How Old ATI GPUs Can Be Faster On Open Drivers

[phoronix]

Phoronix: How Old ATI GPUs Can Be Faster On Open Drivers

A few days ago when publishing the results of benchmarking a lot of graphics cards on their Gallium3D drivers (about a dozen graphics cards) this left a number of people surprised. A number of these results from the open-source Gallium3D drivers illustrated the older graphics processors as being much faster than the newer hardware, even though the newer hardware is far superior to the vintage products. This shouldn't have been a surprise if you stay up-to-date with the Linux graphics news on Phoronix, but it comes down to features found in the older Gallium3D drivers not yet implemented in the newer open-source drivers...

http://www.phoronix.com/vr.php?view=OTA5OQ

[crazycheese]

How can you claim that money from FirePro purchases flows to fglrx but also claim that money from other graphics purchases does not flow to the open source drivers ? In can understand one claim or the other, but not both.

No one can see if I actually use opensource driver with non-workstation purchased card. Could you please clarify where does money for windows driver come, ie how do you (AMD) calculate how much money should be placed to correct upcoming bugs in windows driver? Graphics hardware alone does not work without driver. I'm not claiming you (AMD) should invest less in windows driver, I'm mean only that people are not allowed to (or desired that they) make a choice.

The situation may also be projected onto windows, granted a magic happens and microsoft is suddenly as democratic as ancient Athens, for example windows 9 comes and AMD still produces drivers for windows 8, because they invested lots of resources in it and do not want complete rewrite. So people will buy nvidia for tasks where windows 9 driver model is better. And this is currently the situation in Linux field. So why not to allow people vote with "certificates" proving they invested in the company.

In case it helps, we don't actually make or sell most of what you call "AMD cards". We sell chips, other companies make and sell cards. We don't assign or track card serial numbers, we don't track card sales - we just sell the chips and provide reference designs.

Yes, I know this of course. But those MAC number still lands on Gigabyte' motherboard even if the NIC chip comes from Realtek. So Realtek claiming it is not responsible for the driver, but Gigabyte; or it is Gigabyte who should care about people with issues regarding to the NIC chip... is somehow wrong..

[bridgman]

Fglrx is workstation payed driver. It is supported by people that purchase workstation fireGL.

...

- there is no money to opensource team once people on the street purchase desktop AMD card for desktop use.

How can you claim that money from FirePro purchases flows to fglrx but also claim that money from other graphics purchases does not flow to the open source drivers ? In can understand one claim or the other, but not both.

In case it helps, we don't actually make or sell most of what you call "AMD cards". We sell chips, other companies make and sell cards. We don't assign or track card serial numbers, we don't track card sales - we just sell the chips and provide reference designs.

[crazycheese]

crazycheese: i'll admit at this point to just skimming that monster post you wrote. i don't care enough about this topic to actually read it more carefully anymore.

However, 1 thing that stuck out to me was you think that AMD should give up fglrx and focus entirely on the OSS drivers. It's not going to happen. The only reason fglrx exists is because of the linux workstation market, and fglrx is 10+ years ahead of the OSS drivers for that specific market. It's just for desktop stuff that it sucks. They'd have to drop all the patented stuff from GL3, for example, because they wouldn't be able to license it for an open driver, and that would make their hardware useless for some of their biggest customers.

Np, I will make this answer shorter then.

See I care about this topic, because
- I preferred AMD hardware a lot over nvidia when I used (more precisely fought with) windows
- I prefer open driver from AMD much more to proprietary, because every single experience with fglrx made me want to RMA the card

I was happy to get first ever Radeon instead of Geforce 2(was young and a bit naive), to support ATI. I further used 8500 and later 9700 pro hardware, together with AMD chips. When first socketed Thunderbird started to kick slotted and socketed PIII, it was really fun. But on linux, in desktop use, ATI is just many steps inferior to nvidia, due to driver.

Yes, AMD will not give fglrx up and I do not ask them to do it. However it is not usable for following reasons:
Fglrx is workstation payed driver. It is supported by people that purchase workstation fireGL. Purchase.
When I purchase 6990, I will not support fglrx.

Being trimmed for fglrx, it completely ignores desktop software and focuses on some production software that firegl customers use. Maybe they run gpgpu tasks on it or use Maya and stuff, nothing more. This is not what desktop usage is - windows is for example desktop usage to very large extent.

