RTX 5090 and 5080 leaks! RTX 50 Neural Rendering?!? Strix Halo| Nvidia App issues confirmed | More!!

00:31:47
https://www.youtube.com/watch?v=EM79XC4RtpQ

Resumo

TLDRVideoen fokuserer på flere teknologinyheder, primært inden for gaming og hardware. Det inkluderer forventninger om Nvidia's kommende præsentation af RTX 50-serien ved CES, hvor neural rendering i videospil bliver et stort emne. Neural rendering vil muligvis revolutionere den måde, grafik behandles og vises på. Acer har ved en fejl offentliggjort en gaming-computer konfigureret med RTX 5090 og 5080 grafikkort, som bekræfter specifikationer som 32 GB VRAM for 5090 og 16 GB for 5080. Der er også diskussion af ydeevneproblemer med Nvidia's app, hvor brugere rapporterer reduceret spilpræstation. AMD's Strix Halo-serie og dens potentiale inden for integrerede grafikenheder bliver også nævnt, sammen med enkelte software-opdateringer og spilnyheder.

Conclusões

  • 💻 RTX 50-serien afsløres muligvis ved CES med fokus på neural rendering.
  • 📊 Acer har ved en fejl offentliggjort specifikationer for kommende gaming PC'er.
  • 🕹 Nvidia undersøger ydeevneproblemer i deres app.
  • ⚡ Ny GDDR7 VRAM forventes i RTX 5080 og 5090.
  • 🔍 Nvidia avancerer med neural rendering og AI integration i grafik.
  • 🚀 AMD's Strix Halo APU ser lovende ud for integreret grafik.
  • 🎮 Softwareopdateringer forbedrer spiloplevelser og ydelse.
  • 🔧 Intel Arc B570 review enheder er begyndt at blive sendt ud.
  • 🎲 Witcher 4 projekt bekræfter inklusion af Gwent.
  • 📈 Benchmark leak afslører potentielle evner af AMD's kommende chips.

Linha do tempo

  • 00:00:00 - 00:05:00

    Videoen diskuterer forventede annonceringer ved CES i januar, herunder Nvidia RTX 50-serien med fokus på neural rendering i videospil. Andre emner inkluderer hardwarelækager fra Acer og Nvidia's undersøgelser af reduceret ydeevne ved brug af deres app. AMDs nye APU, Strix Halo, og Intels B570 omtales også.

  • 00:05:00 - 00:10:00

    Fortsætter med at diskutere neural rendering og avanceret DLSS teknologi. Indlæg fra Inno3D forventes at fokusere på revolutionerende grafiske processer, AI-accelereret grafik og forbedret AI i spil- og indholdsskabelsesarbejdsbyrder, herunder kraftværdi AI-assisterede opgaver.

  • 00:10:00 - 00:15:00

    Der diskuteres Nvidia's neural kompressionsteknik for materialeteksturer, der muliggør høj opløsning uden øget lager og hukommelsesefterspørgsel. Videoen refererer også til forskning på dette område fra Siggraph 2023 og 2024, der muliggør brug af filmkvalitetsvisuelle billeder i realtid applikationer.

  • 00:15:00 - 00:20:00

    Nye teknikker og systemer til neural rendering introduceres, herunder realtids neural appearance models, der muliggør hurtigere og mere effektive renderingmetoder. Potentialet i at integrere disse teknologier i videospil og den fremtidige anvendelse af CES præsenteres også.

  • 00:20:00 - 00:25:00

    Acer lækker information om RTX 5090 og 5080 med henholdsvis 32 GB og 16 GB VRAM. Bekymringer om VRAM-grænsen for de kommende GPU'er diskuteres, især med hensyn til fremtidssikring og prispunkt. Leaks om hukommelseshastigheder og specifikationer for den nye RTX-serie nævnes også.

  • 00:25:00 - 00:31:47

    Hardware unboxed tester Nvidias app-relaterede performance problemer og AMD's Strix Halo APU benchmarks. Tekniske detaljer om nye chip egenskaber og potentiale i gaming ydeevne samt en notits om opdateringer af et spil diskuteres, samtidigt med at videoen afsluttes med information om Witcher 4.

Mostrar mais

Mapa mental

Vídeo de perguntas e respostas

  • Hvad forventes der præsenteret ved CES?

    Ved CES forventes der præsentation af RTX 50-serien samt teknologier relateret til neural rendering af videospil.

  • Hvilken information er lækket fra Acer?

    Acer har utilsigtet listet en gaming-PC udstyret med RTX 5090 og 5080 grafikkort.

  • Hvad er neural rendering?

    Neural rendering angår revolutionerende metoder til at processere og vise grafik ved brug af AI-teknologier.

  • Hvad er RTX 5080's forventede specifikationer?

    RTX 5080 er forventet at have 16 GB GDDR7 VRAM.

  • Hvilke problemer er rapporteret med Nvidia's app?

    Nvidia's app reducerer ydeevnen i spil og det er rapporteret, at de arbejder på en løsning.

