PCIe 6.0 Specification finalized in 2021 and 4 times faster than PCIe 4.0

Discussion in 'Frontpage news' started by Hilbert Hagedoorn, Jun 11, 2020.

  1. schmidtbag

    schmidtbag Ancient Guru

    Messages:
    5,748
    Likes Received:
    2,193
    GPU:
    HIS R9 290
    That doesn't change my point at all... Unless you happen to know different, it's basically the GPU equivalent of a paging file. The GPU ideally doesn't depend on that pool of system memory to perform calculations every frame, but it will when the VRAM isn't sufficient. It's woefully inefficient to feed a GPU core from system memory. Just look at how badly crippled the AMD APUs are by memory bandwidth. Their performance scales up almost perfectly proportionately with memory speed. That doesn't always apply with discrete GPUs.
    Of course NVMe drives don't rival DRAM bandwidth, I never said they would. But NVMe drives are already fast enough to be bottlenecked by the x4 lanes that many of them are slotted in. That means NVMe drives can demand at least 25% of an x16 slot. That's significant. So if you have on-board storage for your GPU, that spares you all those lanes for other purposes. Or in the case of my argument, spares those lanes to the point where they're no longer necessary.
     
  2. meanwhile on the pornhub forums... ahh yes... :rolleyes:
     
  3. schmidtbag

    schmidtbag Ancient Guru

    Messages:
    5,748
    Likes Received:
    2,193
    GPU:
    HIS R9 290
    I stand corrected.
    Yes, and I don't see how that negates my point. If anything, it reinforces my point since those demand more compute power than plain textures.
    Again, not sure how this negates my point. I don't disagree.
    Fine, here's your damn proof, despite the fact I already told you what the source was:
    https://www.techpowerup.com/review/nvidia-geforce-rtx-2080-ti-pci-express-scaling/3.html
    Really wasn't that hard to google search it.
    Those are real world benchmarks. Not the stupid theoretical crap Carfax keeps whining about.
    We're talking the best gaming GPU on the market here, and worst-case scenario we're talking a roughly 10% performance loss in framerates that are already in the hundreds.
    PCIe 3.0 is ancient by technology standards and x8 lanes still keeps up.

    Proof by assertion my ass. At least I actually have proof. Where's yours?
     
    Last edited: Jun 19, 2020
  4. Ok hmm so to saturate each lane is to equate to the transfer rate ... then again at Gen 3 past a certain amount gives you enough to yield what may be the minimum for most cards of current gen under 8 lanes for x8 Gen3 so I wonder if 2080TI need even at x16 Gen even saturate all lanes at all and or if it does or if it is able to or "smart" enough to utilize it in a manner of "up to a certain point" per lane + each lanes transfer rate (gen3). Anyways. Interesting. Nice article, it's just more on the surface - I'd probably need to read a white paper from intel on PCI-E spec or some crap
     

  5. Carfax

    Carfax Ancient Guru

    Messages:
    2,915
    Likes Received:
    465
    GPU:
    NVidia Titan Xp
    So after you insulted not only myself but DmitryKo and Alessio 1989, you're complaining about being insulted in turn?

    The old axiom "If you can't take it, don't dish it out" is well warranted here I think.

    Why wouldn't it? A Ferrari la Ferrari and a Honda Accord are both cars with plenty of similarities, but they're obviously quite different as well.

    NVLink uses a hybrid cube mesh framework, which is nothing like PCI's shared bus. Stop spouting nonsense!

    Keep backpedaling. This is what you said:

    Do you know what NVLink actually does? Not even PCIe 6.0 @ x16 is sufficient for what Nvidia wants. They need the bandwidth because the memory is shared. PCIe is not meant to compete with memory bandwidth.

    Also, PCIe works well for workstation and desktop applications. It's in more performance critical sectors that it falls short. AMD uses PCIe for their DX12/Vulkan mGPU technology which allows the GPUs to share and mirror each other's VRAM, and it out scales Nvidia's SLi technology (I don't believe this applies to the Turing generation SLi bridges which use NVLink technology), though the latter has more consistent frametimes.

