2Media Player Classic - Home Cinema LogoComparing 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. How­ever, 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 10701660 / 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.

Think we've missed something? Let us know by commenting below. If you would like to subscribe please use the subscribe link on the menu at the top right. You can also share this with your friends by using the social links below. Cheers.

Leave a Reply

2 Comments

gravatarTerry

Does­n’t the 1070 have 6+ tflops where­as the 1660 has only 5 tflops? Would­n’t this make the 1070 bet­ter for mad­vr, espe­cially when the 1070 comes with 8gb of memory?

Reply
gravatarJon Scaife

Pos­sibly so, altho my guess is that the dif­fer­ence isn’t big enough to enable any addi­tion­al set­tings in mad­vr so may be irrel­ev­ant. The tflops score does­n’t appear to per­fectly pre­dict mad­vr per­form­ance. The only way to know for sure would be to test both but I haven’t found any­where doing dir­ect com­par­is­ons of cards in madvr.

Reply