Ver mais resumos de vídeos

Obtenha acesso instantâneo a resumos gratuitos de vídeos do YouTube com tecnologia de IA!
Legendas
en
Rolagem automática:
  • 00:00:00
    all signs point to CES in January 6 not
  • 00:00:03
    only showing us the RTX 50 Series for
  • 00:00:06
    the first time but also the era of
  • 00:00:09
    neural rendering of video games more on
  • 00:00:13
    that and where am I getting this
  • 00:00:15
    demonstration of it in just a minute uh
  • 00:00:18
    but we have a bunch of other news to
  • 00:00:19
    talk about today including Acer
  • 00:00:22
    accidentally listing a 55090 and 80
  • 00:00:25
    equipped gaming PC ahead of time
  • 00:00:27
    confirming only 16 GB of vram
  • 00:00:30
    on the RTX 580 although it's gddr 7 and
  • 00:00:33
    how fast we have some leaks discussing
  • 00:00:36
    that as well as we have Nvidia saying
  • 00:00:39
    they are investigating the reduced
  • 00:00:41
    performance uh by using their Nvidia app
  • 00:00:45
    and there's some uh Hardware unbox
  • 00:00:47
    testing discussing exactly what in the
  • 00:00:50
    app is causing the issues and confirming
  • 00:00:53
    those issues we also have some exciting
  • 00:00:56
    uh leaks regarding stricks Halo this is
  • 00:00:59
    an amazing amazing looking Apu from AMD
  • 00:01:02
    and then we also have apparently Intel's
  • 00:01:05
    b570 which if you're excited about the
  • 00:01:07
    b580 with 12 GB hey 10 GB on the
  • 00:01:11
    b570 Apparently review samples already
  • 00:01:14
    out to at least one review Outlet uh for
  • 00:01:17
    that
  • 00:01:18
    b570 uh we also have a little bit more
  • 00:01:21
    gaming information towards the end a
  • 00:01:24
    whole bunch more about neural rendering
  • 00:01:25
    in just a second but first unfortunately
  • 00:01:28
    I have to admit that I've been spreading
  • 00:01:30
    misinformation in the comment section or
  • 00:01:32
    at least information I spread is now out
  • 00:01:34
    of date Mr Daniel do you sleep no has
  • 00:01:37
    been upgraded to yes because I am now
  • 00:01:40
    proudly Pro naap honestly guys I moved
  • 00:01:43
    closer to my uh job teaching cuz I also
  • 00:01:46
    do that on top of running the YouTube
  • 00:01:47
    channel fulltime people are like how are
  • 00:01:48
    you doing this power naaps in the
  • 00:01:51
    afternoon are now possible I get home
  • 00:01:53
    from work quick get in a power nap
  • 00:01:54
    before my kids get home and it's been
  • 00:01:56
    amazing now in the past I couldn't nap I
  • 00:02:00
    am now completely blacked out cannot see
  • 00:02:03
    a bit of light and also I can kick on
  • 00:02:06
    some white noise right here with
  • 00:02:08
    completely flat Bluetooth speakers which
  • 00:02:10
    is amazing uh because I'm a sid sleeper
  • 00:02:13
    and headphones in the past have been
  • 00:02:15
    incredibly painful making it difficult
  • 00:02:18
    for me to nap this is also really nice
  • 00:02:20
    when I go to sleep at night a lot of
  • 00:02:21
    times I like to put on like an Audi book
  • 00:02:22
    or podcast and H I can't tell where any
  • 00:02:25
    of you are this is completely blacked
  • 00:02:26
    out anyway I'm take that back off uh but
  • 00:02:28
    the point is um that this has completely
  • 00:02:31
    revolutionized both my actual sleep at
  • 00:02:34
    night as well as my ability to actually
  • 00:02:37
    get in some afternoon naps which is uh
  • 00:02:40
    re really amazing really it has made me
  • 00:02:42
    so much more alert and productive there
  • 00:02:44
    is a wide variety available again this
  • 00:02:47
    is the uh sound version uh which I think
  • 00:02:50
    is really cool but they also have the uh
  • 00:02:52
    pro version they have the silk mask uh
  • 00:02:55
    Manta weighted mask we've got a cool
  • 00:02:58
    mask a steam mask mask uh the original
  • 00:03:01
    Manta sleep mask and a huge thank you to
  • 00:03:03
    Manta sleep for sponsoring today's video
  • 00:03:06
    and allowing me to be proudly prap and
  • 00:03:08
    if you check out the link in the video
  • 00:03:09
    description you can get some amazing
  • 00:03:11
    deals right now on holiday gifts there's
  • 00:03:14
    a 15% off sitewide sale plus a free
  • 00:03:18
    mystery sleep kit uh for those who miss
  • 00:03:21
    the Cyber weekend this is the perfect
  • 00:03:22
    opportunity to grab a great deal on
  • 00:03:24
    sleep accessories for yourself or gifts
  • 00:03:26
    can't emphasize this enough these make
  • 00:03:28
    amazing gifts
  • 00:03:30
    uh and again a mystery sleep kit
  • 00:03:32
    included with the 15% off sitewide sale
  • 00:03:35
    if you click the link in the video
  • 00:03:36
    description so what do I mean about
  • 00:03:39
    neural rendering of video games where am
  • 00:03:41
    I getting this footage and why do I
  • 00:03:43
    think this is very likely to be a big
  • 00:03:45
    Topic at CES with the 50 Series well
  • 00:03:50
    let's dive into my journey down this
  • 00:03:54
    pathway today uh first of all Hardware
  • 00:03:56
    Lux has a article this is a transl from
  • 00:04:00
    German talking about Ino 3D's
  • 00:04:04
    announcement for what they will be
  • 00:04:05
    covering at uh CES now within that post
  • 00:04:10
    if you actually follow up the actual
  • 00:04:11
    in3d post uh they talk about their
  • 00:04:14
    coolers and whatnot but then we get this
  • 00:04:18
    this is where we get into neural
  • 00:04:20
    rendering there's a lot going on here uh
  • 00:04:23
