Radeon GPUs Are Increasingly Competing With NVIDIA GPUs On Latest RadeonSI/RADV Drivers

[max0x7ba]

There is really no reason for any self respecting open-source enthusiast to run NVidia hardware any more. Sold my GTX Titan last year, replacing it with an Rx 480 and have not looked back. All my Steam games play really smoothly on Ultimate settings. Everything just works, and works really well - no more binary blobs or proprietary kernel modules to compile. AMD FTW!

I was exactly that enthusiast with my Vega 64 LC until I needed to do deep learning on GPU. All deep learning frameworks work with CUDA, none with OpenCL or MlOpen. AMD patched the first release of Tensorflow somehow a year ago, but it is woefully obsolete and insufficient now. See the (sad) state of AMD in deep learning over at https://rocm.github.io/dl.html

So, it looks like all big words from AMD about new era of machine learning is just hot air. AMD much touted compute capacity is only applicable to mining, it seems.

[GI_Jack]

From what I gathered during my switch from 1060GTX to RX560 - miners are not that interested in mid-range AMD. Don't know why, they have their own little science/woodoo.

welp, timing could not be better I know my next video card will be AMD.

CPU will likely still be intel because ME can be nerfed, PSP, not so much. That is it.

[treba]

AMD is open source so you can go look..

I was talking about their proprietary driver

[Melcar]

Are these tests done with anti aliasing on? I also can attest for the great performance with open source drivers, until I run AA. Compared to Windows I get dismal performance with AA. I can use up to 4xAA @1440p with my old RX480 on most games post 2013 with acceptable performance, yet under Linux any AA at all makes me look at sub 30fps numbers.

[chithanh]

No, miners care about maximum profits/day, which may or may not be at optimum perf/W. Especially not where electricity is cheap and while the size of mining operations is limited by hardware availability.

[linuxgeex]

Depends. Miners often heavily overclock the cards with little to no care against damaging the hardware. And if a chip is run at highest power level all the time, this accelerates electromigration.

But anyway, GPU vendors are now reportedly selling GPUs directly to miners, who manufacture their own boards. This alone should relax the retail availability in the graphics card market, but spells bad news for the used market.

Actually, pro miners usually underclock and undervolt the cards to achieve peak performance per watt. Novice miners overclock, and when you buy a card from an individual on amazon/ebay you are probably getting your card from a novice, so then yes your assertion is correct. If you get your card used from a reseller who gets them in batches of 100+ from pro miners, then no.

[OneBitUser]

Cards that were used for mining are actually better to buy second-hand.

Silicon chips (CPUs, GPUs) tend to be damaged by temperature changes a lot more than by a consistent high temperature. Having a consistent load on the card actually doesn't damage it as much.

That's true for the GPU itself, unfortunately the same 24/7 uptime absolutely kills VRM and memory chips.

Mining GPUs tend to be run at their "sweet spot" perf./Watt-wise, while during gaming they are usually pushed to their limits, often by the manufacturers themselves (this is especially true for AMD cards).
But miners OC memory, which will take it's toll. It also doesn't help that most graphics cards only have token memory cooling.
The same can be said about the VRMs: ineffective cooling+24/7 uptime wears them out.

As far as I know manufacturers test the endurance of components basically the same way as miners use GPUs, i.e. continuous load 24/7.

[andre30correia]

Running hybrid laptop using Kaveri and Hainan (Sea Island and South Island respectively) with kernel 4.16 snapshot, Mesa 18.0.0 and enabled amdgpu driver on Gnome 3.28 Wayland , the navigation is smooth. Some works need to be done on RADV which currently failed to render Vulkan scene despite passing Vulkan validation for both CIK and SI.

vulkaninfo | grep device
WARNING: radv is not a conformant vulkan implementation, testing use only.
WARNING: radv is not a conformant vulkan implementation, testing use only.
WARNING: radv is not a conformant vulkan implementation, testing use only.
WARNING: radv is not a conformant vulkan implementation, testing use only.
VK_KHR_get_physical_device_properties2: extension revision 1
deviceID = 0x6665
deviceType = DISCRETE_GPU
deviceName = AMD RADV HAINAN (LLVM 6.0.0)
deviceID = 0x130d
deviceType = INTEGRATED_GPU
deviceName = AMD RADV KAVERI (LLVM 6.0.0)
deviceID = 0x6665
deviceType = DISCRETE_GPU
deviceName = AMD RADV HAINAN (LLVM 6.0.0)
deviceID = 0x130d
deviceType = INTEGRATED_GPU
deviceName = AMD RADV KAVERI (LLVM 6.0.0)

nice, ty for feedback

[chithanh]

Cards that were used for mining are actually better to buy second-hand.

Silicon chips (CPUs, GPUs) tend to be damaged by temperature changes a lot more than by a consistent high temperature. Having a consistent load on the card actually doesn't damage it as much.

Depends. Miners often heavily overclock the cards with little to no care against damaging the hardware. And if a chip is run at highest power level all the time, this accelerates electromigration.

But anyway, GPU vendors are now reportedly selling GPUs directly to miners, who manufacture their own boards. This alone should relax the retail availability in the graphics card market, but spells bad news for the used market.

[tajjada]

But please keep in mind: You do not know about the history of the individual card that is sold to you then. Maybe it was running at the outer rim of specs (or beyond) 24/7 for a long time? Can't imagine that this is good for the card.

Cards that were used for mining are actually better to buy second-hand.

Silicon chips (CPUs, GPUs) tend to be damaged by temperature changes a lot more than by a consistent high temperature. Having a consistent load on the card actually doesn't damage it as much.

With gaming, you launch a game, the gpu heats up quickly, you play for a few hours, and then it cools down, rinse and repeat. These temperature cycles are very bad for the silicon and damage it over time (microscopic expansion and contraction of the materials due to the temperature putting repeated strain). Kind of like bending a wire back and forth over and over again. It physically wears out the material and eventually the strain is too much and it snaps.

With mining, the card is running 24/7 and stays at a consistent (albeit hot) temperature. There are no temperature differences to physically strain the materials. This is much more sustainable and better for the gpu, despite the high temperature.

The thing that usually gets damaged on mining cards is the fan. If you by a used card from a miner, you are probably gonna have to replace the fan on the card. Other than that, it is probably going to work great.

Because of this, configuring your gaming system to automatically switch to mining on the gpu whenever you are not playing games could actually prolong the life of your card, because it avoids the temperature cycles by keeping it consistently under load.