Horus-Anhur:

It will be saturated if AMD limits the amount of lanes. Remember that the 6600 had only 8X. And the 6500 only had 4X. I'm pretty sure we'll get a repeat of the same issues a couple of years from now. And with things like Direct Storage, having more bandwidth will improve things a lot.

I was actually considering addressing the point of reduced lanes, but that's why I said "so long as you have enough VRAM" because the 6600 (and especially 6400) didn't. Reduce texture details and AA on those GPUs (which frankly, you ought to anyway since they're not 4K capable) and their limited lanes suddenly isn't a bottleneck anymore. Alternatively, if AMD used more VRAM, that too would have reduced PCIe bandwidth. DirectStorage on PCIe 5.0 will absolutely improve things, but at the same time, it allows the GPU to feed from multiple data sources simultaneously, which means you can get by with less bandwidth. An RX 6400 with DS would most likely not have any PCIe bottleneck, even on gen 4.

DX12 and Vulkan don't reduce usage of PCI-e bandwidth. The things that do are techs tile based rendering, GPU caches, object instancing, memory compression, etc.

Yes... they do. That's one of the "selling points" of using them, in addition to reduced CPU usage. Some (but not all) of the ways this is achieved: * Fewer API calls * Asynchronous compute and data streams * Improved schedulers * As you mentioned, caches and memory optimization (both of these, if done properly, can reduce how much data needs to be transferred)

schmidtbag:

I was actually considering addressing the point of reduced lanes, but that's why I said "so long as you have enough VRAM" because the 6600 (and especially 6400) didn't. Reduce texture details and AA on those GPUs (which frankly, you ought to anyway since they're not 4K capable) and their limited lanes suddenly isn't a bottleneck anymore. Alternatively, if AMD used more VRAM, that too would have reduced PCIe bandwidth. DirectStorage on PCIe 5.0 will absolutely improve things, but at the same time, it allows the GPU to feed from multiple data sources simultaneously, which means you can get by with less bandwidth. An RX 6400 with DS would most likely not have any PCIe bottleneck, even on gen 4.

Take a look at the results in the Direct Storage benchmark. PCI-e Gen3 SSDs are already topping out at 12GB/s. Even if the GPU had more performance to give. Gen4 SSDs are close to the 24Gb/s. So with Gen5 SSDs and GPUs, tis value will be surpassed.

schmidtbag:

Yes... they do. That's one of the "selling points" of using them, in addition to reduced CPU usage. Some (but not all) of the ways this is achieved: * Fewer API calls * Asynchronous compute and data streams * Improved schedulers * As you mentioned, caches and memory optimization (both of these, if done properly, can reduce how much data needs to be transferred)

Fewer API calls, would affect the CPU mostly. At most, DX12 and Vulkan could reduce the amount of Draw Calls the CPU has to pass to the GPU. I doubt that Asynchronous compute has much of an impact on PCI-e usage. Data streams will have, especially texture streaming. Improved schedulers might have an impact, in games that use Sampler Feedback Streaming, as it would more accurately reflect the data the GPU needs to work with. But this also reduces vram usage, meaning more data can be cached on the GPU, resulting in fewer accesses to ram.

So motherboard manufacturers are anticipating selling just one AM5 board for 3ish+ years? Thus, the higher prices? Could be.

Horus-Anhur:

Take a look at the results in the Direct Storage benchmark. PCI-e Gen3 SSDs are already topping out at 12GB/s. Even if the GPU had more performance to give. Gen4 SSDs are close to the 24Gb/s. So with Gen5 SSDs and GPUs, tis value will be surpassed.

Are you referring to a synthetic test or a real-world game/application? Because in synthetic workloads, there will just about always be room for improvement. When benchmarking real game load times, there isn't a noteworthy difference going from 3.0 to 4.0. Obviously, DS works a bit differently than just simply loading game assets, but the underlying point is loading assets isn't a significant bottleneck to games using even last year's NVMe drives. Effectively doubling-up the bandwidth using on-AIB storage pushes the possibility of a bottleneck even further out. The important takeaway is whether the drives used to operate the game can offer enough bandwidth to swap out game assets on-the-fly without causing performance issues. I don't think there's enough data available at the moment to know whether a PCIe 4.0 NVMe drive will be sufficient for that, but when you consider the game can feed data from the system in parallel with DS, I suspect we'll have enough, at least for the next few years. Of course, everything gets far more complicated when you don't have enough VRAM to hold all the game assets, where even PCIe 6.0 might not be enough.

Fewer API calls, would affect the CPU mostly. At most, DX12 and Vulkan could reduce the amount of Draw Calls the CPU has to pass to the GPU. I doubt that Asynchronous compute has much of an impact on PCI-e usage. Data streams will have, especially texture streaming. Improved schedulers might have an impact, in games that use Sampler Feedback Streaming, as it would more accurately reflect the data the GPU needs to work with. But this also reduces vram usage, meaning more data can be cached on the GPU, resulting in fewer accesses to ram.