    And I don't think you understand the significance of the impact that the PS5 and XSX will have on game engines and game design. If this was just an ordinary console refresh with HDDs then I would probably agree with you, but it's not. Having SSDs in the next gen consoles (especially with dedicated hardware decompressors) is a big deal, and lots of things are going to change.

    As I said in a previous post, even the most sophisticated contemporary 3D engines limited their streaming to a few hundred MB/s due to the assumption that the game was installed on a HDD. But with SSDs, they can bump that up to the GB/s range.

    Because progress is unceasing. I'd rather have too much power, than just enough. Also, PCIe 6.0 will have lower latency which has other benefits. It's not just about bandwidth.

    Do you have anything that contradicts that benchmark?

    Thought not....

    Techpowerup benchmarks are irrelevant because those games are not representative of the games that will be available 2-3 years from now when the next gen console generation is in full swing.

    So what did you think I meant when I said, "So although my GPU is only using 2.5GB out of 12GB of VRAM and isn't even close to being maxed out." By not maxed out, I meant in terms of GPU core usage.

    A good example of this is Pathfinder Kingmaker, which was made with Unity. That version of Unity that they used is garbage. Tons of stuttering and slow loading times, despite the fact that it's a slow moving isometric RTS game which I have installed on a Samsung 960 Pro and it barely uses my CPU or GPU power.

    For the fourth time, until we see what kind of impact the next gen consoles will have on game design and game engines, this "prophesizing" is just conjecture and nothing more.

    If the UE5 demo that we've seen is any indication of where gaming is headed, then it's quite conceivable that there would be a large increase in demand for PCIe bandwidth from the GPU.
     