    one of them is specifically neural
  • 00:04:26
    rendering capabilities revolutionizing
  • 00:04:29
    how Graphics are processed and displayed
  • 00:04:33
    so that is one of the highlights for Ino
  • 00:04:36
    3D's C presentation talking about their
  • 00:04:40
    products related to gpus now it's very
  • 00:04:43
    likely that board partners that work
  • 00:04:45
    closely with Nvidia are fully aware of
  • 00:04:48
    what nvidia's big CES talking points are
  • 00:04:51
    going to be and have presentations to
  • 00:04:53
    follow up and expand on how their
  • 00:04:55
    products relate to these so there's the
  • 00:04:58
    neural rendering cap capabilities being
  • 00:05:00
    referenced right here uh along with this
  • 00:05:04
    we also have discussion of advanced dlss
  • 00:05:07
    technology now it's possible that
  • 00:05:10
    they're just talking about dlss
  • 00:05:12
    technology that already exists and the
  • 00:05:14
    fact that it is Advanced it's also
  • 00:05:17
    possible that this is referring to a
  • 00:05:20
    further advancement of dlss Technology
  • 00:05:23
    although it's still just discussing uh
  • 00:05:25
    even better image quality and higher
  • 00:05:27
    frame rates so potentially this is um
  • 00:05:30
    you know expansion of what's kind of
  • 00:05:32
    already out there but it doesn't sound
  • 00:05:33
    like an entirely new thing um that is
  • 00:05:37
    like somehow different than the dlss
  • 00:05:40
    that we already have at a fundamental
  • 00:05:41
    level they also talk about enhanced rate
  • 00:05:43
    tracing improved RT cores um AI
  • 00:05:47
    accelerated Graphics now this is talking
  • 00:05:49
    about improved performance in AI
  • 00:05:50
    assisted tasks and better integration of
  • 00:05:52
    AI in gaming and content creation
  • 00:05:54
    workloads okay but then then we get into
  • 00:05:56
    that neural rendering capabilities
  • 00:05:58
    revolutionizing how Graphics are
  • 00:06:00
    processed and displayed uh they're
  • 00:06:02
    talking about AI enhanced power
  • 00:06:03
    efficiency whether that's something that
  • 00:06:05
    in3d is doing themselves or that'll be a
  • 00:06:08
    wider scale thing for the 50 Series in
  • 00:06:10
    general is unknown at this point uh then
  • 00:06:12
    we have improved Aid driven upscaling
  • 00:06:15
    however this is talking about Beyond
  • 00:06:17
    gaming so this is more for Content
  • 00:06:18
    creators and they also have generative
  • 00:06:21
    AI acceleration so at this point the
  • 00:06:24
    biggest one here that stands out as
  • 00:06:26
    something that is potentially new is
  • 00:06:29
    neural rendering
  • 00:06:31
    capabilities so I know that first of all
  • 00:06:34
    that rung a bell talking about some
  • 00:06:36
    Nvidia white papers that I've reported
  • 00:06:38
    on in the past uh from sigraph
  • 00:06:41
    presentations so we've already talked
  • 00:06:44
    about um uh neural texture compression
  • 00:06:49
    in previous videos at sigraph
  • 00:06:52
    2023 Nvidia presented research on the
  • 00:06:55
    neural compression of
  • 00:06:57
    materials however I went in to see if
  • 00:07:00
    they had any newer presentations related
  • 00:07:03
    to neural rendering at the 2024 sigraph
  • 00:07:08
    and they do we have realtime neural
  • 00:07:11
    appearance models and that's the video
  • 00:07:15
    that I was showing you here this is from
  • 00:07:17
    their materials available uh for the
  • 00:07:20
    2024 sigraph presentation on neural
  • 00:07:24
    materials you can see the diagonal bar
  • 00:07:27
    across the screen here where where this
  • 00:07:29
    is the ground truth reference model in
  • 00:07:31
    the bottom right corner and the top left
  • 00:07:34
    corner is the neural material pathway
  • 00:07:37
    now this just switched to the 2x 32
  • 00:07:40
    neurons Which is less accurate than the
  • 00:07:42
    2x 64 that we saw
  • 00:07:45
    initially uh but you can see that even
  • 00:07:48
    at this level there's uh quite a lot of
  • 00:07:51
    accuracy here also it's important to
  • 00:07:54
    point out that this does not mean I I'm
  • 00:07:57
    sure you guys have seen videos talking
  • 00:07:59
    about entirely AI generated games where
  • 00:08:03
    you're literally playing a game that is
  • 00:08:05
    basically being hallucinated by an AI
  • 00:08:07
    and that is not what I think we're going
  • 00:08:09
    to be getting from Nvidia when we look
  • 00:08:12
    at neural rendering capabilities I think
  • 00:08:14
    it's more along the lines of uh these
  • 00:08:17
    neural compressions that we've seen in
  • 00:08:20
    uh previous
  • 00:08:21
    videos as well as potentially uh this
  • 00:08:25
    neural appearance models so what is
  • 00:08:27
    going on here I have fully read the the
  • 00:08:30
    available information here and to be
  • 00:08:32
    honest a lot of this is rather indepth
  • 00:08:36
    regarding the actual implementation of
  • 00:08:38
    how this works uh the abstracts
  • 00:08:42
    themselves provide a bit more of a kind
  • 00:08:45
    of summary of what is going on here that
  • 00:08:49
    perhaps we can delve into a bit so again
  • 00:08:51
    the big neural uh uh rendering stuff
  • 00:08:54
    that I could find research on is these
  • 00:08:56
    neural materials sorry neur neural
  • 00:08:59
    appearance models and then neural
  • 00:09:01
    compression so neural compression uh of
  • 00:09:05
    material textures seems a little simpler
  • 00:09:07
    to wrap your head around uh the idea
