NVIDIA RTX 2060 / 2070 / 2080 SUPER Linux Gaming Performance - 26 GPUs Benchmarked

[tuxd3v]

This sucks... Now NVIDIA releases a card that:
- is better than the 5700 XT (AMD's best offering)
- consumes less power than the 5700 XT

AMD come on, release that 5990 XT already with 128 cores...

I believe that, the processing power in the AMD cards, is superior,
But the drivers are not yet there..

They need to mature, and still reduce latency a bit more( even tough that is has already been done in previous releases,, )
The biggest thing to do now is power consumption profiles..
Because you can run a lot of this games with less power consumption, but the profiles are the ones that need improving right now..

If that happens, you will notice that the AMD cards will be more power efficient in this games comparison.

@Michael,
It would be nice to do a test comparison for GPGPU computing( the last one you have done, was some time ago.. ),
Even tough that Rocm haven't seen a break trough, at least that I now of..

[tildearrow]

I believe that, the processing power in the AMD cards, is superior,
But the drivers are not yet there..

I see, but do you really think a 2x+ improvement would be feasible? (so that the 5700 XT can defeat the TITAN RTX on DiRT)

(What about ray-tracing?)

They need to mature, and still reduce latency a bit more( even tough that is has already been done in previous releases,, )
The biggest thing to do now is power consumption profiles..
Because you can run a lot of this games with less power consumption, but the profiles are the ones that need improving right now..

Exactly. They apparently regressed because I could watch a video using mpv at the card's lowest core clock, but now for some reason the driver prefers raising the core clock first (before raising the memory one).

If that happens, you will notice that the AMD cards will be more power efficient in this games comparison.

Really? I thought NVIDIA was the power consumption king? (bar Polaris)

[tuxd3v]

I see, but do you really think a 2x+ improvement would be feasible? (so that the 5700 XT can defeat the TITAN RTX on DiRT)

(What about ray-tracing?)

2x improvement, seems a lot, I don't see that coming in near future..
But the drivers are the ones that need to mature, with that should come optimisations..
If we look back, we see that AMD is now with graphics performance in the middle, it was a lot worst in the past..
The idea of developing for mainline kernel, I think helped a lot also..

Ray tracing is not the best feature that some want in a Graphics card..
I for example, want decent OpenCL support..

Exactly. They apparently regressed because I could watch a video using mpv at the card's lowest core clock, but now for some reason the driver prefers raising the core clock first (before raising the memory one).

Unfortunately..
I hope AMD put in place a strategy to decently deal with Power Consumption...
I think with the drivers, should be there a user-space tool,
To do calculations of requirements.. like pushing frequencies up, for each 25mv steps, analyse were graphics card crash
In this way the user could better tune its cards, and the data would be to a database file for example, and be used at startup.

Power Consumption is one of the 'must have' features for some years now, in Europe( the cost of electricity is incredible high.. ).

Really? I thought NVIDIA was the power consumption king? (bar Polaris)

NVIDIA have indeed the best power consumption in a overall analysis,
Because NVIDIA has a nice power managment hardware/Software match, but I think also due to VLIW( which by default doesn't need to do lots of processing in the frontend, its parallelism by definition.. ).
There are situations were AMD cards can consume less power, but you need to tune them..
In Games I don't really know how it will play( I was talking about GPGPU compute tough )

I don't know of a better way.. for now I have a script that I launch on start, if any one has a advice,
I would be grateful( here 2 Poraris rx580 ), its far from ideal, but a starting point..

#! /usr/bin/env bash

### Rx 580 8GB
### Reference: /sys/kernel/debug/dri/0/amdgpu_pm_info
## HardWare Monitor
  # Power Cap 95 Watts
  echo 95000000 > /sys/devices/pci0000:00/0000:00:02.1/0000:01:00.0/hwmon/hwmon2/power1_cap
  # PWM Fan Tune Enable
  echo 1 > /sys/devices/pci0000:00/0000:00:02.1/0000:01:00.0/hwmon/hwmon2/pwm1_enable
  # PWM Fan Set to 90%
  echo 90 > /sys/devices/pci0000:00/0000:00:02.1/0000:01:00.0/hwmon/hwmon2/pwm1
## Power DPM Force Performance state
  # State Setto Manual
  echo "manual" > /sys/devices/pci0000:00/0000:00:02.1/0000:01:00.0/power_dpm_force_performance_level

  echo "s 0 300 800" > /sys/devices/pci0000:00/0000:00:02.1/0000:01:00.0/pp_od_clk_voltage
  echo "s 1 466 840" > /sys/devices/pci0000:00/0000:00:02.1/0000:01:00.0/pp_od_clk_voltage
  echo "s 2 918 880" > /sys/devices/pci0000:00/0000:00:02.1/0000:01:00.0/pp_od_clk_voltage
  echo "s 3 1167 925" > /sys/devices/pci0000:00/0000:00:02.1/0000:01:00.0/pp_od_clk_voltage
  echo "s 4 1239 950" > /sys/devices/pci0000:00/0000:00:02.1/0000:01:00.0/pp_od_clk_voltage
  echo "s 5 1282 1000" > /sys/devices/pci0000:00/0000:00:02.1/0000:01:00.0/pp_od_clk_voltage
  echo "s 6 1326 1025" > /sys/devices/pci0000:00/0000:00:02.1/0000:01:00.0/pp_od_clk_voltage
  echo "s 7 1366 1075" > /sys/devices/pci0000:00/0000:00:02.1/0000:01:00.0/pp_od_clk_voltage
  echo "m 0 300 900" > /sys/devices/pci0000:00/0000:00:02.1/0000:01:00.0/pp_od_clk_voltage
  #echo "m 1 2000 875" > /sys/devices/pci0000:00/0000:00:02.1/0000:01:00.0/pp_od_clk_voltage
  echo "m 1 2100 950" > /sys/devices/pci0000:00/0000:00:02.1/0000:01:00.0/pp_od_clk_voltage
  echo "c" > /sys/devices/pci0000:00/0000:00:02.1/0000:01:00.0/pp_od_clk_voltage
## Profile Mode "Compute"
  echo 5 > /sys/devices/pci0000:00/0000:00:02.1/0000:01:00.0/pp_power_profile_mode
## Tune SCLK to Level 2( 918 Mhz )
  echo 2 > /sys/devices/pci0000:00/0000:00:02.1/0000:01:00.0/pp_dpm_sclk
## Tune MCLK to Level 1( 2100 Mhz )
  echo 1 > /sys/devices/pci0000:00/0000:00:02.1/0000:01:00.0/pp_dpm_mclk