  6. schmidtbag

    schmidtbag Ancient Guru

    Messages:
    5,748
    Likes Received:
    2,193
    GPU:
    HIS R9 290
    First of all, you guys are the ones attacking my point of view first.
    Second, where am I complaining? Call me whatever you want; I don't care. My point is you're not contributing anything toward proving yourself right when you're just mindlessly throwing insults. You might as well call me a "stinky poo-poo head" because that's as productive as your previous comment was.
    Right, and those differences are important. Important enough that even those who know absolutely nothing about cars knows a Ferrari and a Honda aren't really comparable.
    The differences between NVLink and PCIe are significant enough that Nvidia felt the need to make NVLink.
    Using the marketing terminology doesn't change what it really is...
    Do you have reading comprehension issues? Where's the backpedaling? I sincerely don't understand where you think I am. I basically said the same thing both times: NVLink was made because PCIe doesn't have enough bandwidth, not even 6.0 @ x16.
    Where did I say PCIe doesn't work well for workstation and desktop applications? It's an absolute necessity for x86 PCs these days.
    I'm aware how PCIe is used for mGPU setups. I don't see how that affects my point, especially since both brands have basically given up mGPU at the consumer level. Once again, I don't understand what it is you think you're arguing.
    Again, where am I saying otherwise? I agree. But that still doesn't refute my point. Next-gen consoles aren't going to cause PCIe bandwidth needs to suddenly explode, because GPUs are moving toward on-board NVMe storage. That frees up a significant amount of bandwidth.
    As of today, PCIe 5.0 @ x8 will be "too much power" for future consumer-grade hardware and I assure you, it will remain that way for quite some time. I say future hardware because PCIe specs run at the slowest device, and there are 3.0 devices today that could use more than x8 lanes (but don't need the full x16).
    Um.... you do realize that the only thing I'm arguing here is about bandwidth, right? I have no problem whatsoever with PCIe 6.0. I look forward to it when we need it. Same with 5.0. My point from the very beginning, and the only point I've been trying to make, is that electrical x16 slots will no longer be necessary.
    Your presumptuousness and ego is unwarranted. First of all, I already mentioned a source to you that contradicts it, which I also already posted (not to you). Here it is again:
    https://www.techpowerup.com/review/nvidia-geforce-rtx-2080-ti-pci-express-scaling/3.html
    "hurr durr but the theoretical performance blah blah blah".
    It is entirely relevant. The 2080Ti will be the last high-end GPU on PCIe 3.0 and it's barely demanding more than x8 lanes (sometimes not at all). On PCIe 4.0, we're doubling the bandwidth, yet we are most certainly NOT doubling the performance of the GPUs. And as I stated several times already, on-board NVMe storage will further reduce bandwidth.
    So what the techpowerup benchmarks show is actually the worst-case scenario. x16 will only become less relevant as time goes on.
    Remember, 4.0 is already obsolete; Intel is going straight to 5.0. At the rate GPUs are improving, it could be a decade until 5.0 @ x8 will be saturated, and this is accounting for the way next-gen games will be handled.
    Supposing that is what you actually meant: you're still too dense to understand that:
    1. If a game is only using 2.5GB, it wasn't optimized to handle your beefy GPU. It was optimized for something much weaker. So obviously it isn't going to max out anything when the entire GPU isn't a bottleneck. As you yourself pointed out, games are not built to handle future advancements in technology.
    2. Let's say your GPU wasn't overkill, and had 4GB. If the core still isn't maxed out, well, that's why using low-res textures at a distance helps.
    3. In most situations with AAA games, the GPU is the bottleneck. Every little thing to improve performance is worthwhile. Using low-res textures in far away objects yields a minimal visual difference but a significant decrease in GPU load.
    To give you the benefit of the doubt, I'm going to assume v-sync is off. If the game barely uses your CPU or GPU, the only thing it exemplifies is how to make a game poorly. A well-made game (with an uncapped frame rate) will keep pumping out frames until some part of the hardware is bottlenecked.
    So, what exactly does this game do to prove your point? Remember, you're trying to prove to me why x16 slots in the future will be necessary. I doubt this game even uses x4.
    We know the demands of real-world results. We know the pace at which GPU technology is advancing. We know how fast NVMe storage is. We know that PCIe bandwidth roughly doubles per-generation. We almost have a complete picture even without having the hardware to test with. The only thing that is uncertain is any breakthrough in GPU design.
    If what I say sounds like prophecy, I don't think you are putting enough forethought into where technology is headed in the next several years.
    Not really. It will demand more bandwidth when loading in new assets, but the bottleneck at that point will be NVMe storage, not PCIe.
     
    Last edited: Jun 20, 2020
  7. DmitryKo

    DmitryKo Master Guru

    Messages:
    376
    Likes Received:
    122
    GPU:
    ASUS RX 5700 XT TUF
    Again, dedicated on-board NVMe or flash memory storage for desktop GPUs is neither announced, nor needed. It's a solution to a problem that does not exist.

    Both AMD (Zen) and Intel (Rocket Lake) increased the number of PCIe lanes in their recent processors to include both NVMe x4 and GPU x16 links directly on the CPU, in addition to dedicated links to the chipset.


    It's because current PCIe 3.0 is much slower than both HBM2/GDDR6 local video memory (1/40th) and system DDR4 memory (1/4th).

    PCIe 6.0 improves this bandwidth by almost an order of magnitude - that would be 1/2th to 1/3th of the memory bandwidth of a typical mid-end discrete GPU today.

    And a 10x difference between a high-end hard disk drive and a mid-end NVMe drive (200 Mbyte/s vs ~ 2 GByte/s) is what enables next-gen game consoles to radically reduce loading times.


    No. System shared is just another graphics memory pool, a kind of slower "far memory" for the GPU. It can hold whatever the programmer directs it, including extra textures that do not fit into video memory, though typically it's used for command buffers, compute buffers, constant buffers, and other resources infrequently read from the GPU.

    It's not a paging file, since you cannot allocate physical memory in a paging file, and it's not using external storage either.


