Radeon glmark2 scores

Below are to be found some Radeon scores from the glmark2 benchmark. Version 2014.03 of the benchmark was used, compiled on OpenSuSE 13.1 (gcc 4.8), and run on Ubuntu 16.04 LTS (kernel 4.4.0, Xorg 1.18.3, module version 7.7.0 of free radeon driver, Mesa 11.2.0, DRI2, LLVM 3.8). The default window size of 800x600 was used on a 1920x1080 display, and the window manager was fvwm2 (i.e. free of compositing effects). The machines were otherwise idle.

All of the cards were passively cooled, and were purchased for under £50 (some second-hand).

Scores are in frames per second (FPS).

Some of the scores are repeated, using the same binary, but running under Ubuntu 18.04LTS (kernel 4.14.0, Xorg 1.19.6, Mesa 18.0.5, DRI3, LLVM 6.0). For the Terascale cards, these show scores almost a factor of two faster, showing how much of a difference software can make even for a rather mature hardware product. The column headed 18.04 gives these scores.

So far tests with Ubuntu 20.04LTS (kernel 5.4.0, Xorg 1.20.8, DRI3, Mesa 20.08, LLVM 10.0.0) have shown little difference on Terascale cards, but the performance of the GCN 1 cards is back to 16.04LTS levels, the 15% regression of 18.04LTS having disappeared. Tests with Ubuntu 22.04LTS (kernel 5.19.0, Xorg 1.21.1, DRI3, Mesa 22.0.1, LLVM 13) are similar to 20.04.

Ryzen 7600 HC 3.8GHziGPURDNA 21282200(128 bit DDR5/4800)7672
Kaby Lake QC 3.7GHzPCIe v3 x8GT 1030Pascal384150064 bit GDDR5/600071923
Haswell QC 3.1GHzPCIe v3 x8RX5504GCN 45121206128 bit GDDR5/600066827336
Haswell QC 3.1GHzPCIe v3 x8R7 250EGCN 1512800128 bit GDDR5/4500470340324689
Ivy Bridge QC 3.3GHzPCIe v3 x8HD 7750GCN 1512800128 bit GDDR5/450046103984
Haswell QC 3.1GHzPCIe v3 x8R7 2405GCN 1320(?)780128 bit DDR3/16002050
Ivy Bridge QC 3.3GHzPCIe v2 x8HD 6670T'scale 2480800128 bit GDDR5/40001757
Ivy Bridge QC 3.3GHzPCIe v2 x8HD 5670T'scale 2400775128 bit GDDR5/40001719
Ryzen 5 QC 3.2GHzPCIe v2 x16HD 5670T'scale 2400775128 bit GDDR5/400014942975
Haswell QC 3.1GHzPCIe v2 x8HD 6570T'scale 2480650128 bit DDR3/1600 1944
Haswell QC 3.1GHzPCIe v2 x8HD 6570T'scale 2480650128 bit DDR3/133396817291709
Core2 QC 2.4GHzPCIe v2 x16HD 5570T'scale 2400650128 bit DDR3/10006341371
Haswell QC 3.1GHzPCIe v2 x8R5 230 FlexT'scale 216062564 bit DDR3/1600542940910
Haswell QC 3.1GHzPCIe v2 x8HD 6450T'scale 216062564 bit DDR3/1333483
Core2 QC 2.4GHzPCIe v2 x16HD 5450T'scale 28065064 bit DDR3/1333388
Haswell QC 3.1GHzPCIe v2 x8HD 5450T'scale 28065064 bit DDR3/1333378
Sandy Bridge QC 3.2GHzPCIe v2 x8HD 5450T'scale 28065064 bit DDR3/1333378802
Kaby Lake QC 3.5GHzPCIe v2 x8R5 230T'scale 216062564 bit DDR3/1066377
Athlon II DC 3.0GHzRS880/HD 4250T'scale 140500UMA & sideport559
Core2 DC 2.4GHzPCIe x16X550RV3704400?128 bit DDR/500?2981,2
Core2 QC 2.83GHzPCIe x16HD 3450T'scale 14060064 bit DDR2/800250
Core2 QC 2.4GHzPCIe x16HD 4350T'scale 18060064 bit DDR2/800246488
Athlon II DC 3.0GHzRS880/HD 4250T'scale 140500UMA483
Core2 DC 2.4GHzPCIe x16X1550RV5004550?128 bit DDR2/800?14111411
Athlon II DC 3.0GHzRS880/HD 4250T'scale 140500Sideport179
Core2 DC 2.4GHzPCIe x16X300SERV3704325?64 bit DDR/400?841,2

T'scale: Terascale
GCN: Graphics Core Next

Memory speeds given in MT/s. Memory clock speed is one half this (one quarter for GDDR5).