  • 00:09:10
    here and and I've I've again I've shown
  • 00:09:12
    this in uh videos you know back after
  • 00:09:14
    the 2023
  • 00:09:15
    sigraph uh is that the textures on the
  • 00:09:18
    left here are uh neural textures on the
  • 00:09:21
    right is BC uh bcx textures that's a
  • 00:09:24
    more uh traditional method of doing this
  • 00:09:27
    and you can see that these are similar
  • 00:09:29
    texture size at 3.6 like storage size
  • 00:09:31
    3.6 megabytes on the left versus 3.33 on
  • 00:09:34
    the right however I play the video uh
  • 00:09:36
    you can see as it scans through here on
  • 00:09:39
    the left you have massively more detail
  • 00:09:42
    at just a slightly larger uh file size
  • 00:09:45
    that's cool however there's a major
  • 00:09:48
    downside to this which is the uh 1.15
  • 00:09:52
    milliseconds to uh to do this on an RTX
  • 00:09:55
    4090 GPU at Native 4K whereas this one
  • 00:09:58
    was 0. for 9 milliseconds in the video
  • 00:10:01
    if you listen to it they describe uh you
  • 00:10:03
    know rounding off as about uh 1
  • 00:10:05
    millisecond of overhead now a
  • 00:10:08
    millisecond doesn't sound like a lot but
  • 00:10:09
    in terms of uh if you're rendering
  • 00:10:11
    frames at a high frame rate uh every
  • 00:10:14
    millisecond counts however they also
  • 00:10:17
    describe in this process that in a if
  • 00:10:20
    you're rendering a larger scene
  • 00:10:22
    including using like Ray tracing and
  • 00:10:24
    path tracing uh there there's you know
  • 00:10:27
    the compute time that's happening for
  • 00:10:28
    those other processes so during that the
  • 00:10:32
    rest of the computation of the frame you
  • 00:10:35
    know this extra 1 millisecond can kind
  • 00:10:37
    of be hidden by happening in parallel to
  • 00:10:40
    those if I'm understanding it correctly
  • 00:10:42
    meaning it might not actually slow you
  • 00:10:44
    down a millisecond for your frame times
  • 00:10:47
    uh because other things would be taking
  • 00:10:49
    at least a millisecond as well and this
  • 00:10:51
    could kind of be computed at the same
  • 00:10:53
    time as some other things like the ray
  • 00:10:55
    tracing for example that's at least how
  • 00:10:58
    I understand this to be clear I am not
  • 00:11:01
    like like like I'm I'm a little out of
  • 00:11:03
    my depth reading papers like this like I
  • 00:11:05
    have a math degree but not specifically
  • 00:11:07
    in like computer graphics and this style
  • 00:11:09
    of stuff anyway uh so that's that so the
  • 00:11:13
    abstract for neural uh textures uh
  • 00:11:16
    neural neural compression of textures is
  • 00:11:18
    the continuous advancement of photo
  • 00:11:20
    realism and rendering is accompanied by
  • 00:11:21
    a growth in texture data and
  • 00:11:24
    consequently increasing storage and
  • 00:11:25
    memory demands so the idea here is to
  • 00:11:28
    have higher resolutions without
  • 00:11:30
    increasing storage and memory demands so
  • 00:11:33
    to address this issue we propose a novel
  • 00:11:35
    neural compression technique
  • 00:11:36
    specifically designed for material
  • 00:11:38
    textures we unlock two more levels of
  • 00:11:40
    detail IE 16 times more texal using low
  • 00:11:43
    bit rate compression with image quality
  • 00:11:46
    that is better than Advanced image
  • 00:11:47
    compression techniques such as avif and
  • 00:11:49
    jpeg XL at the same time our method
  • 00:11:52
    allows on demand realtime decompression
  • 00:11:55
    with random access similar to block
  • 00:11:58
    texture compression on gpus enabling
  • 00:12:01
    compression on disk and memory the key
  • 00:12:03
    idea behind our approach is compressing
  • 00:12:05
    multiple material textures and their M
  • 00:12:08
    map chains together and using a small
  • 00:12:10
    neural network that is optimized for
  • 00:12:12
    each material to decompress them finally
  • 00:12:15
    we use a custom Training implementation
  • 00:12:17
    to achieve practical compression speeds
  • 00:12:20
    whose performance surpasses that of
  • 00:12:22
    General Frame Works like Pi torch and uh
  • 00:12:25
    by an order of magnitude so that's a
  • 00:12:28
    general idea all my sources will be
  • 00:12:30
    linked in the video description so those
  • 00:12:31
    of you more technically minded uh could
  • 00:12:34
    dive into some of the additional details
  • 00:12:37
    about that now again that's from
  • 00:12:40
    2023 so uh not only has it given more
  • 00:12:43
    time to potentially get this into at
  • 00:12:45
    least one video game potentially uh
  • 00:12:48
    there's also like I said this update for
  • 00:12:50
    2024 I've seen real-time neural
  • 00:12:52
    appearance models so this seems a bit
  • 00:12:54
    more advanced uh so this is where I was
  • 00:12:57
    showing you that that you know uh
  • 00:12:59
    diagonal uh side by side at the
  • 00:13:01
    beginning here this is the neural
  • 00:13:02
    material idea so let's go ahead and read
  • 00:13:06
    at least the abstract of what's going on
  • 00:13:09
    here so they're saying we present a
  • 00:13:10
    complete system for realtime rendering
  • 00:13:13
    of scenes with complex appearance
  • 00:13:16
    previously reserved for offline use
  • 00:13:19
    meaning you couldn't do this in like
  • 00:13:20
    real-time rendering uh this is achieved
  • 00:13:23
    with a combination of algorithmic and
  • 00:13:26
    system level Innovations our appearance