[tildearrow]

2x improvement, seems a lot, I don't see that coming in near future..

Then we will not be able to defeat NVIDIA...

But the drivers are the ones that need to mature, with that should come optimisations..

This is exactly what I meant.

If we look back, we see that AMD is now with graphics performance in the middle, it was a lot worst in the past..
The idea of developing for mainline kernel, I think helped a lot also..

Exactly, but this also means future improvements aren't going to be as massive as they used to be...

Ray tracing is not the best feature that some want in a Graphics card..
I for example, want decent OpenCL support..

Let's really hope that ray-tracing will just be a fad.

Unfortunately..
I hope AMD put in place a strategy to decently deal with Power Consumption...
I think with the drivers, should be there a user-space tool,
To do calculations of requirements.. like pushing frequencies up, for each 25mv steps, analyse were graphics card crash
In this way the user could better tune its cards, and the data would be to a database file for example, and be used at startup.

OK

Power Consumption is one of the 'must have' features for some years now, in Europe( the cost of electricity is incredible high.. ).

Like how high? Here there have been similar complaints, but not sure if it is higher here.

NVIDIA have indeed the best power consumption in a overall analysis,
Because NVIDIA has a nice power managment hardware/Software match, but I think also due to VLIW( which by default doesn't need to do lots of processing in the frontend, its parallelism by definition.. ).
There are situations were AMD cards can consume less power, but you need to tune them..
In Games I don't really know how it will play( I was talking about GPGPU compute tough )

Yes. And even after tuning, they may become unstable...

## Profile Mode "Compute"
echo 5 > /sys/devices/pci0000:00/0000:00:02.1/0000:01:00.0/pp_power_profile_mode

How many profile modes are there? Is there a profile mode called "Come On, Raise The Memory Clock First And If That *Really* Does Not Suffice Raise The Core Clock"?

[tuxd3v]

Like how high? Here there have been similar complaints, but not sure if it is higher here.

It depends of the zone, but around 0.3-0.32/KWh

How many profile modes are there? Is there a profile mode called "Come On, Raise The Memory Clock First And If That *Really* Does Not Suffice Raise The Core Clock"?

usually 4 profiles in rx580 4GB, the rx580 8GB there are 5.

But there are 8 Power State Modes[ 0 - 7 ]
I think it works in regards to the gfx clocks.. the cards rise first the core clock
found a paper about that, page 15:

[tildearrow]

It depends of the zone, but around 0.3-0.32/KWh

Pretend I'm saying something here about my electricity cost here, but I just don't have time to check the prices right now. (This will be edited out later)

usually 4 profiles in rx580 4GB, the rx580 8GB there are 5.

OK

But there are 8 Power State Modes[ 0 - 7 ]
I think it works in regards to the gfx clocks.. the cards rise first the core clock
found a paper about that, page 15:

They used to raise the memory clock first before 2019. What happened here?

[tildearrow]

By the way, power consumption numbers (core first vs. memory first) when playing 4K60 video:

Core clock first

amdgpu-pci-0300
Adapter: PCI adapter
vddgfx:       +1.12 V   
fan1:        1960 RPM  (min =  400 RPM, max = 4900 RPM)
temp1:        +38.0°C  (crit = +89.0°C, hyst = -273.1°C)
power1:      118.00 W  (cap = 220.00 W)

Memory clock first

amdgpu-pci-0300
Adapter: PCI adapter
vddgfx:       +0.96 V   
fan1:        1882 RPM  (min =  400 RPM, max = 4900 RPM)
temp1:        +38.0°C  (crit = +89.0°C, hyst = -273.1°C)
power1:       40.00 W  (cap = 220.00 W)

Huge waste of energy right there.

[tuxd3v]

By the way, power consumption numbers (core first vs. memory first) when playing 4K60 video:

Huge waste of energy right there.

Its indeed impressive!

I haven't look into the algorithm but if we take the slides, has something serious, and not only PR stuff,

‒ Graphics activity >0% will cause SCLK to go to the highest frequency
‒ Graphics inactivity (0% busy) will cause SCLK to go to the lowest frequency

I agree with you, at least an hybrid solution..

[tildearrow]

Its indeed impressive!

I haven't look into the algorithm but if we take the slides, has something serious, and not only PR stuff,


I agree with you, at least an hybrid solution..

Exactly. A good idea would be:

- If load is higher than 20%, increase memory clock.
- If that does not suffice (e.g. load at 70%), increase core clock.