware without waiting a year or more for the drivers to mature. But I wouldn't blame you if you wanted to wait for the price to drop as it's been 8 months, 22 days since launch and Vega is just around the corner. Vega should (fingers crossed) also have good launch-day support and rumour has it that it can beat the GTX 1080 (with the closed drivers... and we've heard a similar story to this one before, lol)

PS Michael the RX 480 review lacks an affiliate link to Amazon. hint hint.

[bridgman]

I've suggested that given AMD's track record, consumers who care about open drivers should buy older hardware to help ensure they get good support and good value for the money. I provided examples to support my suggestion.

bridgman has said that you should ignore AMD's track record, and just buy the new hardware. Just trust him. He knows, after all he works for AMD, and has a completely unbiased opinion on this topic. He believes that all the problems are behind us now.

I said nothing of the sort. With respect, you are not even summarizing your own statements accurately, let alone mine.

You are saying that prospective buyers should focus on "AMD's track record" (specifically open source driver performance while the code was under development) and base their buying decisions on that, saying that "company track record" is more important than recent behaviour. You are also ignoring the point that we were shipping closed source drivers with higher performance all that time. You acknowledge that Polaris was well supported on launch day but are treating that as a one-time exception rather than as a reflection of the maturing code base.

I am saying that prospective buyers should focus on more recent experiences with the open source drivers (good and bad) because that is a better predictor of what is likely to happen in the future, since most of the work done for existing hardware carries forward to new hardware as well.

I can live with something in between those viewpoints, particularly given the size of the Vega code changes, but I do think your focus on performance averaged across 10 years (while we + community were building the current code base) is missing the point.

You are treating the open source driver evolution as a quality problem (where the expectation of "reverting to older behaviour" is not unreasonable) rather than as a development effort (where expectations evolve with the code), and I guess that is the main thing I am disagreeing with.

[Kendji]

Guys, as a potential buyer of either an RX 480(the card I prefer) and GTX 1060, the thing I care about is not any track record, but what is the sacrifices(pain points) I have to make with AMD compared to nVidia? I don't give a shit if AMD had poor driver support previously, but if they can now push a lot improvements it makes it worth my consideration. The player on the AMD driver scene that I'm watching is Steam, they have recently hired people to improve the drivers etc. Key sacrifices for me would be HDMI sound and certain games not working at all. I do not care if I use open-source or partially open-source, sure I prefer the open-source, but if AMDGPU-Pro is better that is what I'm going to use (with the goal of eventually going for the open-source version, while I wait for the drivers to mature). I'm holding back on buying because I don't seem to get a straight enough answer, though my HDMI answer seems to be awn sered, AMDGPU-pro currently supports, open-source not (point and click version of drivers, you need to mess with kernel etc. to install them). What games don't work or does all games across the board now work? Also I'm using virtualization and I use 3d applications there, how does the drivers affect this?

[linuxgeex]

I said nothing of the sort. With respect, you are not even summarizing your own statements accurately, let alone mine.

I summarized my own statements very accurately. If you read something else from them, it was because you were not hearing what I was saying, or because you forced me to digress into pointing out some less-than-completely-honest things you had to say.

You are saying that prospective buyers should focus on "AMD's track record" (specifically open source driver performance while the code was under development) and base their buying decisions on that, saying that "company track record" is more important than recent behaviour. You are also ignoring the point that we were shipping closed source drivers with higher performance all that time. You acknowledge that Polaris was well supported on launch day but are treating that as a one-time exception rather than as a reflection of the maturing code base.

I am saying that prospective buyers should focus on more recent experiences with the open source drivers (good and bad) because that is a better predictor of what is likely to happen in the future, since most of the work done for existing hardware carries forward to new hardware as well.

And that, by itself, would be a fair statement to make, if you didn't make-believe that the Open Driver experience hasn't been a hideous one for consumers, resulting in performance from their new hardware being slower than hardware 2 entire generations older. I didn't even ask you to own up to that fact. I just politely asked you to stop bald-faced lying about it. But polite hasn't worked. So now you get the more direct approach.

I can live with something in between those viewpoints, particularly given the size of the Vega code changes, but I do think your focus on performance averaged across 10 years (while we + community were building the current code base) is missing the point.