Opensource driver is already usable as 2D and semi 3D driver, however and most importantly two critical points are missing:
- there is no money to opensource team once people on the street purchase desktop AMD card for desktop use.
- it does not support newly purchased cards, so even if bought for that driver, without monetary feedback, it is not usable till several years; and not fully usable for further 5 years.

I see this situation simply as unacceptable and question if AMD cares about gold midrange users at all. Not low-end hardware hackers and 2D IGP users; and not ultra high-end workstation. I simply do not clearly understand the company: they (ATI) exist solely due to selling the cards, however they do not care about selling them for linux - even if setting such mechanism up will carry absolutely <minimal> cost.

From previous post, even if approximation may be very incorrect, I projected at least 70% current nvidia card buyers come from this golden midrange (GT 450 - 570, GT 240 - 280) and use the cards for desktop linux & gaming, where only 10% of current amd users buy midrange cards (5670 - 6970) for this use. Improving fglrx driver for desktop use will bring only minimal card sells (due to nvidia being already there for many years) and will just remove more user base from opensource driver team (concurrency within ati itself).

[smitty3268]

crazycheese: i'll admit at this point to just skimming that monster post you wrote. i don't care enough about this topic to actually read it more carefully anymore.

However, 1 thing that stuck out to me was you think that AMD should give up fglrx and focus entirely on the OSS drivers. It's not going to happen. The only reason fglrx exists is because of the linux workstation market, and fglrx is 10+ years ahead of the OSS drivers for that specific market. It's just for desktop stuff that it sucks. They'd have to drop all the patented stuff from GL3, for example, because they wouldn't be able to license it for an open driver, and that would make their hardware useless for some of their biggest customers.

[Drago]

Someone can tell me :

I want to upgrade to the latest kernel GIT and still use S3TC.
Do I still need to add the "return 0;" somewhere in the cs checker stuff or I can use it without any patching?

With the latest kernel git, you will receive a snapshot of 2.6.38. You need stuff for 2.6.39, which lie in *-next branches. You need drm-next branch specifically. Please consult someone that had done it already.

[RavFX]

Someone can tell me :

I want to upgrade to the latest kernel GIT and still use S3TC.
Do I still need to add the "return 0;" somewhere in the cs checker stuff or I can use it without any patching?

[Qaridarium]

Interesting. I suppose it may be different drivers we have.

that?s it it all comes to the driver quality. he do have an intel core cpu and a hd2600 gpu more details are not in my mind.

yes true on other hardware with other drivers it works well.

I use Win2008 for development at work and routinely leave it running for a month at a time with no problems. Actually, the fact that i have to reboot once per month because of the monthly security updates is one of the annoying things i don't like about it.

yes linux do need less restarts-

There's really not any inherent reason Linux can't be hit with spyware anymore - you get prompted in windows before you can install anything just like you would in linux. However, since most malware targets windows you are correct that Linux and OSX tend to be much safer for less advanced users. More advanced users are smart enough to avoid malware on any OS.

sorry i don't believe you the security standard is is much higher and the fragmentation of the desktop does the rest one uses gnome the other use kde and another one use another desktop and another one use BSD kernel in debian6.

some people use moving targets like debian unstable you can not Hit Linux (gnu/OS-system) in the same way like windows.

I haven't tried 4.0 in linux yet, so maybe it does improve things. However, OpenGL acceleration is disabled on the OSS drivers, which is at least one thing that is disappointing. I know you can override that with an environment variable, though, so maybe some testing will show that it works on my system. Chrome probably is more state of the art, but I still like Firefox better and don't want to switch unless i really have to.

if you use a nvidia card+closed source drivers firefox4 work well like the same as in windows7.
yes sure not all hardware drivers in linux do support this right now.

but firefox4 does not run in windows faster for general its just a individual driver situation.

I've never used WINE on windows, although DosBox is pretty useful. Another good thing is the 32-bit compatibility Windows has. That probably lines up with a downside about how much disk space the OS requires, a lot of that is probably old 32bit libs.

believe me if you count ALL OLD games wine beat windows in compatibly.

[crazycheese]

I tend to say this sums it up perfectly.
This topic polarizing forever is just stupid - "haters gonna hate"...

Keep up the excellent work!

To me, every single sum up was useless and shining at same time.

Like someone said, "A conclusion is the place where you get when you get tired of thinking."
And once you stop thinking, you sleep and comply.
If you comply you accept the situation.