  • 00:13:29
    model utilizes learned hierarchical
  • 00:13:31
    textures that are interpreted using
  • 00:13:34
    neural decoders which produce
  • 00:13:36
    reflectance values and important sample
  • 00:13:38
    directions to best utilize the modeling
  • 00:13:41
    capacity of the decoders we equip the
  • 00:13:43
    decoders with two Graphics priors the
  • 00:13:46
    first prior transformation of directions
  • 00:13:48
    into learned shading frames facilitates
  • 00:13:51
    accurate reconstruction of mesoscale
  • 00:13:53
    effects the second prior a microfacet
  • 00:13:56
    sampling distribution allows the neural
  • 00:13:58
    decoder to perform important sampling
  • 00:14:01
    efficiently the resulting appearance
  • 00:14:03
    model supports an I don't know how to
  • 00:14:05
    pronounce this word guys anisotropic
  • 00:14:07
    sampling I know I see it in graphic
  • 00:14:09
    settings menus but I don't always I I'm
  • 00:14:11
    not confident how to pronounce it anyway
  • 00:14:13
    and level of detail rendering and allows
  • 00:14:16
    baking deeply layered material graphs
  • 00:14:18
    into a compact unified neural
  • 00:14:22
    representation by exposing Hardware
  • 00:14:24
    accelerated tensor operations to Ray
  • 00:14:27
    tracing shaders we show that it is
  • 00:14:29
    possible to inline and execute the
  • 00:14:32
    neural decoders efficiently inside a
  • 00:14:35
    real-time path Tracer we analyze
  • 00:14:38
    scalability with increasing number of
  • 00:14:40
    neural materials and propose to improve
  • 00:14:43
    performance using Code optimized for Co
  • 00:14:46
    for coherent and Divergent execution our
  • 00:14:49
    neural material shaders can be over an
  • 00:14:52
    order of magnitude faster than non
  • 00:14:54
    neural layered materials this opens up
  • 00:14:57
    the door for using film qual quality
  • 00:14:59
    visuals in real-time applications such
  • 00:15:01
    as games and live previews so this if if
  • 00:15:05
    if a lot of this your eyes were glazing
  • 00:15:07
    over um I think in the end you're like
  • 00:15:09
    okay so what is the point of any of that
  • 00:15:11
    what am I supposed to take away so the
  • 00:15:12
    idea here is that the neural material
  • 00:15:15
    shaders can be can be right can be not
  • 00:15:17
    is always right over an order of
  • 00:15:19
    magnitude that's a lot faster than non
  • 00:15:23
    neural layered materials and that this
  • 00:15:26
    would open the door for using film
  • 00:15:28
    quality visuals in realtime applications
  • 00:15:30
    such as games and live previews um okay
  • 00:15:35
    so that's that now the idea here again
  • 00:15:38
    they they talk about the neural B brdf
  • 00:15:40
    model I don't think we need to try to
  • 00:15:42
    make sense out of this diagram because
  • 00:15:43
    you guys are probably not the ones
  • 00:15:45
    programming this stuff um but here's
  • 00:15:48
    kind of the idea here so so if we look
  • 00:15:51
    at this diagram we see a scene with four
  • 00:15:54
    materials that we approximate using the
  • 00:15:57
    proposed neural BRD F the first three
  • 00:16:00
    columns correspond to different
  • 00:16:02
    configurations of the brdf decoder from
  • 00:16:04
    fastest to the most accurate so it's uh
  • 00:16:08
    uh fastest on the left getting more
  • 00:16:10
    accurate right from going from 16 uh 16
  • 00:16:13
    neurons to 32 neurons to 64 neurons but
  • 00:16:17
    this time also going up to three layers
  • 00:16:19
    instead of two layers on the first two
  • 00:16:21
    then they have the ground truth
  • 00:16:23
    reference on the right hand side
  • 00:16:26
    um so uh then they have flip error
  • 00:16:30
    images in the corners timings quantify
  • 00:16:34
    the cost of rendering a one spp image of
  • 00:16:36
    the scene we show images with 8,192
  • 00:16:40
    samples to suppress path tracing Noise
  • 00:16:44
    Okay so in other words what does it mean
  • 00:16:45
    about the the the flip image so you guys
  • 00:16:47
    see in the in the corners here it might
  • 00:16:49
    be a little hard to see in this video by
  • 00:16:50
    the way if you're watching on a phone
  • 00:16:51
    you can pinch to zoom on the YouTube app
  • 00:16:53
    anyway uh you can see that so this is
  • 00:16:56
    basically like a a a map where the more
  • 00:16:59
    stuff is showing up here I think is
  • 00:17:00
    being uh showing up as error when
  • 00:17:03
    compared to the ground truth image so
  • 00:17:05
    you can see that the um the faster ones
  • 00:17:08
    have more error and there's less error
  • 00:17:11
    showing up on the more accurate ones you
  • 00:17:14
    can see the cost in milliseconds of
  • 00:17:17
    rendering time so here we have 3.15
  • 00:17:20
    milliseconds and then you get a lot more
  • 00:17:22
    accuracy or in other words a lot less
  • 00:17:24
    error uh going up to 32 neurons and only
  • 00:17:27
    going up to 3.71 Mill seconds compared
  • 00:17:29
    to
  • 00:17:30
    3.15 uh but then jumping up to 64
  • 00:17:32
    neurons with three layers increases all
  • 00:17:34
    the way up to 6.31 milliseconds however
  • 00:17:37
    their ground truth reference uh was a
  • 00:17:40
    total of 13.25
  • 00:17:42
    milliseconds so all of that sounds like
  • 00:17:45
    a lot of milliseconds to me so then the
  • 00:17:48
    idea is so so how are you able to get
  • 00:17:51
    this running uh you know you know fast
  • 00:17:53
    enough for a video game or live preview
  • 00:17:56
