I'm curious about how ray tracing affects performance. Essentially all of the ray tracing is done on dedicated hardware alongside a normal dedicated GPU, correct? So essentially when RTX is rendering your lighting/shadows it would be similar in performance to a normal pascal GPU today running a game without calculating shadows or ray tracing at all, correct?
This brings me to my second point of confusion. How do you scale ray tracing to different levels of performance? The way the technology was described you either are ray tracing or you are not. The only thing I can see is different elements being turned on or off (like GI or shadows or AO) one at a time. It seems like either it is computing and working or it is not. Like you cannot ray trace at a lower quality, so does speed come into account?
OR are the only differences between the 2070-2080 ti the number of regular computing cores, which means that regular game compute performance is different regularly between them but the RTX differential would be the same?
It's just shitty 12nm stopgap GPUs with >MUH GIGARAYS marketing, nothing really to see here.
Nvidia increased the power limit and die size compared to pascal to get like 20% more performance.
Brandon Cox
I believe you can lower the number of rays output by sources to reduce the raytracing load.
Carson Ortiz
Google it. A company had this technology in hardware about 10 years ago. They provided far more detail about how they did it and how the hardware works. It's essentially a dedicated pipeline that is fed the volume data in parallel to the traditional GPU pipeline and produces results in parallel with rasterization just in time for the final frame.
Nvidia wasn't the first to develop this technology. The algorithms are still the same as they were many years ago.
>It's essentially a dedicated pipeline that is fed the volume data in parallel to the traditional GPU pipeline and produces results in parallel with rasterization just in time for the final frame. so then what I said is true? That it will reduce the load on the rasterizing GPU and improve performance? Because shaders at this point in time take up about 50% of GPU calculations. That means a 1080ti + ray tracing would turn from a 4k60hz GPU to a 4k120hz GPU overnight. That's pretty huge, but only if the ray tracing can compute at that level as well. I cannot imagine the nightmare scenario of being bottlenecked solely by ray tracing
yes good nigger goy, that is how it will perform in classical graphical computation. Well done, you can do simple maths fuck off retard, this is a technology thread
Cooper Stewart
but that's not how it's computed. It's traced from the camera backwards, so how can you reduce the load of rays from the source when the calculations don't even begin from the source?
Ryder Hall
3 types of cores now- RTX Tensor CUDA
Cuda is the older compute cores, they're good for FP32 compute and that's it. Tensor is a newer core type, they're good for FP32 compute too, as good as CUDA cores, but they're also really good for FP16 (twice as fast as FP32!) and INT8 (Twice as fast as FP16!) and now INT4 (Twice as fast as Int8!) RTX is the new 'ray tracing' cores, they're also just compute cores, but they're restricted to INT4 and INT8
The new ray tracing software could run on CUDA cores, it would just run ~8x times slower clock for clock. The Ray Tracing is running on both the RTX cores AND the Tensor cores (both are equally as fast for the job) but the Tensor cores use more power to do the same thing (but are flexible in they can also do the work of CUDA cores)
4bit ray tracing will be super noisy (anything less than 64bit will be) and part of Nivida's secret sauce is their denoising filter.
David Wood
thank you very much. So I'm guessing that CUDA core counts are actually the exact same as Pascal but with RTX and Tensor cores added? That would be incredible. It would mean that Pascal would essentially get a price slash and if you aren't interested in Ray Tracing just yet then you can essentially ignore RTX
Cameron Hughes
>rtx and tensor cores do matrix calculations therefore they are literally the same thing is just nvidia calls them rt cores for pr plus all of the cores at nvidia side ARE ALWAYS depricated
Caleb White
was this that euclideon company?
Blake Martin
PASCAL will not get any fucking price slash . NOVIDEA has no competition .
Aiden Jenkins
The RT core essentially adds a dedicated pipeline (ASIC) to the SM to calculate the ray and triangle intersection. It can access the BVH and configure some L0 buffers to reduce the delay of BVH and triangle data access. The request is made by SM. The instruction is issued, and the result is returned to the SM's local register. The interleaved instruction and other arithmetic or memory io instructions can be concurrent. Because it is an ASIC-specific circuit logic, performance/mm2 can be increased by an order of magnitude compared to the use of shader code for intersection calculation. Although I have left the NV, I was involved in the design of the Turing architecture. I was responsible for variable rate coloring. I am excited to see the release now.