    The point is, anisotropic fitering needs anisotropic mipmaps (just like trilinear filtering needs regular mipmaps), since it works by tapping samples from different mipmap levels, to eliminate blurriness or shimmering on inclined surfaces; these mipmaps are generated automatically either by authoring tools or by the API runtime, and they only take a fraction of the full texture. They are not some special low-res textures designed to reduce memory bandwidth.

    AF doesn't need more compute power either - TMUs are fixed function units, and texture cache is optimized for memory access patterns typical in AF.


    They didn't bother to run PIX tool or Radeon Memory Visualizer to capture graphics memory usage. If resources were created in local video memory, what's the point of measuring the impact of PCIe speed when you don't actually transfer textures over PCIe?

    If they could tweak the game to oversubscribe texture memory, and/or test a graphics card with smaller video memory, in order to make automatic eviction possible, then they could claim that PCIe 3.0 doesn't make any difference.


    What do you want me to prove exactly? That 8x the bandwidth in PCIe 6.0 allows increased useage of shared system memory? Unfortunately I don't have a PCIe 6.0 system or graphics applications from 2025 optimized for PCIe 6.0 bandwidth.

    I also have neither time nor desire to perform the tests with real-world games. I consider it an obvious point that faster PCIe bus would improve graphics performance when video memory is overcommited.


    3DMark TimeSpy Extreme only uses 3.3 GBytes in local heap - and that's supposedly a high-end test with 4K textures.

    I don't currently own any recent graphically demanding game to capture memory usage with Radeon Memory Visualizer https://gpuopen.com/rmv/ but I suppose it's common to target 4 GB cards.
     
    Last edited: Jun 22, 2020
  8. DmitryKo

    DmitryKo Master Guru

    Messages:
    376
    Likes Received:
    122
    GPU:
    ASUS RX 5700 XT TUF
    The differences could be significant enough for their proprietary supercomputers, but this does not apply to PCIe 6.0 and HEDT PCs with two or three x16 slots.
    It was made because PCIe 3.0 didn't have enough bandwidth, and because Intel processors didn't have enough PCIe lanes, so it was easier/cheaper to design their own cross-bar controller for a simpler point-to-point protocol with native cache coherence.

    PCIe is still there in NVidia DGX-2, which uses twelve PCIe 3.0 switches to connect each pair of GPUs with an PCIe x16 slot for NIC and RAID cards and one of the two CPUs.

    Desktop video cards are not moving toward onboard NVMe storage, unlike HPC/supercomputers.
    PCIe x16 links are not going anywhere - not with PCIe 4.0, and not with PCIe 5.0 or 6.0 either.


    It's you who started calling names when your arguments were found unconvincing.

    It's because you don't change your points.
    True in that your argumentation is a complete collection of intentional fallacies.
    At least it's not about onboard NVMe x4 drives as the reason for dumping PCIe x16 links.
     
