3Comparing madvr GPU performance

My cur­rent HTPC has an nVidia GeForce GTX 1050Ti 4Gb with which I am able to upscale all con­tent to 1080p with fairly high qual­ity set­tings, and am also (just) able to play­back 8k 60fps con­tent down­scaled to 1080p. ਪਰ, I have long wondered what the bene­fits of a more power­ful card might be on the abil­ity to use high­er set­tings. Sadly, no-one seems to provide any bench­marks for mad­vr, so how can cards be compared?

After much search­ing, it seems that the strongest indic­at­or of mad­vr per­form­ance is the shader per­form­ance of a card. This is effect­ively meas­ured in FLOPS. There are some lim­it­a­tions to this as a doub­ling of shader per­form­ance does­n’t necesar­ily mean a halv­ing of render time so the rela­tion­ship isn’t necesar­ily lin­ear, but once oth­er factors (e.g. memory quant­ity and speed) are equal, the shader per­form­ance will be a good guide to the per­form­ance of a card.

It is highly recom­men­ded to get a GPU with at least 4Gb of VRAM for mad­vr. Assum­ing a new card is being obtained, the memory is likely to be at least GDDR5, and may well be GDDR6. In these cir­cum­stances the shader per­form­ance is a use­ful guide.

I also used a few spe­cif­ic pieces of inform­a­tion to com­pare dif­fer­ent bench­marks to see which type appeared to be the best fit. One piece of info I picked up is that my 1050Ti is faster than the Radeon RX 570 in mad­vr which is not reflec­ted in lots of GPU com­pute bench­marks which sug­gests these are not a good type of bench­mark to use.

Anoth­er dir­ect com­par­is­on sug­gests that the 1060Ti is faster in mad­vr than the RX580 so this should also be reflec­ted in any bench­marks used, which rules out most of them! In the same thread there is a sug­ges­tion that the Radeon RX 5700 XT is sim­il­ar in per­form­ance to a GeForce RTX 2070 Super

Some oth­er facts from the doom9 for­ums

• The AMD RX480 is slightly faster than the GTX970
• The GTX1070 is around 30% faster than the RX480

As I was­n’t able to find any bench­mark that matched these data points, instead I’ve star­ted by look­ing at all the cards avail­able in the £100 — £200 brack­et which are…

• GTX 1660 £195
• RX590 £190
• RX5500 XT £160
• GTX 1650 Super £150
• RX580 £145
• GTX 1050Ti £130

Is there a sur­pris­ing win­ner, or does the price pre­dict the performance?

The info I can find sug­gests that the GTX 1650 (non super) is around 25% faster than the GTX 1050Ti in mad­vr, and the 1650 Super is faster again (more cuda cores and a high­er clock speed). It also appears the the RX5700 XT is a little faster than the GTX 1080 and that the RX5700 (non XT) is sim­il­ar to a GTX 1070.

After lots of search­ing to com­pare these cards I found a very use­ful post on red­dit which ranks the nVidia cards approx­im­ately. These rank­ings (unsur­pris­ingly) do fol­low the price of the nVidia cards.

• GTX 1050 / 1050 Ti
• GTX 1650
• GTX 1650 Super
• GTX 1060
• GTX ¹⁰⁷⁰⁄₁₆₆₀ / 1660 Super / 1660 Ti
• GTX 1080 / RX 2060
• GTX 1080Ti / RTX 2070

The bet­ter cards (towards the bot­tom) tend to be bey­ond the price range. I can see a GTX 1070 on ebay for around £230 which is more than a new GTX 1660 so second hand options don’t appear to help. There are also 6Gb GTX 1060’s for around £170 on ebay (3Gb isn’t enough, 4Gb is a min­im­um really so be care­ful with these) — but are they really £20 bet­ter than a new GTX 1650 Super?

It sounds like the AMD Polar­is gen­er­a­tion cards wer­en’t good per­formers in mad­vr. This has appar­ently changed with Navi, how­ever there have been driver issues with Navi so most people still recom­mend nVidia cards for mad­vr. On top of this it is rumoured that future ver­sions of mad­vr may make use of the nVidia tur­ing archi­tec­ture, but these rumours are at least a year old with noth­ing to show so far. If they ever come true it may be that AMD’s “ray-tra­cing” com­pat­ible cards may be just as well suited too.

I have a sus­pi­cion that the RX 5500 XT may be a good value deal as it may offer GTX 1660 like per­form­ance in mad­vr, for £35 less, but I can­’t find con­crete evid­ence or a dir­ect comparison.