    which is what they mentioned uh so again
  • 00:17:58
    again if we go back to the abstract
  • 00:18:00
    where a lot of this might not have made
  • 00:18:02
    sense um the idea is is of how they're
  • 00:18:07
    getting it running is talking talked
  • 00:18:08
    about right here so by exposing Hardware
  • 00:18:12
    accelerated tensor operations to rracing
  • 00:18:15
    shaders we show that it is possible to
  • 00:18:19
    inline and execute the neural decoders
  • 00:18:22
    efficiently inside a realtime path
  • 00:18:26
    Tracer so again while a lot of this is
  • 00:18:29
    also over my head so you're not the only
  • 00:18:31
    one a little confused here as I'm trying
  • 00:18:33
    to read that what I'm kind of hearing is
  • 00:18:37
    that this can happen so if you're doing
  • 00:18:38
    this inside of a path tracing situation
  • 00:18:41
    again the path tracing takes time to
  • 00:18:43
    compute as well so in other words you
  • 00:18:46
    wouldn't necessarily by running it in
  • 00:18:48
    line uh right to show that it is
  • 00:18:50
    possible to in line and execute the
  • 00:18:52
    neural uh uh decoders efficiently inside
  • 00:18:55
    the real-time path Tracer I'm reading
  • 00:18:57
    that as you don't
  • 00:18:59
    add these these frame time costs to the
  • 00:19:03
    uh the rest of the frame just as a sum
  • 00:19:05
    right if some of this is happening in
  • 00:19:07
    line I think that means like at the same
  • 00:19:09
    time so the path tracing itself has a
  • 00:19:12
    frame time cost and perhaps and this has
  • 00:19:15
    a frame time cost but the cost of doing
  • 00:19:17
    both together isn't necessarily their
  • 00:19:20
    sum if it's happening at the same time
  • 00:19:23
    and not necessarily um competing for the
  • 00:19:26
    same resources on the GPU if if that
  • 00:19:29
    makes any sense so in other words uh
  • 00:19:33
    whether or not this is actually what
  • 00:19:35
    in3d is referencing uh on their CES
  • 00:19:39
    presentation here this is definitely
  • 00:19:42
    research that Nvidia is doing so I don't
  • 00:19:46
    know if this is coming to games in 2025
  • 00:19:48
    but I have a feeling we are going to see
  • 00:19:50
    discussion of these types of neural
  • 00:19:52
    rendering techniques at CES
  • 00:19:55
    2025 and that's why it's showing up here
  • 00:19:57
    as neural rendering capabilities
  • 00:19:59
    revolutionizing how Graphics are
  • 00:20:01
    processed and displayed on uh on this
  • 00:20:04
    post here okay now let's go ahead and
  • 00:20:09
    move along to some additional leaks so
  • 00:20:12
    videoc
  • 00:20:13
    cards.com caught Acer posting their
  • 00:20:17
    Predator origin 7,000 series gaming PCs
  • 00:20:21
    with their 5090 and 80 listed so if
  • 00:20:25
    there was any doubt now that the 5090
  • 00:20:28
    would have 30 2 GB of gddr 7 memory Acer
  • 00:20:31
    should have now just completely removed
  • 00:20:34
    any doubt also if there was any doubt
  • 00:20:37
    that the 580 would have only 16 GB of
  • 00:20:41
    vram and it is gddr 7 again Acer has
  • 00:20:45
    removed any doubt so it's basically a I
  • 00:20:49
    would consider it pretty much 100%
  • 00:20:51
    confirmation maybe we'll call it 99% I
  • 00:20:54
    guess Acer could have just put up some
  • 00:20:56
    placeholder text here but that doesn't
  • 00:20:58
    seem like what this is right um they're
  • 00:21:00
    getting their products ready to unveil
  • 00:21:02
    for CES it's on their website goodjob
  • 00:21:04
    video cards.com for catching them all
  • 00:21:07
    the sources will be in the video
  • 00:21:08
    description so a 80 with 16 gbt what do
  • 00:21:11
    you guys think about that like I I've
  • 00:21:13
    seen a lot of comments on my channel uh
  • 00:21:16
    pretty upset about the idea of a 580
  • 00:21:18
    with 16 GB personally I'm more upset
  • 00:21:20
    that the 470 is rumored to have 12 and
  • 00:21:24
    that the 4060 is rumored to have eight
  • 00:21:27
    because eight is definitely a problem
  • 00:21:29
    for rendering maximum texture settings
  • 00:21:32
    in a lot of games right now today and 12
  • 00:21:35
    is fairly borderline and there are some
  • 00:21:38
    examples of games spilling over that now
  • 00:21:42
    16 GB on the other hand in realistic
  • 00:21:46
    gaming situations that don't involve a
  • 00:21:48
    whole bunch of modding and a b or other
  • 00:21:50
    crazy stuff I haven't seen 16 GB being
  • 00:21:53
    an issue however a 580 is a very
  • 00:21:58
    powerful graphics card that could
  • 00:21:59
    theoretically be used many years into
  • 00:22:02
    the future in which case uh I could see
  • 00:22:06
    it still being a viable GPU when we get
  • 00:22:09
    you know the next generation of consoles
  • 00:22:11
    even uh at which case if those consoles
  • 00:22:14
    come with an increased uh you know Ram
  • 00:22:15
    capacity compared to current gen you
  • 00:22:17
    know it's like vram usage tends to go up
  • 00:22:20
    over time and 16 gbes is still quite a
  • 00:22:23
    bit uh it's not immediately presenting
  • 00:22:26
    massive concerns to me however um the
  • 00:22:30
    longevity of a card this powerful uh
  • 00:22:32
    being potentially impacted by having 16
  • 00:22:35
    rather than something like 24 or 32 is a
  • 00:22:39
    bit frustrating especially when you when
  • 00:22:41
    you throw price into this right because
  • 00:22:44
    if the 5080 is at least
  • 00:22:46
    $11,000 or maybe even uh maybe uh even
  • 00:22:50
    $1,200 $1,300 may we don't know the