    Last edited: Jun 22, 2020
  9. schmidtbag

    schmidtbag Ancient Guru

    Messages:
    5,748
    Likes Received:
    2,193
    GPU:
    HIS R9 290
    I don't disagree, but it doesn't change the fact that it helps reduce bandwidth.
    Yes, and your point is?
    So is PCIe 4.0, and 5.0, and 6.0. Do you know how fast HBM2 and GDDR6 are? Even PCIe 6.0 @ x16 isn't fast enough. HBM2 and GDDR6 will be obsolete by the time PCIe 6.0 is released. More data has to be used over the bus than just sharing memory, so even if x16 were fast enough, it's still going to get bottlenecked.
    And yet you say on-board NVMe storage for a desktop GPU isn't needed? A bit contradictory there.
    Ugh you're one of those literalist types...
    Forget I said anything - this is not a rabbit hole I'm in the mood for and it's not important to the discussion. For the record, I didn't say it is a paging file, I said like a paging file, because I know they're different, but the fundamental principles are similar.
    I know what AF is but I don't understand what it has to do with the discussion. I get the whole "pop up" thing but like you said, AF takes basically no compute power, so that isn't what was being referred to. The "popping" textures is due to things coming in closer from a distance, regardless of viewing angle.
    What difference does it make? The only thing that matters is real-world results while you're in the game or application. The whole point of this discussion is whether x16 slots are necessary for consumer motherboards in PCIe 5.0+. As the tests show, not even PCIe 3.0 @ x8 is being consistently maxed out. PCIe 3.0 @ x16 is still fast enough to fill up a 2080Ti's VRAM in 1 second. Are you so impatient that you couldn't have all assets loaded in 1 second? Not that it matters, because the bottleneck there would be storage, not the PCIe bus.
    I see what you're saying but at that point your system memory will most likely become the bottleneck. PCIe 3.0 @ x16 offers a lot more bandwidth than DDR4.
    What? No. My point is that unless a major breakthrough in GPUs occurs, 6.0 @ x8 will offer sufficient bandwidth for consumer-grade hardware.
    Only if the system memory is fast enough, which currently, it isn't.
    Have you seen memory scaling benchmarks of APUs? No matter how fast you push the memory, the performance scales up almost perfectly proportionately, suggesting they're utterly starved for bandwidth. These are crappy low-end GPUs and they're nowhere close to reaching their full potential, and yet you think feeding a high-end GPU through system memory is going to be a problem for PCIe?

    I don't know how many times I have to say that PCIe 4.0, 5.0, and 6.0 still aren't enough to compete with NVLink's bandwidth. Remember: even if PCIe 6.0 @ x16 was enough, the 2 linked GPUs don't get the entire bus all to themselves. The CPU is going to talk to them too, and that also needs bandwidth.
    Why would I? To do so means I have no confidence in my stance. To do so, I would be called a backpedaler. This statement of yours is so weird I don't even know if it's negative, because nobody would ever complain about such a thing.
    Where? Point out 1 time where I did that. Just 1.
    Meanwhile, here you are rambling about AF and DGX-2 as though it has anything to do with what's being talked about.
    Whenever I'm asking you things like "what's your point?" or "how does this refute my point?" that's you using an intentionally fallacy.
     
    Last edited: Jun 22, 2020
  10. Carfax

    Carfax Ancient Guru

    Messages:
    2,915
    Likes Received:
    465
    GPU:
    NVidia Titan Xp
    This should put an end to your clucking. PCGH.de did a test with PCIe 3.0 vs 4.0 x8 on a 5500XT with both 4GB and 8GB of VRAM. Their findings support what DmitryKo has been saying, that a faster PCIe connection will increase graphics performance when the VRAM is scarce.

    This isn't because the GPU is using the system memory to buffer frames. It's because the data can be swapped out much faster between the system RAM and the VRAM. The benchmark with Wolfenstein Youngblood is the most salient, and shows a whopping 85% difference between the 5500XT 4GB with PCIe 3.0 8x vs PCIe 4.0 8x. Far Cry New Dawn shows 78% advantage to PCIe 4.0 as well:

    [​IMG]

    [​IMG]

    I want some of whatever you're smoking. PCIe 3.0 x16 is 16GB/s or 32GB/s bidirectional. High performance dual channel DDR4 setups are getting around 50GB/s, and quad channel around 100GB/s depending on memory speed.
     
    Last edited: Jun 22, 2020
    DmitryKo and Alessio1989 like this.

  11. PrMinisterGR

    PrMinisterGR Ancient Guru

    Messages:
    7,652
    Likes Received:
    581
    GPU:
    Inno3D RTX 3090
    It's beautiful how in this thread we all know more than the people who build the hardware and the game engines.

    And they all say that NVMe and interconnect bandwidth matters a lot. Why is there even a dispute about this I truly don't get.

    Probably because it's not that readily apparent yet, in the "it hits you in the face" kind of way. But even Microsoft realized they need to do something about it, hence DirectStorage. Next Gen GPUs will be able to access PCIe storage directly, and hundreds of engineers are working on it.