I just can't accept the "RTX 2080" name, it's somehow even more dumbfucking gaymer-ey than "GTX 1080". No other product in any other industry except Huang's spew will be named "RTX 2080" or "RTX 2080 Ti". It's so fucktarded, jesus christ. I hope they sell 0 units.
Also, I'm surprised they went with naming it "RTX", since that's so close to AMD's "RX". Will they get sued by AMD or will subservient c*ck AMD change their GPU branding for the second time in three years?
Aiden Hughes
>So I'm guessing that CUDA core counts are actually the exact same as Pascal but with RTX and Tensor cores added? No, the RTX cards have more CUDA cores and the cores are a different architecture, according to Jacketman in the presentation they perform 1.5x better than Pascal cores.
Kevin King
R is there for raytracing, that much should be obvious
Brandon Cooper
Its kind of ironic that Nvidia went back to their roots, but then brought along all the advancements to start over; creating a hybrid architecture.
GPUs started out as fixed function pipelines, and then moved into programmable shader territory. The Turing uArch is basically a hybrid fixed-function core with programmable sub-elements.
RTX, Tensor, and CUDA all have specialized functions; but overall, they are programmable. More than anything, I think this is the greatest achievement they've managed to do with the GPU. They've also become effectively the first to redefine how GPUs will be designed pretty much everywhere. Hyperspecialization of the GPU architecture is now THE future and forever will be. Any approach to creating a unified architecture that does it all, will be considered second class and a waste of man power, capital, and material resources on a massive scale.
This is actually amazing. I hope Navi follows a similar path, or AMD is catastrophically fucked.
Cameron Parker
Why don't they tell us the number of rtx cores? Or are they included in the Cuda count?
Aiden Jenkins
RTX is done to purposefully segment the market and give the impression that anything that is GTX based is effectively and absolutely obsolete to the RTX tech. All due to a variety of game developers introducing it into their titles and the appearance of significant change, reduction of LoE and costs leading to a greater degree of visual fidelity and depth in lesser amount of TIME.
The impression Jensen wanted to leave with the crowd was that with RTX everywhere, games that took 4 years to make can now be made within 1.5-2 years, due to the benefits of RayTracing. This is obviously not the case, as a reduction wouldn't be that much--but one can expect anywhere from 6-8 months reduction overall due to the amount of time less an artist would need to generate a material, texture, apply it to a model, and then view how light behaves within a specific scene. Get 4 of those Quadro 8000s in an NV-Link configuration, and you can basically take a full BF5 or BLOPS4 map, with all objects, geometry, particles, alpha-effects, PBR, everything; render it all at once with RT and get a scene without needing to resort to baking light maps, shadows, etc. Being able to prototype vertical slices to publishers (for those that can afford these high end workstation GPUs), will be massively easier.
And if more developers & publishers onboard, then more gamers will as they want to play with the latest and greatest. Its mind games. Wait for benchmarks and vote with your wallet, don't subscribe to the hype, preorder, etc.
Gavin Lee
...
Jayden Wright
What do the T, the X, the 80, and the Ti stand for, dumbass?
Ethan Sullivan
RT is for "Ray Tracing" X is there to otherwise match their well-known GTX branding 80 has been standard for their high-end product offering for multiple generations, so they're sticking with it "Ti" is also branding for their enthusiast products, IIRC it actually stands for "Titanium" and goes back at least to the GeForce 3 days
I hope this alleviates your great confusion over this simple matter
Blake Flores
GPU has always been hyperspecialized its the most parallel task a computer can make basically.
RTX is a market term for hardware specialized compute units made for raytracing. Raytracing needs fast as balls cores and a lot of them but not accurate. These RTX cards will be 8x faster than the non RTX cards for raytrace compute. Problem is very few games are going to use raytracing real time until most of the market has this type of specialized hardware.
Current AMD cards are much better than current geforce cards for raytracing tasks as they have higher raw compute performance. Maybe nvidia caught wind of this or upcoming cards from RTG that uses modular design similar to ryzen which they can place specialized core complexes for w/e task the card is made for.