  • 00:22:53
    price right but depending on the price
  • 00:22:57
    you know it starts to feel like penny
  • 00:22:58
    pinching cuz could you just do a uh a
  • 00:23:01
    clamshell design and get 32 GB on here
  • 00:23:04
    for the sake of um you know longevity uh
  • 00:23:08
    and just uh okay if you're spending over
  • 00:23:10
    $1,000 here how much more does 16 gab of
  • 00:23:13
    vram cost uh in the grand scheme of
  • 00:23:15
    things like it feels like it could be on
  • 00:23:17
    there if Nvidia wanted it to be right
  • 00:23:21
    but they don't want it to be and this is
  • 00:23:22
    confirmation that the 80 only has 16 it
  • 00:23:26
    is gdr 7 though and how fast is it well
  • 00:23:29
    this is I would say a less solid leak uh
  • 00:23:33
    but this leak is pointing to 30 GB per
  • 00:23:35
    second sorry not gigabytes gigabits per
  • 00:23:38
    second there's a difference there times8
  • 00:23:40
    big difference okay um and this is again
  • 00:23:43
    coming from video cards.com but they're
  • 00:23:45
    getting the information sourced from
  • 00:23:47
    bench life the translation is according
  • 00:23:51
    to our sources that being bench life
  • 00:23:52
    sources at this stage in addition to the
  • 00:23:55
    GeForce RTX 5080 using 30 GB per second
  • 00:23:57
    ddgr 7 memory models including the
  • 00:24:00
    GeForce RTX 5060 series are also
  • 00:24:02
    configured with 28 GB per second gdr 7
  • 00:24:05
    memory so now video cards.com has this
  • 00:24:08
    helpful table kind of summarizing the
  • 00:24:10
    information we've had leaked so far uh
  • 00:24:13
    regarding these gpus and their memory
  • 00:24:15
    speeds uh and just other specs that
  • 00:24:17
    we've had leaked previously so it's
  • 00:24:19
    really starting to look like this is
  • 00:24:20
    where we're at with things again my
  • 00:24:22
    biggest issue with vram is here on that
  • 00:24:25
    5070 um this 5070 TI could poten be
  • 00:24:28
    interesting if it's uh at the right
  • 00:24:31
    price but if it's super expensive
  • 00:24:33
    that'll put a damper on that and the
  • 00:24:35
    5090 does look like an absolute monster
  • 00:24:38
    especially compared to the rest of the
  • 00:24:39
    product stack uh since the 5080 looks to
  • 00:24:41
    be like half of it now does that mean
  • 00:24:43
    it's half as fast especially if it has
  • 00:24:45
    slightly faster memory speed compared to
  • 00:24:46
    the 90 um you know we'll have to see
  • 00:24:49
    what what actually Bears out in actual
  • 00:24:51
    gaming
  • 00:24:52
    performance now uh in yesterday's video
  • 00:24:55
    I discussed the fact that there were
  • 00:24:57
    reports that uh nvidia's new app was
  • 00:25:02
    reducing gaming performance and there
  • 00:25:04
    were charts available testing this from
  • 00:25:06
    Tom's Hardware however as I mentioned in
  • 00:25:08
    that video uh they had the app open but
  • 00:25:11
    they didn't try out like which specific
  • 00:25:13
    settings in the app or what if the app's
  • 00:25:15
    closed what if you turn off overlays all
  • 00:25:17
    of that kind of stuff since then uh
  • 00:25:20
    we've had a couple of things one is
  • 00:25:22
    we've had Nvidia themselves state that
  • 00:25:25
    they are aware of a reported performance
  • 00:25:27
    issue related to to game filters and are
  • 00:25:29
    actively looking into it you can turn
  • 00:25:31
    off game filters from the Nvidia app
  • 00:25:33
    settings features overlay game filters
  • 00:25:35
    in photo mode and then relaunch your
  • 00:25:37
    game and we've also had Hardware unboxed
  • 00:25:40
    uh do a video testing this out and again
  • 00:25:43
    they have confirm that it's not the
  • 00:25:45
    entire app that's causing issues it
  • 00:25:48
    appears to just be the photo and filters
  • 00:25:52
    uh needing to be set to off in some
  • 00:25:54
    games like cyberpunk there's a small but
  • 00:25:57
    measurable performance difference uh in
  • 00:26:00
    some other games the performance
  • 00:26:01
    differential is much much higher uh for
  • 00:26:05
    example in Hogwarts Legacy uh you can
  • 00:26:08
    see a a much more significant
  • 00:26:10
    performance uh uh uh differential with
  • 00:26:14
    the with the filters turned on anyway uh
  • 00:26:18
    so that confirms this and again the
  • 00:26:20
    performance hit uh changes based on the
  • 00:26:22
    game you could look at Hardware unbox
  • 00:26:24
    video which will be linked in the video
  • 00:26:25
    description if you want to see all of
  • 00:26:27
    their testing
  • 00:26:28
    uh basically what you should take away
  • 00:26:30
    from this is if you have the Nvidia app
  • 00:26:31
    uh installed you should probably turn
  • 00:26:33
    off photos mode SL filters uh unless
  • 00:26:37
    you're specifically using that feature
  • 00:26:39
    at that time you also require a full
  • 00:26:41
    game restart in order for the
  • 00:26:43
    performance to go back to normal now
  • 00:26:45
    we've seen a benchmark leak for AMD
  • 00:26:48
    ryzen Max Plus 395 stricks Halo now this
  • 00:26:52
    is interesting because stricks Halo is
  • 00:26:54
    supposed to be a ridiculously massive
  • 00:26:57
    APU U when it comes to the GPU inside of
  • 00:27:00
    it it's looking like a um a 40 compute
  • 00:27:04
    unit uh GPU according to the the leaks
  • 00:27:08
    which means you could have some
  • 00:27:09
    significant gaming performance on an
  • 00:27:12
    integrated uh Graphics based laptop mini
