Goodbye ATI

[artivision]

Measuring general flops is wrong, the right thing is measuring mac-flops. Example: madd(x=a*b) is 1 general flop because does a *action or a +action, but is 2-mac-flops because does action in 2 objects a,b. fmac(x=a*b+c) is 2-general-flops because does 2 fused actions a* and a+, but is 3-mac-flops because does action in 3 objects a,b,c. So fmac is 1,5-2 times faster than madd. Examples:

gtx280: 240mimd-cores(32bit)*1,4ghz*2-3(madd+mul)= 1-mac-tflop

radeon6900: 384vliw4-cores*4-32bit-executions*900mhz= 1,35-general-tflops(32bit) *2(madd)=2,7mactflops(32bit)

gtx580: 512mimd(dual issue)cores*1-64bit or 2-32bit executions*1,55ghz*2-3(fmac)= 1,6-general-tflops(64bit) or 3,2-general-tflops(32bit) or 4,7-mac-tflops(32bit)

radeon7000: 512simd4-cores*4-32bit-executions*900mhz*2-3(fmac)= 3,8-general-tflops(32bit) or 5,7-mac-tflops(32bit)

gtx600: 2xCores(gtx580)*2xBitrate(128bit quad issue, at the same transistors)= 6,5-general-tflops(64bit) or 13-general-tflops(32bit) or 20-mac-tflops(32bit).

Wile amd gains 2xFlops/watt per generation, nvidia gains 4x after gtx280. Also amd has mediocre opengl and bad d3d to ogl translation, regardless of the generation. Amd is not even close in programmability to fermi and kepler, no full native integer, no full native 64bit, no good vm like cuda for wide many language support. Amd is not cheap, I can buy a gtx-460 new for 100 box in my country, overclocking to 1,8-1,9ghz and I have near gtx580 or 75%-radeon7000 performance, and I use wine-mediacoder_cuda for h264 encoding.

[Dandel]

I've also switched to nvidia (GTX 560). Solved problems (using closed nvidia driver):

• games under wine work,
• accelerated video decoding without problems (more formats supported),
• hibernation/sleep problems solved (under flgrx it usually worked),
• vsync works,
• KDE works without screen problems (seems like uninitialized video memory)

Before I used HD4870, this card is great but only under Windows (it is almost as powerful as my current nvidia card). Using fglrx wasn't so bad, problem was that there were many many small issues that made the overall experience quite hard.

The vertical sync issue seems to be getting a bug fix within the next two releases (See the Unofficial bug report)

[artivision]

FMA is not instruction set and it is general, but not certain. It is just the two functions of Madd(*,+) fused. You must have a well written and clean code in order to achieve more than 95% fusion. The BitcoinXXX lacks, here some benchmarks: http://www.xbitlabs.com/articles/gra...k_5.html#sect3 In Unigine with heavy tessellation, Fermi crushes Radeon by 1.5-2x!!! Any way I want to correct my previous post for gtx600. I have new information: gtx600 has only 3.5 billion transistor 28nm against 3bt 40nm of gtx580, has 1024 64bit cores (gtx500 kind), and 1.8ghz clock with "turbo bust" near 2ghz, with only 180w (30% less than gtx580). Offers 2.5x the gtx500 performance and 3.4x the per/watt efficiency (4Tflops@64bit - 8Tflops@32bit - 12macTflops). Also offers 2.1x the radeon7970 performance with less watt, that means nVidia can give a 1536 cores card in the near future (3 months later). Any way here is PHORONIX-linux(libre), we not speak for something that does not work on Linux. So why some of you speak about Radeon cards here?

[artivision]

New radeon7970 has FMA replacing old Vector Vliw5 with new scalar arch. Radeon7970 is only 20% faster than gtx580 and less programmable, what to do against ftx600(2,5x faster). Tessellation is a control engine, not an execution unit, the main program runs on shaders, more flops=more speed. Many GB-ram has nothing to do with speed. And radeon does not work on Linux.

[artivision]

And a Radeon7970 has 1.1ghz maximum clock not 1.4ghz. http://www.tomshardware.co.uk/AMD-Ra...ews-37352.html

[artivision]

I don't count open source drivers because they don't exist, even the gallium r600 driver has closed firmware. About 1.4ghz radeon, that was with ice cooling, the true maximum clock is 1.1ghz: http://www.tomshardware.co.uk/AMD-Ra...ews-37352.html Radeon7970 is only 20% faster than gtx580: http://www.xbitlabs.com/articles/gra...c_6.html#sect3 The end gtx600 specs are 1536(64bit) cores @1.8+ghz, in 28nm, 3.5Btrans, 200+-w: http://www.techpowerup.com/162500/GK...Explained.html and http://www.techpowerup.com/162504/NV...-Detailed.html How a card 3.6x faster than gtx580 and 3x faster than radeon7970 will lose to someone? Even the 1/3cores 130box mini Kepler will be magnificent!

[DaemonFC]

I don't count open source drivers because they don't exist, even the gallium r600 driver has closed firmware. About 1.4ghz radeon, that was with ice cooling, the true maximum clock is 1.1ghz: http://www.tomshardware.co.uk/AMD-Ra...ews-37352.html Radeon7970 is only 20% faster than gtx580: http://www.xbitlabs.com/articles/gra...c_6.html#sect3 The end gtx600 specs are 1536(64bit) cores @1.8+ghz, in 28nm, 3.5Btrans, 200+-w: http://www.techpowerup.com/162500/GK...Explained.html and http://www.techpowerup.com/162504/NV...-Detailed.html How a card 3.6x faster than gtx580 and 3x faster than radeon7970 will lose to someone? Even the 1/3cores 130box mini Kepler will be magnificent!

Nouveau has free and open source firmware. They asked Nvidia if they could distribute theirs and Nvidia said no, so the Nouveau project came up with their own.

I guess it's an odd example of a company being openly hostile to free and open source software where some good came out of it.

I doubt we'll ever be able to use most of our devices without blob firmware. Sometimes it's hard to convince a company to allow their blobs to be redistributed. So many distributions like Debian and Fedora give up and try to rationalize blob firmware as being somehow different than proprietary software so that they can look like they are not violating their own policies.

[RealNC]

Hardware is worth nothing if the drivers suck

[curaga]

That goes both ways, drivers are worth nothing if the hardware sucks

[RealNC]

Indeed. Good thing the hardware doesn't suck ;-)