Nicholas Ross
sounds like youre making shit up on the spot, nigga
James White
You will NEVER live to see realtime raytracing, and when it does become possible somewhere down the line, it sure as fuck won't be used in video games if they even still exist as we know the medium today
Wyatt White
not him but what? he is right on everything lol
They said that but yet its here real time RT requires an assload of compute power but we have that now
Full scene exclusive RT for all lighting? Thats a little ways off 10~ years or so
Elijah Turner
AMD's ultimate end goal with Radeon is to make a Zen GPU. Design a Core Complex type GPU, and then stick together many of these dies on an interposer connected by an InfinityFabric like interconnect. The yields on Vega suck. If they were able to make a Zen-like GPU, their yields would shoot through the roof, and having many-die implementations, would allow them to compete with Nvidia on implementation, price, and most importantly, performance with near parity.
Imagine something like Polaris, say the RX 570 class of performance/power (at its most efficient voltage); take a socket the size of Threadripper or EYPC, and place 1 dies on each side for entry level, 2 dies on each side for mid-range, and 3 dies on each side for high/ultra; all made at 7nm's maximal power efficiency process. With a Zen-like GPU redesign, their yields will be through the roof and they'll be able to make use of majority of the dies per wafer. They won't have to clock the GPUs insanely high to compete or overvolt them to maintain within 80% perf window.
They could probably clock them back down to 800-900MHz targets, and while Nvidia releases 1.5GHz ~5k shaders, AMD can release 900MHz 8-9k shaders that can access 16-24-32GB of GDDR6 within the same power target, heat target, and cost as Nvidia's competing high end.
With just 2 dies on an interposer, they can basically say 1 core is pure rendering, other core is pure RT and DNN/AI denoiser. With 3 dies on either side, they can say 4 cores (and all related shaders) are pure rendering and 2 cores are RT & DNN/AI denoiser. AMD can then effectively compete with Nvidia's own game, with better yields, more raw throughput at a fraction of the cost.
so wait is this just basically dedicated hardware for ray-triangle intersections or some shit
how do the APIs look like that actually interface with this shit
Eli Morris
How will 2 1080TIs do against a single 2xxx series card in ray tracing in sli and in multi gpu say for Vulkan and DX 12?
Bentley Jenkins
NVIDIA were looking into MCM GPUs as well (they released some paper about it a while ago), but it doesn't look like that's in the cards any time soon from either company. Apparently there's some difficulty in making such an arrangement appear and perform like a single monolithic GPU and that's what the goal is, they want to avoid requiring something like CF and SLI profiles.
Kayden Nelson
AMD has been working on this for a while NAVI was supposed to have "scalability" Infinity fabric works better with faster ram so GDDR6 or HBM2/3 will improve the performance over Zen IF We have already had this with SLI/crossfire sort of. Graphics workload is extremely parallel so more compute the better. We can keep reducing the precision as well which is what nvidia is doing with tensor cores and RTX cores
Blake Bennett
>Apparently there's some difficulty in making such an arrangement appear and perform like a single monolithic GPU
How? Compared to CPU, I'd think it'd be much easier to split a GPU into multiple chips. It's basically a huge collection of tiny processors anyway. Nvidia just lacks based&redpilled Infinity Fabric (tm) patented technology.
>in ray tracing in sli They're going to get slaughtered, can't beat the specialized hardware. In practice I imagine the games supporting ray tracing won't even be able to use it on non-RTX cards.
As for multi-GPU in general, Pascal (and below) suffer from bandwidth bottlenecks, as the old SLI bridge is pitiful and almost everything has to go over PCIe. It's even worse if the CPU/motherboard doesn't offer 2 PCIe x16 slots, SLI can really choke on x8/x8. RTX 2080 (Ti) however use NVLink which is vastly faster. That sounds like a good thing but we haven't seen how much things improve in practice.
Benjamin Peterson
>I'm curious about how ray tracing affects performance It doesn't because playing video games is a waste of time.
Mason Wood
Zen figured it out with the butterdonut configuration. I imagine that a Zen-like GPU would similarly benefit from such a configuration. That said, this is grossly simplifying the effort--but I'd expect to see an MCM approach that's very much TR/EYPC like from AMD first. Nvidia has no reason to go Zen-like, because they have a near monopoly on the market. They can go as big as they want, and fab partners will cater to them, because they have the coin. They could go smaller towards MCMs, but that would require segmentation of their uArch, and they've got a goose that lays the golden egg each time; no point in trying to kill it to try and get more gold out of it too early.