  • 00:27:15
    PC that kind of thing now uh it's
  • 00:27:18
    looking like Asus will be launching a
  • 00:27:21
    laptop with this chip and it's looking
  • 00:27:26
    like we have seen this Rog flow z13 show
  • 00:27:30
    up on uh a benchmark here so if we take
  • 00:27:33
    a look at this Benchmark we get single
  • 00:27:35
    core and multicore performance in other
  • 00:27:37
    words it's a CPU Benchmark oh no I
  • 00:27:40
    wanted to see what the GPU could do
  • 00:27:43
    however that being said it does have
  • 00:27:45
    some confirmation of the part here we
  • 00:27:48
    can see the uh AMD
  • 00:27:50
    ryzen um uh AI Max uh plus 395 with
  • 00:27:55
    radon 860s look at that radon 860s is
  • 00:27:59
    apparently what they're calling the GPU
  • 00:28:01
    in this thing uh which is uh you know a
  • 00:28:03
    new naming scheme is this is the low end
  • 00:28:06
    of the 8,000 series integrated Graphics
  • 00:28:08
    you know interesting thing to look at
  • 00:28:10
    there uh but then again we are also
  • 00:28:13
    seeing the um uh the CPU that's inside
  • 00:28:16
    of this is a 16 core
  • 00:28:19
    CPU anyway so we can see some of the
  • 00:28:21
    specs there and all of that and the
  • 00:28:23
    overall Benchmark result is a good
  • 00:28:26
    result for the CPU
  • 00:28:28
    uh apparently coming out significantly
  • 00:28:31
    faster than a um 7945 HX which scores
  • 00:28:37
    2736 single core and
  • 00:28:39
    15896 multicore in the same uh geekbench
  • 00:28:42
    test uh which is
  • 00:28:45
    interesting um but anyway and it also
  • 00:28:47
    looks like it beats the 3D vcash version
  • 00:28:49
    of the part which scores 16393 so it
  • 00:28:52
    looks like the CPU is going to be uh
  • 00:28:54
    very effective but then also again I'm
  • 00:28:57
    really interested what this GPU is going
  • 00:28:59
    to be able to do with 40 compute units
  • 00:29:01
    that is absolutely massive uh compared
  • 00:29:05
    to what we've seen in previous uh mobile
  • 00:29:08
    chips we've seen 16 compute units as
  • 00:29:10
    kind of the top end there so far now a
  • 00:29:13
    lot of times uh integrated Graphics seem
  • 00:29:15
    constrained by system memory so uh I
  • 00:29:19
    don't know I'm just curious how that
  • 00:29:20
    ends up all playing out also the Intel
  • 00:29:23
    Arc b570 has apparently shipped out
  • 00:29:26
    review units to at least one reviewer
  • 00:29:29
    this being Andreas sheiling saying uh
  • 00:29:32
    what have we got here the Intel Graphics
  • 00:29:35
    Arc b570 from ASRock arrived earlier
  • 00:29:38
    than expected now again that doesn't
  • 00:29:40
    mean that the uh benchmarks will be uh
  • 00:29:43
    allowed to be posted anytime soon um but
  • 00:29:46
    maybe it's indicating that U Intel could
  • 00:29:49
    be thinking about getting this
  • 00:29:51
    information ahead of any big competitor
  • 00:29:54
    announcements at CES so maybe uh despite
  • 00:29:57
    the card's release date being a while
  • 00:29:59
    out maybe they're trying to get
  • 00:30:00
    information out there sooner than um uh
  • 00:30:03
    than the Nvidia new GPU announcements
  • 00:30:06
    now uh I've tested out Indiana Jones in
  • 00:30:08
    the great circle on this channel I want
  • 00:30:10
    to mention that update 2 is out bringing
  • 00:30:12
    support for gpus that have less than 8
  • 00:30:14
    gbes of vram because previously they
  • 00:30:16
    were unsupported this game is very vram
  • 00:30:18
    hungry but apparently a little bit less
  • 00:30:21
    so now with less than 8 GB of
  • 00:30:23
    vram um and Global illumination
  • 00:30:27
    improvements to the console versions if
  • 00:30:29
    you actually read the full system p uh
  • 00:30:32
    uh patch notes uh you get PC specific
  • 00:30:34
    fixes including solving some dlss
  • 00:30:37
    related issues uh specifically also uh
  • 00:30:40
    Nvidia frame generation had issues with
  • 00:30:43
    HDR support and there's still a known
  • 00:30:47
    issue that Nvidia frame generation may
  • 00:30:49
    be temporarily disabled if HDR is
  • 00:30:51
    activated for the first time to work
  • 00:30:54
    around this issue you would disable and
  • 00:30:55
    then reenable dlss and that only needs
  • 00:30:58
    to be done once and then that should be
  • 00:31:00
    solved there's some other issues uh kind
  • 00:31:02
    of solved here as well and they also
  • 00:31:05
    fixed an issue where Global illumination
  • 00:31:07
    may be completely disabled when playing
  • 00:31:09
    with below minimum vram graphics cards
  • 00:31:11
    so previously if you had below 8 gabes
  • 00:31:14
    of vram the global illumination would
  • 00:31:16
    just be disabled so you wouldn't get any
  • 00:31:18
    real lighting since all the lighting in
  • 00:31:20
    the game was based on the ray tracing uh
  • 00:31:23
    last thing I want to mention is that The
  • 00:31:24
    Witcher 4 devs have confirmed gwent so
  • 00:31:27
    if you uh concerned that the new gwent
  • 00:31:30
    expansion might not actually have gwent
  • 00:31:32
    they have confirmed gwent so there's
  • 00:31:35
    that I hope you guys enjoyed today's
  • 00:31:37
    video and uh uh I hope that you have an
  • 00:31:41
    excellent day that's what I was thinking
  • 00:31:43
    uh don't know why I struggled to say it
  • 00:31:45
    hope all of you have an excellent day
Etiquetas
  • CES
  • RTX 50 Series
  • Neural Rendering
  • Acer
  • Nvidia
  • AMD Strix Halo
  • VRAM
  • GDDR7
  • Intel Arc