Ray Tracing on Nvidia, AMD and Intel. What's the differences?

you have probably heard that Ray tracing on different PC gpus and consoles Works in slightly different ways some even say that RT on AMD is not a real RT or something like that of course all those statements have nothing to do with reality but different approaches to RT implementation can vary in terms of their efficiency and specific tasks this video contains a brief explanation of those differences and a bit of testing of Ray tracing performance on Nvidia AMG and Intel gpus hello and welcome back to Hardware lab children [Music] speaking Ray tracing is a method of visualization of an image in Ray tracing Rey is nothing else than the vector form of data that is traced from the viewer's camera perspective and location this Ray intersects three dimensions geometry considering physical properties and color of that geometry the ray can bounce and continue bouncing multiple times so all the polygons in Geometry that Ray bounces off or not bounces off should be calculated this is called array per triangle intersection calculation but just imagine how expensive calculating the rate per triangle intersection could be in modern games with a lot of insanely detailed geometry real-time RT requires some clever acceleration for that purpose and this is where the bounding volume hierarchy concept plays its role yes we were speaking about bvh in the previous video but we're about to bring quickly repeat this explanation here the best way to understand the bvh concept is to think about it as a tree of possible rays and triangles intersection test outcomes all geometry of a scene is found inside a bvh structure BBH is a tree where the scene is apparent or a root and objects are children or leaves each object in term is apparent for its own nested items which are boxes or triangles the main idea of bvh is testing intersection with boxes first rather than triangles to speed up intersection calculations the bvh algorithm first traces an intersection with the main box of an object and if that happens the algorithm continues testing the nested boxes inside the main box after a hit with one of these boxes the ray is traced into the nested boxes of this box and so on until the ray reaches a triangle or a real polygon of a 3D model the boxes which haven't been shot are not taken into account these calculations are simply called box intersection calculations real-time Ray tracing in most cases is done via either Vulcan RT or dxr library that already include instructions for building acceleration structures and shooting in tracing Rays talking about actual numbers for this test we have three different mid-range gpus Intel Arc a770 GeForce RTX 3060 and Radeon RX 6700 XT yeah 6700 XT is definitely not a direct competitor for the other two but 6600 XT in this test would be not a good showcase for amd's RT approach since its vram subsystem is a huge limiting factor while 6700 XT is a well-balanced card it is also worth mentioning that we are looking at the performance disadvantage of tested cards with turn on rtfx not at the actual FPS alright numbers Intel are gray accelerator can calculate 12 race per box intersections per cycle and one Ray intersection per triangle per cycle as for NVIDIA unpair there is no data about intersections per box but it's at least 12 as well and one triangle intersection talking about AMD their approach provides only four intersections per box per cycle and as well as the others one intersection per triangle the reason for this is that BBH structural calculations on AMD rely hardly on regular shaders rather than on array accelerator unlike Nvidia and Intel vvh gigres per second amount doesn't look particularly good on AMD side with even more budget cars from competitors over performing RX 6700 XT while the triangle intersections look weak on Nvidia side due to still low overall frequency and RT blocks amount Intel arc on the other hand combines both high frequency and C use amount with a high number of BBH structures calculations moreover these Peak speeds can often be achievable only when the RT tasks are sorted properly and Intel Arc gpus have a special thread sorting unit and async RT system you can learn more about that from our Intel Arc a770 full review by following the link in the description the game list for this video includes some representative games that will cover a lot of different RT usage scenarios and test RT subsystems of our gpus from different perspectives this table shows the RT effects implemented in those games but let's look closer match your ex's enhanced Edition features a fully Dynamic high quality Ray tracing Global illumination system this game is fully based on RTR doing and cannot be run on a non-rt GPU so in order to test the performance impact of Ray tracing in Metro the original game's non-rt results will be compared with enhanced version results Spider-Man Miles Morales was firstly released on PS5 so the game's RT implementation was done with amd's rt approach in Mind Ray tracing in Spider-Man features some really huge distance of object rendering in RT Reflections and also features the high quality RT Shadows the Calista protocol is one of the most recent games with rt the list of protocols rendering not only includes common RT Reflections and shadows but also a unique effect Ray tracing transmission that improves enemy models and their transparent Parts appearance Forza Horizon 5 doesn't feature a complete game-changing rtfx but it has some nicely done RT car self-reflections improving the overall vehicle models look with retracing turn off these Reflections can be weird and improperly drawn Bridge breaker is a game where around a half of a frame rendering time is spent on Ray tracing when it's turned on the game artwork and dynamic light system is based around using local and Global RT Shadows so when RT is turned off the game looks faded in addition the game also has an RTA o that produces the most accurate invent occlusion effect cyberfunk 2077 has most of Ray tracing effects available in high quality it has both local and Global RT Shadows raytracing GI system and high quality RT Reflections so the game relies a lot on RT Computing and provides one of the best Ray tracing experiences available however cyberpunk is one of the earliest RT titles that was released even before rdna 2 appeared on the market so compatibility with