    No why would they do that? :p
     
    TheSissyOfFremont likes this.
  12. schmidtbag

    schmidtbag Ancient Guru

    Messages:
    5,748
    Likes Received:
    2,193
    GPU:
    HIS R9 290
    Nope. Not even a little.
    Yeah, obviously performance will greatly suffer when you're running something very demanding and memory intensive on a 4GB GPU on PCIe 3.0 @ x8. I never said otherwise. I know 3.0 @ x16 is necessary (and 4.0 @ x16). 3.0 @ x8 barely has enough bandwidth for modern GPUs, and you think you proved your point by having that slot feed a GPU core from system memory?
    Meanwhile, the 4.0 @ x8 results show that this 4GB GPU, while reading from RAM, is barely slowed down compared to the 8GB model. There isn't even enough bandwidth to read from DDR4 at that speed, let alone all the other communication that must happen each frame, and yet, the performance loss is minimal. PCIe 4.0 is already obsolete, this is worst-case scenario, and the results are still pretty good. You only further proved my point. Yet you still insist 5.0 @ x16 and 6.0 @ x16 will be necessary based on these results...
    Remember, this discussion is about the needs of FUTURE slots. Not 3.0 or 4.0 slots.
    That I will admit was a mistake on my end, since I was only looking at single-channel.
     
  13. Carfax

    Carfax Ancient Guru

    Messages:
    2,915
    Likes Received:
    465
    GPU:
    NVidia Titan Xp
    I thought not. Well I have to hand it to you, you clearly like having the last word even more than myself, something that I thought was impossible!

    I already explained this in my previous post. The GPU is not reading directly from RAM. The RAM is still being used as a buffer for the GPU's VRAM, but because the PCIe bandwidth is higher and the latency lower in x8 4.0, the VRAM can be refreshed at a much faster pace.

    This illustrates the beneficial role of having a faster PCIe connection quite well, because it allows GPUs with smaller VRAM pools to achieve higher performance than they would otherwise have on a slower PCIe connection.

    On the contrary I did not prove your point. I proved DmitryKo's point about faster PCIe increasing performance when VRAM is being utilized very heavily.

    Well I'm glad you can admit you were wrong about something! :D
     
  14. Carfax

    Carfax Ancient Guru

    Messages:
    2,915
    Likes Received:
    465
    GPU:
    NVidia Titan Xp
  15. schmidtbag

    schmidtbag Ancient Guru

    Messages:
    5,748
    Likes Received:
    2,193
    GPU:
    HIS R9 290
    Indeed it does, not sure how that conflicts with my point.
    I'm well aware that if you're out of VRAM, you will need more bandwidth to not get a major performance loss. I never denied that; I didn't say his point was wrong. Getting that out of the way, yes, you did still prove my point. My point is that x16 lanes on PCIe 5.0 and 6.0 are unnecessary for consumer-grade hardware based on the trends of hardware progress and upcoming technologies. That's the only point I've been trying to make. The test results you showed prove that even in a realistic worst-case scenario, x8 lanes is sufficient even on 4.0. Bump up the lane count and you'd still see roughly the same amount of performance loss, because a GPU communicating to system memory is going to be a bottleneck.
     

  16. Celcius

    Celcius Master Guru

    Messages:
    325
    Likes Received:
    435
    GPU:
    Orchid Pro Designer
    I've read time and time again how most people are moving away from desktops and rely primarily on various mobile devices and their smart phones; an item, (my thanks to the Creator), which I've never owned. And then, there is that funky thing known as The Cloud.

    Sure, there are significant exceptions to this trend in terms of business, science, industrial and educational use, among others. But, with regard to even just PCIe 4.0 adoption, once you leave the server room behind, and then examine the typical consumer and their PC usage habits, I don't get a sense there has been a big demand from the unwashed masses for anything beyond PCIe 3.0 on the desktop. I'm not sure many people even know about PCIe 4.0. Or, maybe even what a NVMe drive is. But they are people who buy and use computers.