1 does not support GL version 3.0 and unable to run the Terrain benchmark from this suite. Score is average of those it could run, which will flatter the result as the Terrain benchmark is generally the slowest.

2 failed to complete four other benchmarks correctly, substituting a dummy shader, and failed to run the Jellyfish correctly. Again, this flatters the results as the dummy shader is less work than a proper shader.

3 using nvidia's binary drivers. With the open-source noveau drives, this scores under 300 on 18.04. (This passive card is both newer and more expensive than the Radeon ones.)

4 the card used was actively cooled, but it is quite low power and a passively-cooled version would be technically possible. This card used the AMDGPU driver, not the radeon driver.

5 most cards based on this entry-level GCN 1 GPU have a maximum resolution of 1920x1200 on their digital outputs, whereas most of the older Terascale 2 cards can manage 2560x1600.

The performance of the RS880, part of a common AMD chipset c 2010, depends strongly on the exact memory configuration. It mostly uses UMA (unified memory, shared with the CPU), but it may have a small amount of "sideport" memory mostly for use as a framebuffer. If used this enables the CPU to enter a lower powerstate whilst the video signal is still active, and slightly increases the performance. The sideport memory is DDR3/1333, and it appears to be just 16 bits wide, though some sources say 32 bits. As a frame buffer, the highest bandwidth it needs to support is 330 Mpixels/s over a DVI-DL connector, which, at four bytes per pixel, is 1.33GB/s. DDR/1333 at 16 bits has a theoretical peak of 2.66GB/s.


The benchmark does seem to be testing things which are entirely offloaded to the graphics card, in that very significant changes in CPU power have no impact.

The benchmark seems to be heavily influenced by GPU memory speed, particularly when running Ubuntu 16.04. The HD 5670 would be expected to be about 20% faster than the HD 5570 if the constraint was GPU processing power. It is over 2.7 times faster, presumably reflecting a memory subsystem four times as fast. Similarly the only difference between the R5 230 and the HD 6350 is that the memory is 20% faster on the former, leading to a score over 12% higher. Conversely the HD6670 has much better GPU performance than the HD5670 (20% more cores and 3.2% higher clock speed), yet the same memory, and scores just 2.2% better. One should be careful when purchasing cards: I purchased the HD 5570 believing it to have DDR3/1333 memory, which would probably have increased its score by about a hundred FPS if it were true.

The GCN-based card seems to be able to make much better use of the available memory bandwidth.

The Terascale and GCN-based cards switch between different power states, changing GPU and memory speed, and GPU voltage. The Terascale-based cards seem incapable of transitioning from their high-power state whilst a 2560x1440 monitor is active, save for the XFX HD5670 (the other cards were mostly Sapphire).

Upgrading from Mesa 11.2.0 (Ubuntu 16.04) to 12.0.6 made no significant difference to the above scores. The Haswell with the HD 5450 achieved an identical score, for instance.

Comments on the different DVI, HDMI and DisplayPort standards.

Video Decoding

The above Radeon cards all include some version of AMD's "Unified Video Decoder" (UVD). This may sit on the same die as the GPU, but it is separately clocked, and is mostly independent in generation from the rest of the GPU. Firefox and Chromium are notorious for not using it, but mplayer and VLC can. UVD 2 can decode two H264 1080p streams simultaneously, and is found on the HD4000 series and later. Only very recent cards (and none of the above) can decode H264 at 4K, H265 or VP9.

The line in /sys/kernel/debug/dri/0/radeon_pm_info reading

uvd    vclk: 0 dclk: 0

refers to the current clock rates of the UVD. Here, quiescent. The program vdpauinfo will describe the capabilities of the UVD.