Of course most of these will have a minimal impact, my point was many of new improvements, not limited to the ones I mentioned, collectively make a noteworthy reduction on bandwidth. Remember: so long as VRAM isn't overfilled, PCIe bandwidth is already pretty minimal even on DX11/OGL. Techpowerup has tested this over and over again - take a high-end GPU and just keep slashing the lane count and you hardly see any difference. Worst-case scenario, loading screens take a little longer, but we're talking a worst-case scenario of a few additional seconds. At the consumer level, we just simply don't need that much bandwidth.

Embra:

So motherboard manufacturers are anticipating selling just one AM5 board for 3ish+ years? Thus, the higher prices? Could be.

AM5 for 3+ years sure but they'll be different chipsets that use AM5. I'm sure X770 or w/e will be out by end of next year.

schmidtbag:

Are you referring to a synthetic test or a real-world game/application? Because in synthetic workloads, there will just about always be room for improvement. When benchmarking real game load times, there isn't a noteworthy difference going from 3.0 to 4.0. Obviously, DS works a bit differently than just simply loading game assets, but the underlying point is loading assets isn't a significant bottleneck to games using even last year's NVMe drives. Effectively doubling-up the bandwidth using on-AIB storage pushes the possibility of a bottleneck even further out. The important takeaway is whether the drives used to operate the game can offer enough bandwidth to swap out game assets on-the-fly without causing performance issues. I don't think there's enough data available at the moment to know whether a PCIe 4.0 NVMe drive will be sufficient for that, but when you consider the game can feed data from the system in parallel with DS, I suspect we'll have enough, at least for the next few years.[ Of course, everything gets far more complicated when you don't have enough VRAM to hold all the game assets, where even PCIe 6.0 might not be enough.

At this point we only have synthetic benchmarks. The first game to use DS will be released next week, Forspoken. Currently we are bottlenecked by storage API and file system, not by the hardware. That's why a Sata SSD loads games almost as fast as an nvme drive. But with DS this will change.

schmidtbag:

Of course most of these will have a minimal impact, my point was many of new improvements, not limited to the ones I mentioned, collectively make a noteworthy reduction on bandwidth. Remember: so long as VRAM isn't overfilled, PCIe bandwidth is already pretty minimal even on DX11/OGL. Techpowerup has tested this over and over again - take a high-end GPU and just keep slashing the lane count and you hardly see any difference. Worst-case scenario, loading screens take a little longer, but we're talking a worst-case scenario of a few additional seconds. At the consumer level, we just simply don't need that much bandwidth.

Overall, I doubt the differences in APIs, will do much in PCI-e bandwidth usage. On GPUs that have 16X lanes, PCI-e 4 to 5, will have little to no difference, for the next years. But my point is that on cut down GPUs that have only 8X or 4X lanes, it can become an issue. Even on a low end GPU, like the 6400, changing from Gen3 to Gen4, makes a significant difference.

Horus-Anhur:

At this point we only have synthetic benchmarks. The first game to use DS will be released next week, Forspoken. Currently we are bottlenecked by storage API and file system, not by the hardware. That's why a Sata SSD loads games almost as fast as an nvme drive. But with DS this will change.

I'm not trying to be antagonistic (I'm genuinely curious) but what suggests the storage API is the issue? Many games are built to read from spinning rust or spinning plastic, which obviously doesn't require much performance tuning. However, even games that are more SSD-focused can still take a while to load because they're performing other calculations not necessarily related to loading asset data. My point is, if a game is optimized for SSDs yet a SATA drive can keep up, I find it hard to believe that both Vulkan and DX12 have the same issue and that reworking the storage API for them would fix anything. I assume this is why DS isn't necessarily just storing raw asset data.

Even on a low end GPU, like the 6400, changing from Gen3 to Gen4, makes a significant difference.

It makes a significant difference because it only has x4 lanes with 4GB of VRAM, so it basically has to feed itself from system memory every single frame. Gen 3, doesn't have enough bandwidth to even keep up with a decent single-channel DDR4 module, so of course the performance is going to suffer. That's why I was saying even with a PCIe 6.0 NVMe drive, there wouldn't be enough bandwidth to fix that problem. It's worth pointing out too that when you're out of VRAM, you're not just slowed down by the rather long pipeline reaching system memory but you waste cycles erasing and re-populating the VRAM too.

schmidtbag:

I'm not trying to be antagonistic (I'm genuinely curious) but what suggests the storage API is the issue? Many games are built to read from spinning rust or spinning plastic, which obviously doesn't require much performance tuning. However, even games that are more SSD-focused can still take a while to load because they're performing other calculations not necessarily related to loading asset data. My point is, if a game is optimized for SSDs yet a SATA drive can keep up, I find it hard to believe that both Vulkan and DX12 have the same issue and that reworking the storage API for them would fix anything. I assume this is why DS isn't necessarily just storing raw asset data.

DirectStorage overview

schmidtbag:

It makes a significant difference because it only has x4 lanes with 4GB of VRAM, so it basically has to feed itself from system memory every single frame. Gen 3, doesn't have enough bandwidth to even keep up with a decent single-channel DDR4 module, so of course the performance is going to suffer. That's why I was saying even with a PCIe 6.0 NVMe drive, there wouldn't be enough bandwidth to fix that problem. It's worth pointing out too that when you're out of VRAM, you're not just slowed down by the rather long pipeline reaching system memory but you waste cycles erasing and re-populating the VRAM too.