    I mean, currently, we have just two desktop motherboard platforms that support 4.0 natively, and neither one is likely being used in any mass-produced OEM box. (I wonder if even half of the X570, B550 owners opted for a 4.0 drive?) I actually hope that the $20,000 bleeding-edge consumer graphic cards that are waiting in the wings might actually benefit from PCIe 4.0, because without a genuine need, I doubt we're going to have the necessary surge in demand to drive down the cost of 4.0 enough to make it more commonplace.

    OK, so let's assume Intel makes good on it's PCIe 5.0 promise, and large companies order 800 trillion new Dells for their employees. Sure, while I can see them, perhaps, sporting NVMe drives, I'd be surprised if they would be anything beyond PCIe 3.0. Really, why would they need to be?
     
  17. wavetrex

    wavetrex Maha Guru

    Messages:
    1,409
    Likes Received:
    1,045
    GPU:
    Zotac GTX1080 AMP!
    (Much) Faster PCI-e can bring unified non-volatile memory design to reality.

    DDR, next gen NVMe / Xpoint, GPU RAM, they will all be seen as a big stack of data storage, with a fancy futuristic memory manager simply allocating blocks to whatever part makes sense.
    This can bring instant-on/always-on high-performance computers to life, not something relegated to low-power mobile devices.

    Imagine if you were in some big RPG game, then you hear the doorbell, the game auto-pauses when the computer senses you left, a friend is at the door, you two go out to do some stuff, come back, sit back and the computer instantly resumes where you left it, without any delay, saving, loading.

    No need to boot, no need to launch Steam, no need to start your game, nothing. Everything is just there, instantly available !

    That's the promise of this insanely fast interconnect bus.
     
    PrMinisterGR likes this.
  18. Dribble

    Dribble Member Guru

    Messages:
    196
    Likes Received:
    76
    GPU:
    Geforce 1070
    That's just sleep mode which PC's have had for many years. Don't need non-volatile memory as you can keep a little power trickling to normal memory.
     
  19. PrMinisterGR

    PrMinisterGR Ancient Guru

    Messages:
    7,652
    Likes Received:
    581
    GPU:
    Inno3D RTX 3090
    Consoles already have this, and the next gen will even allow you to pause multiple games like this.
     
  20. DmitryKo

    DmitryKo Master Guru

    Messages:
    376
    Likes Received:
    122
    GPU:
    ASUS RX 5700 XT TUF
    HBM2/GDDR6 won't become obsolete with faster and even more expensive memories. Fast VRAM has always been expensive, while system RAM is cheaper but much slower, and external storage is the cheapest and the slowest of them. New memory types won't change this balance.

    It's not like a paging file or external storage, it's just local memory for a remote processor unit (think 'far' node in NUMA).

    Using 4K textures on remote objects is not going to overwhelm the GPU, the runtime would use lower-level mipmaps. Popping could result from non-optimal texture filtering modes or texture scaling factors combined with abrupt changes in geometry LOD - and UE5 demo uses 8K textures with hundred billions of pixel-sized triangles (each model takes several dozen millions of triangles) even for the outdoor areas.

    There is no need to bother with expensive PCIe x16 switches and M.2 slots on the gaming video card when you have fast dedicated links on your CPU and motherboard.

    Dual-channel DDR4 @3200MHz is rated at 50 Gbytes/s, four times the bandwidth of PCIe 3.0 x16.
    Dual DDR5 @8000MHz would be 128 Gbytes/s, the same as PCIe 6.0 x16.
    128 bit GDDR5/6 is rated at 200-250 Gbytes/s in current mid-range videocards (GTX 1660 / RX5500XT).

    A difference between measuring the effect of faster PCIe bus, and not measuring it.

    'Limiting' is not the same as 'reducing'.


    Nothing of the above makes any sense.

    Entertain yourself with an excercise in self-reflection.
     
    Last edited: Jul 1, 2020
    Carfax and angelgraves13 like this.

Share This Page