Levi Butler
>they want to avoid requiring something like CF and SLI profiles. That will be done on the API side of things IIRC DX12 allowed cross brand cross gpu SLI/crossifre
Depends on how many cores are dedicated to RTX each RTX core is exactly 8 times faster than a cuda core for raytracing this assumes the same clock speed
Core config for 2080 ti is 4352 but we dont know how many of that or if its more on top for the tensor/RTX Also ray tracing is not 100% of what the GPU computes it still has to render all of the normal shading lights shadows ect
Kevin Perez
>IIRC DX12 allowed cross brand cross gpu SLI/crossifre And fucking nobody implements it, since that requires actual dev time/effort and no consumer runs such a setup.
Landon Foster
>he is right on everything lol >he isn't making it up yeah, I know, but it sounds like he is because of how retarded Huang's naming scheme is; it's such a shitty, pandering name that's so obviously made to sound like COOL NEW GFX CARD BRO, GOTTA BUY THIS BRO to little boys and onions-indulgers.
Huang should be put to the furnace tbqh desu
Jaxon Diaz
Sli doesn’t choke on 8x what are you on crack? Also HB bridge has plenty of bandwidth for today’s games. You only lose 3% performance at most with 8x vs 16X. I think pascal can maybe do some low ray tracing just not as good as rtx
Caleb Wright
>Most of us had thought AMD would start to use its Infinity Fabric interconnect to join smaller GPUs together in a single package to create a high-performance multi-chip module (MCM) with the new Navi graphics silicon. But we recently spoke with David Wang, the new SVP of engineering for AMD’s Radeon Technologies Group (RTG), and there’s pretty much zero chance that’s going to be worked into next year’s Navi GPUs.
People below 13 isn't allowed in these parts And yes, what he said was true, but the Ti part actually dates back way before by the name Geforce 2 Ti
Daniel Morgan
No consumer runs a set up like that or even have multi gpu because DEVS NEVER DO IT RIGHT. “Guys multi gpu never works waaaahh”, hmm I fucking wonder why?? Devs are at fault and are holding us back from pushing hardware with better software. I’ve met a lot of people who want to run multi gpu but hate the support.
>dev thinks it’s a small market >”not work our time for multi gpu” >Tells gamers it’s too small of a market >wonders why it is a small market
2018 Jow Forums anyone else fed up and tired of this autism? This is still tech related in general
This is why I said "was" in the above comment Navi is a replacement for polaris
AMD having the patents in the special innterconnect infinity fabric is why they could build the first multi chip GPU
Camden Perry
>appear like a single monolithic GPU plausible >perform like a single monolithic GPU no even if they had optical interconnects, simply no. make all your operations act like L3 cache why don't you
Matthew Rivera
I ran a quad SLI setup back in the day SLI sucks because devs suck It has to be done on the API side of things DX12 fully supports it but only if the engine supports it too AND the game supports it It needs to be seemless on the API basically the API should see all available graphics cores as a single unit
Nicholas Torres
That's why THIS: needs to happen. Once it does, SLI/CF like concepts will return and devs will have NO choice but to cater to multi-GPU markets. The only reason why Nvidia keeps making massive GPUs, is because they can and nobody can stop them--and since they have near complete control to the market, devs have no reason to not support anything but single-gpu efforts.
If suddenly there was a MCM GPU that matched the 2080Ti in performance tomorrow, you bet your ass developers would cater to it even with Nvidia's massive market-share. They're not going to say no to like 30% of the market's money, their publisher overlords will fire anyone who says "I don't wanna support it" and replace them with "hell yeah I wanna support it."
Lucas Brooks
>tfw dev >tfw marketing and project managers are pushing impossible deadlines in my throat >tfw have to deal with quirks in hardware and software because fellow devs and engineers are pushed to meet impossible deadlines by their own project managers >tfw it's impossible to do anything good in timeframes given >tfw managers just throw more devs at it because they think it'll go faster >tfw end users hate me because they think its my fault
so yea mGPU is coming guys just wait haha.
Wyatt Peterson
Just put shadows to medium, up some anti-aliasing and there you go ...Raytracing.
Ryder Martin
navi was supposed to be that, apparently it's not. zen like MCM on GPU is a clusterfuck for drivers rumor has it that vega prim shaders was a step for it, but another rumor said that hardware bug got into fabs and that idea got fucked we can only hope it is a hardware bug in vega and prim shaders are actually possible on driver level, why would driver dev lie about that I still want to keep hope.