Exactly, that's why having Gen5 now is important for the future. What is happening today with Gen3, will happen a few years from now, but with Gen4.

Horus-Anhur:

DirectStorage overview

Skimming through the Existing Issues section, seems to me it's hardly an API problem. Rather, it seems to me MS is trying to use DS as a way to compensate for either their own negligence or poor development practices of game devs. If you care to dive into it: High CPU usage, as the article suggests, can be alleviated using more cores. While DS absolute could help alleviate this, it's treating the symptom and not fixing the cause, because either: A. It requires developer intervention to use, in which case, why couldn't the developers just make the game use more cores when handling data? B. It doesn't require developer intervention to use, in which case, there is a slew of ways to take advantage of untapped system resources. For instances of insufficient bandwidth, that makes a lot of sense for DS (and if bottlenecks are removed then disk bandwidth demands will go up). The inability to prioritize disk requests is another bogus point of treating-the-symptom. If this problem only exists with DX12 (not sure if that's the case) then MS needs to get their act together and fix how the base code functions so all products made with DX12 can improve, regardless of DS. If this problem exists in Vulkan too then that suggests this isn't an API issue. In that case, it either means the game developer is the problem (not unreasonable to assume) but also possibly an OS issue. After all, MS is awful at making CPU schedulers, and NTFS is archaic enough to be the only modern filesystem that still needs defragging (it's just not worth the write cycles to do it on an SSD), so I wouldn't at all be surprised if Windows is terrible at prioritizing disk requests too. Assuming game developers aren't the problem, that means MS ought to be fixing their other products before they use DS to duct tape this problem. If Windows and DX12 aren't to blame, then DS is a good solution because there's not much MS can do about crappy 3rd party devs. The inability to cancel disk requests is also something that DS is not necessary to fix. Again, that's something that either DX12, the OS, or game devs should fix. If canceling disk requests is a common enough issue to noticeably impact game performance, devs should reconsider how they go about loading assets. I could potentially see how DS could assist with hardware-accelerated decompression, except when you consider 99% of the time, the GPU is working super hard while the CPU almost always has cycles to spare. So if anything, DS might actually exacerbate this problem since it means more load on the GPU. Despite everything I said, I do strongly believe DS is a good technology that absolutely will improve performance.

Exactly, that's why having Gen5 now is important for the future. What is happening today with Gen3, will happen a few years from now, but with Gen4.

What difference does it make though? No matter what, you're going to get crap results, so why bother upgrading when it's still a sub-par experience? You are far better off using a GPU with another couple GB of VRAM than to get a motherboard with the next gen of PCIe. Even with PCIe 5.0 @ x16, an RX 6400 is still going to be bottlenecked, because at that point system memory isn't fast enough (remember - the system memory needs enough bandwidth for everything else the PC is doing too, not to mention the additional CPU load). Even with a 4090, PCIe 2.0 @ x16 is only 8% slower than 4.0. A 4090 is fast enough that even the CPU can get bottlenecked at times, yet, the PCIe spec from 2007 can keep up reasonably well. https://www.techpowerup.com/review/nvidia-geforce-rtx-4090-pci-express-scaling/28.html TL;DR: If you have enough VRAM, the PCIe bandwidth hardly matters - 4.0 will be all we'll need for many years to come. If the GPU is really starving for VRAM (like the 6400) then more/faster lanes isn't going to fix the problem.

Undying:

But you will need it in the future. Thats why smart buyers think ahead. Next gen nvme drives and gpus will use pcie5.

By the time you need PCIe5.0 you'll be upgrading all your PC again. Most people don't need it for the next 5 years. GPUs still run fine at PCIe3.0x16 I don't see the point of PCIe5.0. You still don't have games that really push direct-storage and you're forced to use PCIe5.0 SSDs. No, we don't need it. It's just marketing.

Horus-Anhur:

That is true today. But AM5 is supposed to last 5 years. That is one of the key points AMD made about the platform. A couple of years from now, we'll have Gen5 as standard on all GPUs. And let us remember that AMD likes to cut on PCI-e lanes on mid and low range GPUs. So having Gen5 will become very important.

AM4 was also supposed to last for so many years and first came with PCIe3.0 and then upgraded to 4.0. I don't see why AM5 can't have PCIe4.0 and 5.0 boards if that makes the adoption more affordable and faster. AMD is losing here. GPUs still work with older gens, you don't have to worry. The performance difference between PCIe3.0x8 and PCIe4.0x8 is in single % digits, I doubt you'd notice any difference between PCIe4.0x8 and PCIe5.0x8 for a long long time.

Undying:

Motherboards are the real issue building a am5 system. I remember when i bought x570 for 200$. Those ware the times amd was getting some ground and now when they have it prices skyrocketed.

I have a GB b550 aorus pro and it's a very solid board. Paid around the equivalent of 140-150$ USD for it.