IMG and of course Intel can be questionable Quake 2 RTX was one of the tech demo projects for RTX cards from Nvidia it features fully packed traced rendering combined with assets from a very old original Quake too but reward for PBR anyway it means that all non-rt computations in Quake 2 take hundreds times less time of completion so basically all rendering time is spent on RT here in terms of Quake tests we won't compare it with turn off RT and just show you the RT results for obvious reasons finally the last test will be done in amd's hybrid reflection sample in the case of this sample we're about to compare the rendering with just a pure SSR component of hybrid Reflections and the rendering with full usage of hybrid RT Reflections technology with a brief understanding of the effects that we are about to test we can find finally move on to the test segment right after telling you about our sponsor's amazing deals keysfan.com is an online store when you can purchase a cheap Windows 10 Microsoft Office and a bunch of other keys for a reasonable price Keys fan sells a lot of Windows and Mac OS software including creativity software anti-virus licenses Advanced download tools and game keys for PC and consoles moreover Keys fan holds a Christmas sale now with a lot of amazing discounts up to 66 percent there are three coupon codes that can be activated during the Christmas sale mt66 for a 66 discount on Microsoft Office which delivers you a price of 22 US dollars for office 2090 Pro and 24 bucks for office 2021 Pro mt50 for a half price for Windows 10 home and professional it allows you to purchase Windows 10 key for 7.59 US Dollars and windows 11 for just 13 bucks and finally nt62 provides 62 discount for Windows plus office bundles starting at just 26 and a half dollars check out Keys fan amazing deals using the link in the description and let's move to the test segment [Music] foreign [Music] Exodus shows some insane results in terms of a770s performance it's faster than the more expensive 6700 XT in both enhanced RT and non-rt modes even despite the sensor on a770 that reports about under load talking about the performance drop by using Ray tracing a770 shows the smallest drop around 14 and amd's car shows the highest around 25 percent RTX 3060 holds in the Middle with 19 to 20 percent loss but the overall performance is kinda poor foreign [Music] foreign [Music] trace and performance impact looks very soft on both Intel and Nvidia sites around seven percent as for AMG there is a 20 impact and the more expensive RX 6700 XT with turned on RT shows the on par results with cheaper Intel and nvidia's cards foreign [Music] foreign [Music] protocol shows some weird behavior on all of the cards the last video game update makes it possible to run RT on Intel Arc but something's definitely going wrong here I wouldn't call it runs actually it looks more like crawling than running RTX 3060 doesn't surprise here at all the games at times look like CPU limited however 6700 XT with this same CPU provides much better results FPS drop calculations in our tests are based not on the lowest but on the average frame rate and it's obvious that Crystal protocols 3 tracing prefers amd's approach with 22 performance drop instead of 26 on the Nvidia side [Music] [Applause] [Music] thank you [Applause] [Music] [Applause] [Music] [Applause] long Forza Horizon 5 doesn't feature any really demanding RT effects so overall performance punishment for turning on RT is not that big however although Reflections in force include just a small amount of geometry shading in those Reflections is done quite well so the demand on proper sorting is huge this is probably why RT enforcer works better on the arc than on the RTX in fact even radium's performance drop here is smaller than g-forces however it is that small in general that it can be explained with measurement uncertainty [Applause] foreign [Music] the performance drop in Reef breaker is the smallest on AMD side however 6700 XT is the only GPU that is slightly CPU Limited in this test without RT so it can be connected to that anyway RT is working great here on AMD surprisingly Intel arcs dropped from rtfx in this game is higher than g-force's one foreign [Music] [Music] [Music] Tel's GPU shows some weird behavior in one more game cyberpunk 2077 the sensor reports on the load despite all the other cards running on higher frame rate are completely GPU limited this is normal behavior for a770 in some games though the performance drop from turning on rtfx is the smallest of all three on the arc side 30 percent but the overall performance is not as impressive RTX drop is the more significant around 45 and rdna 2 gpus don't like cyberpunk's RT implementation at all losing a massive 63 percent of frame rate more expensive RX 6700 XT is rated here under RTX 3060 with rt foreign [Music] RTX Works significantly better on Nvidia than on both AMG and Intel the cheapest of all cards today are tx3060 easily outperforms both a770 and 6700 XT Does it show the Pure retracing Performance and mean that Nvidia is just better than that I wouldn't really say so but the result is interesting and quite unexpected foreign [Music] demo AMD showed the worst result in terms of turning on our tfx and Nvidia the best 43 35 and 21 loss respectively overall 360s performance is not really competitive but we are still looking at percentage differences and this demo definitely prefers nvidia's approach foreign [Music] there are a few obvious conclusions that we can make based on the tests above first we can't be wrong about dvh calculations on the Nvidia side it can be even higher than 12 but remember these are just assumptions there is no data about BBH box Computing on Nvidia second the difference in RT processing on AMD Nvidia in Intel are quite huge so we cannot say that one is better than another Intel's approach looks pretty decent while AMD cards are likely to lose more performance when retracing is turned on but it always depends on the games RT implementation and you cannot be sure what to expect just from the raw or t performance anyway this is it for this video Merry Christmas Hardware enthusiasts and as usual stay up to date with HL [Music] foreign [Music] [Applause] [Music] [Music] [Applause] [Music]

Channel: Hardware Lab

Views: 17,549

Keywords:

Id: kANMoO-TvDs

Length: 21min 49sec (1309 seconds)

Published: Thu Dec 29 2022