[crazycheese]

crazycheese:

Hey, you can be human.

Please explain how this would actually accomplish anything more than the yearly surveys Phoronix already does.

Gladly.
Lets see Phoronix survey from 2010 - the brand, the driver

1.Look at brand. I will write here the numbers for major three, just for comfort.
Intel: 2639
AMD: 3794
Nvidia:4272

What can we conclude - major player is nvidia, with amd following closely and intel having 3/5 of nvidia marketshare.
1.1) Note that everyone: intel, nvidia and amd, do have opensource drivers; but all are only good for 2D, maybe basic weak composite. We will return here to nvidia later.
1.2) Also note within all that values are IGP/HTPC and performance discrete cards scattered.

1.3) Now remember you "do NOT buy 300$ discrete amd card for opensource drivers", keep it in mind please. Even if AMD opensource drivers currently have most attention, barely anyone want to use opensource for playing quake4. He want, but cannot - he will use catalyst. Lets bind him to the brand (he will install windows for this, choiceless, but still)

2.Now look at particular driver. Again I write down values for major three.
Intel: 2514
AMD-cat: 1645
AMD-oss: 2074
AMD-rhd: 495
Nv-blob: 3923
Nv-nv: less than 80

Lets combine the driver+brand as well, to get real card usage, not someone using vesa:
I also have no idea who uses radeonhd, but lets include it too.
Intel: 2514
AMD: 4214*
Nvidia: ~4003

2.1) Note, AMD driver is used on non-AMD cards. Yeah, driver total is HIGHER than brand itself. Either incorrect data, incorrect voter behavior or someone using several drivers - at same time. I doubt radeonhd and radeon are used at same time, probably someone switches between catalyst and radeon/radeonhd. Lets cut the value of total use to real number of real cards: 3794
2.2) Also note, someone uses nvidia with.. vesa or even vmware or similar. Outdated cards? May be. But lets cut the real number of cards to match amount in drivers, so we get 4003 of nvidia users that actually use nvidia driver.
2.3) Note the amount of nvidia-noveau(nv-nv) cases is minor. Since nvidia sells chipset IGP much lower currently, primary discrete, probably single exception is ION; and mobile usage is not included in this test - lets assume almost all cases with nvidia are discrete cards with small added percentage of nvidia blob on ION(and others, like 6200/300) IGP for purpose of VDPAU.

So the corrected values of real cards, backed up by with real drivers are:
Intel: 2514
AMD: 3794
Nvidia:4003

Now Intel is close and AMD - Nvidia are nearly equal.

3.1) Note Intel is not suitable for anything gaming, performance or other things you occasionally find on linux desktop. Yes linux gaming lives and there are commercial windows games bought and played on linux. However intel is not suitable for it at all - intel "cards" are probably laptops and htpc or similar 2D. That was 2514.

3.2) Lets move to AMD - 3794 cases. Look at amount of intel "cards".Remember AMD also has various IGP northbridge motherboards(4200 etc). AMD also sells laptops. This will spread IGP and discrete to roughly 50/50 - around 1800/1800
AMD video decoding is hardly present(windows omitted from survey), so barely anyone use AMD for HTPC - Laptop or Desktop should dominate. Catalyst is not help for video decoding either. Lets cut HTPC AMD use. Will not affect current values, but may affect driver choices further.
So we are left with Laptop and Desktop. Technology-dependent (from my observations) current intel laptop solutions go more preferable over AMD due to energy drain. The optimus case however did "positively" influenced that. Also, remember quality of Catalyst, add driver switchers and people that cannot use open drivers for gaming and also do not buy $300 card to use with open drivers. Because 100% of laptops users will use opensource driver - thats 1800 for laptop, 1800 are projected discrete cards. Look at amount of catalyst users - its confirmed.

3.2.1) Now why people still use catalyst? Only lastest hardware is supported. Error rate is .. biger than competition. You cannot really game with catalyst. At least not WINE(choice given - people take nvidia for that). Barely any opensource (except xonotic) will require anything past (current) opensource drivers.
Humble Indie bundles players maybe? People that (like in 1.3) stuck with discrete AMD card? No, they will just get nvidia. But take a look.
Intel dual head? No.
Nvidia dualhead? Very painful, but possible(2d).
Noveau - possible.
Catalyst dualhead - possible, but buggy.
Radeon dualhead - possible.