Actually, you've been missing the point since I made the buy recommendation for consumers who care about Open Drivers to purchase year-old AMD hardware. in fact you've been punishing me for making that recommendation. So I will withdraw it.

May I remind you that in every rebuttal I have made, I have stressed that I am talking about consumers who care about Open Drivers. And I have Capitalized it persistently to help drive the point. You keep ignoring it. Case in point:

You are also ignoring the point that we were shipping closed source drivers with higher performance all that time. You acknowledge that Polaris was well supported on launch day but are treating that as a one-time exception rather than as a reflection of the maturing code base.

That is exactly what I mean. I have said over and over again - for consumers who care about Open Drivers. Your sometimes stable, sometimes not, sometimes supported by the kernel, sometimes not, sometimes supported by X11, sometimes not, fglrx drivers, were never an option for someone who either ran on a rolling release (where AMD never managed to release a driver that would link with the kernel, and XCB was broken about 30% of the time) or someone who cared about using Open Drivers. You want to stick your head in the sand and pretend that because you offered a binary alternative, that somehow that made everything alright. I have not "ignored" it. I repeatedly pointed to the alternative - that is not the same as ignoring it.

Over and over I have said, this recommendation is for people who care about Open Drivers. Over and over, you have not heard it.

That speaks volumes.

I reiterate: Polaris has done well. AMD is doing well now. Stop sullying your success with lies.

You are treating the open source driver evolution as a quality problem (where the expectation of "reverting to older behaviour" is not unreasonable) rather than as a development effort (where expectations evolve with the code), and I guess that is the main thing I am disagreeing with.

No, I'm weighing the benefits and advantages of your products on the merits of your products, from the point of view of someone who both cares about the Open Drivers, and has had a dozen years of dual-booting to a second Linux installation to run FGLRX only to play games, because it was incompatible with an up-to-date system. And that is all.

My best recommendation, actually, is to run the high-end Broadwell Iris Pro... it's stable, supported, open, and nobody is trying to ram false claims about its current or past performance history down your throat.

I'd like to begin to recommend AMD again. Polaris really has been a well executed launch on the Open Drivers, there's a good chance that it will continue to be good value moving forward. But coming from a company whose spokesperson who has claimed (and not retracted, even in the face of evidence) that their hardware has for years performed the same or better level relative to its hardware capabilities on the Open Drivers, same day, versus their older products... well... that is making me feel pretty slimey about his employer.

[bridgman]

I'd like to begin to recommend AMD again. Polaris really has been a well executed launch on the Open Drivers, there's a good chance that it will continue to be good value moving forward. But coming from a company whose spokesperson who has claimed (and not retracted, even in the face of evidence) that their hardware has for years performed the same or better level relative to its hardware capabilities on the Open Drivers, same day, versus their older products... well... that is making me feel pretty slimey about his employer.

Let me take another look. Your examples were not as good as you believe, since they focused on the one-time events I said were *not* part of the overall pattern, but one of them reminded me that the transition from radeon to amdgpu resulted in not having dynamic power management for a while.

I had been thinking about the CI to VI transition as being the first one where existing code worked almost equally well on newly launched HW (which I still think is the case), but I had forgotten about amdgpu and DPM. If the delay there was a long one relative to Tonga/Fiji launch then I will retract the "last year or two" part of what I said and agree with you about the history.

[jnrivers]

I'm a little disappointed with the results. I was expecting Team Red to place a little better with the shader cache enabled, which was the only reason for reading more than just the first page of this article.



The RX480 does seem to be very good value, and obviously there's still a lot of room for improvement as it's performing quite a bit stronger on Windows at the moment.

There's a lot of speculation about how the actual compute capabilities of the RX480 are better than the 1060, and I have a lot of faith in the RadeonSI team... it's just speculation but down the road it may overtake Team Green's competitive offering when it ceases to be in NVidia's best interest to keep optimising drivers for what has become an older piece of hardware. So I feel that what is good value today will just keep getting better. But that's just my rampant AMD fanboi optimism speaking lol. Chances are very good that by the time that happens I will have moved on with different hardware, as has happened several times already in the past.

I have to agree with your "rampant AMD fanboi optimism" AMD cards do seem to get better with age. Next desktop build will have a AMD card assuming the prices eventually bounce back after this cryptocurrency mining craze.