It's probably the PM code in fglrx
Hawaii introduced a new scheme for this in their windows drivers. It probably isn't fully there yet on their linux driver.
Details: Anandtech R9 290X PowerTune info
And yes, the high temps are expected. AMD designed these parts to run hot.
Sorry if mentioned in the article, but were you running the tests in Quiet mode, or in Uber mode?
Thank you for good review. Great to see more apps and especially more demanding apps reviewed.
One thing that bothers me. I did not find anything about noise. On windows side this card got driver update just before release which raised fan speed from 40% to 47%. Reason for this was to compete with gtx780 which got price cut before release of r9-290. With new powertune these hawaii chips throttles clocks down to maintain targeted 95°C. So if linux default fan speed is 40% it throttles like hell. So what are the linux defaults? Is there settings for fan speed in catalyst control center?
There is no quiet/über modes in r9-290 non-x. There are 2xsame bios with same fan settings.
Originally Posted by devguy
At 95c, these cards are designed to fail, now at triple rate. Post High K Dielectric processes on the sub 45nm level aren't heat-friendly. From the past, we have learned the basics of the cooling - lower temperatures = lower wear, higher temperatures = higher wear.
The regression isn' t the driver, the regression is the chip itself. Its like nvidias 4xx, but made even dumber.
I wonder what temperatures do FirePro variants will run at.
Exactly. And this has been done to compete with Nvidia.
Originally Posted by brosis
This is an overclocked card. Period. 95ºC is not acceptable. At this rate, gfx cards are going to consume more energy than an air conditioner.
The R9 290 (especially the non-X variant that Michael owns) had a lot of improvements for clocking/fan speed/performance with Windows Beta 8 and Beta 9.2 which hadn't been released for Linux. The next Catalyst should hopefully catch up.
Its is a safe benchmark. R290 has 5Tflops of 32bit SP. GTX680 has 3.2Tflops of 64bit SP "or" 6.4Tfops of 32bit SP (each 64bit core has 1ALU and 2FPUs). On the other side AMD has Drystone. Integer execution on Radeons 5000+ are close to DP/FP, its actually 2-3 times that of Nvidia, see GPU Coin Mining. Now a game its mostly FP+ASIC, but an inside Driver FX like Antialising can be optimized for Integer, and many Graphics Engines can do the same duo to consoles. Any way Nvidia cannot have big difference. The difference that you see on closed Linux vs closed Windowz drivers are probably that Linux Catalyst has Hacks/cheating on only for Source and Unigine. Any way all those things don't matter for Linux. We want good quality Libre "Gallium only" Drivers (not targeting PTX). And we want finally some Distro to have Mesa and Wine with the D3D9 State Tracker on their repositories. After that, we will find bugs on that State Tracker and we will correct them, but not before. We don't want to face a so delayed merge, because some of you you are not ready. What is the problem, you have a pain in the ass and we all waiting to pass???
95 C is hot for microelectronics, not all solid state devices
In fact, while 95 C would be extremely hot for a CPU and is hot for a GPU, there are other, much larger fab process devices for which this is only warm. Most of the large RF transistors used in radio transmitters are good to 150 C, which is a standard for silicon discrete transistors. Some ruggedized devices will even tolerate 200C, temperatures in both cases being limited by crystal regrowth that tends to destroy junctions over time. Radio amateurs often push them hard, as a ham needing to replace a $2 IRF 510 once a year could often care less. I've seen a mismatched transistor get hot enough to boil water on its case for short intervals with no damage.
Originally Posted by schmidtbag
That being so, if AMD found a way to reduce electromigration and crystal regrowth, I could see them raising temperatures and reducing heatsink size/mass and fan noise. All the same, I don't like what hot metal masses due to circuit boards over time, they too must be ruggedized if high temperatures are to be run.
Nice set of benchmarks
That was a great set of benchmarks, Michael. Comprehensive, wide selection of applications, plus performance per watt.