If we remove rough proportion of opensource radeon to catalyst users, from that amount as well as cut minor usable noveau from that value we will land on almost exactly 1600. Maybe it would be less. Maybe people just lied in the survey and less of them use it with catalyst, lets assume 1200 only and 400 are not - using catalyst for other purposes than gaming on modern discrete cards.

3.3) I guess you understand why opensource driver has so much more use in AMD,it is used on outdated discrete cards as well. But outdated cards means people do not buy amd hardware and use old hardware(how does amd survive? No wonder it is pessimistic about linux.) Or they have no choice, but to use Linux+Windows combo (for gaming)? <<This would be a good question for survey 2011. Some may still hesitate and just use Catalyst for some titles (that 400 in 3.2.1).

3.4) Now nvidia. noveau is so extremely small - it takes almost no impact. Almost everyone use nvidia blob. What can nvidia blob do?
Gaming - yes
Gaming via wine - yes
Video acceleration - yes
GPGPU - yes
Multimonitor - almost complete fail
There are estimated 4003 people with nvidia hardware, 90% of which are discrete cards.
Those 90% make nearly the equal amount of people that use AMD solutions for various reasons.

Now look here at this "barely survey".

This is OpenGL. How much people have we calculated are "stuck" to catalyst for OpenGL reasons? 400.
Remember, its catalyst. If people just get catalyst debugged to nvidia state, the number will not increase past that of nvidia. Barely anyone will just go sell nvidia card and get amd card, because amd all of the sudden has achieved same driver state. There will be a internal fight ("shelf fight" if I recon correctly) between AMD catalyst and AMD radeon. AMD will gain nothing from it.

And if opensource driver is updated instead? How about 400 become 3800 (4000 minus ten percent for HTPC(see 2.3) minus five percent for GPGPU plus 400)? No internal fight. Almost tenfold increase in card sales. If they manage some video acceleration without getting into that always annoying, always cracked DRM/HDCP swamp or OpenCL (which are combinable if UVD cannot be touched(still what a shame) )?

See, it is normal world practice for someone who desires the product(in AMD case - discrete card) to buy that product.
Currently (based on my rought unprecise calculations) barely any consumer buys modern amd card for catalyst if he wants gaming, gpgpu or vidaccel(htpc).
There is no chance to support opensource driver developers with money, by normal practice, and if you do do that, they recieve nothing as they are ignored by sales team.

So people with desktop, representing more than third of total graphical build, just go buy nvidia card. Maybe still with a tear in the eye (remembering =( AMD opensource "attempt"). AMD allows this to happen.
Or they buy AMD card because they dual boot, looking at linux every time with a tear in eye (and remembering >=( AMD opensource "attempt"). People allow this to happen.
Either proprietary driver and linux, or proprietary OS and linux. No linux standalone.

The last thing I want to conclude is that my whole this article, whole calculations are ... garbage.
We need real world data, precision, acceptance from people, readiness to vote, readiness to take actions.

Until then, we will have regimen of Kazakhstan or Egypt.
Current phoronix(or any other, phoronix has nothing to do with it) surveys are of exact use.
You want democracy? YES!!!
What do you do? Uhhh...
You deserve what you do.
People should be allowed to vote, directly. With some money.

If the distros had a way of recording what all their users were doing and sending that back upstream that might actually be more useful, but that seems unlikely to ever happen, at least by default, because of privacy concerns.

Implementing the hardware registering via distros seem very foreign idea to me.
Why should they, if it is AMD cards. If AMD does this:
- AMD will become a moral medal from purchasers, not distros.
- the data would be secure, way more reliable
- the votes will be distro-agnostic
- since it is centralized, it will be much easier and hence quicker to accomplish
- everyone has a browser, regardless of OS. Why make it complex?
- it is actually not related to kernel, since its not bug report. It is related to AMD stat, marketing and sales team only.
- yes, privacy concerns. Would you personally be happy to register your card using only its SN(without any other data) on official AMD site?

I hope this happens. I will devote any of my time I can squeeze to this idea, so it can receive enough kicks, punches, beatings from everyone till it is formed into the acceptable state. Beaten up enough to carry AMD tag, hehe. If they won't find it ugly in the end, hehe. But if so, reasoning (hence more beating) would be also nice, not "Yr ugly, get lost"- type.

And a small remark. Dear Steve, if you read this, you can sanely ignore it. You leak lots of information back home(and not only) anyway, so you miss nothing.