Are we even close to saturating PCIe 3.0 yet?

BLEH!:

Are we even close to saturating PCIe 3.0 yet?

For current-gen AMD GPUs, probably not. But, I imagine Volta GPUs demand more bandwidth than 3.0 @ x8 can offer. EDIT: As far as I'm concerned, the real benefit of higher-bandwidth PCIe lanes comes down to everything that uses fewer than x4 lanes (like M.2, x1 slots, Thunderbolt, etc). We need more bandwidth there; at this point, we could probably ditch x16 slots because I'm not sure we're really ever going to see hardware (of the same generation) take advantage of those ever again.

BLEH!:

Are we even close to saturating PCIe 3.0 yet?

It's not only about saturating but efficiency too. Current SSD can already max PCIe3.0 4x for example. Business needs this badly to improve performance. As for us consumers, CPUs have a limited number of lanes because they're expensive to make. So, if we need less lanes to power the same things, we get cheaper products or more feature rich products for the same price! Currently some boards with 2 nvme slots have one running at 3.0 and the other at 2.0 through the chipset. An SSD connected through the chipset will have its performance degraded by half.

schmidtbag:

For current-gen AMD GPUs, probably not. But, I imagine Volta GPUs demand more bandwidth than 3.0 @ x8 can offer. EDIT: As far as I'm concerned, the real benefit of higher-bandwidth PCIe lanes comes down to everything that uses fewer than x4 lanes (like M.2, x1 slots, Thunderbolt, etc). We need more bandwidth there; at this point, we could probably ditch x16 slots because I'm not sure we're really ever going to see hardware (of the same generation) take advantage of those ever again.

Gen 4 X 4 lanes could handle just about anything we need. It could allow for standard 4 X risers in cases to more conveniently mount GFX cards.

BLEH!:

Are we even close to saturating PCIe 3.0 yet?

Seeing as both the DMI interface on Intel, and the AMD equivalent, both use effectively PCIe 3.0 x4 interfaces, yes is the answer. 🙂 Having a PCIe 4.0 x4 interface between the CPU and the chipset will be a godsend to motherboard makers everywhere.

schmidtbag:

For current-gen AMD GPUs, probably not. But, I imagine Volta GPUs demand more bandwidth than 3.0 @ x8 can offer. EDIT: As far as I'm concerned, the real benefit of higher-bandwidth PCIe lanes comes down to everything that uses fewer than x4 lanes (like M.2, x1 slots, Thunderbolt, etc). We need more bandwidth there; at this point, we could probably ditch x16 slots because I'm not sure we're really ever going to see hardware (of the same generation) take advantage of those ever again.

PCIe can't come soon enough. On both sides of barricade southbridges are limited by PCIe lanes invested. On AMD side it is pretty severe limitation. Some Disabled SATA slots in case you use M.2. And that M.2 only in PCIe 2.0 mode anyway. Sometimes disabled PCIe slot if you use M.2 slot. ASUS has one middle of the road solution, where 2nd M.2 slot is not on SB, but shares lanes with Graphics. Therefore it runs in 3.0 x8 mode moment you occupy 2nd M.2 slot. PCIe 4.0 instantly removes all those compromises. With PCIe 5.0, you could almost feel that it delivers sufficient bandwidth for 2 GPUs to share VRAM without crazy penalty. Or have PCIe 5.0 x2 to some external connector (Like USB 3.0) delivering same bandwidth as PCIe 2.0 x16. Only thing I was disappointed with Ryzen was that AMD did not skip PCIe 3.0.

nosirrahx:

Gen 4 X 4 lanes could handle just about anything we need. It could allow for standard 4 X risers in cases to more conveniently mount GFX cards.

Oh whoops I meant x4 lanes or lower. Overall I agree, but I think there are some situations where a next-gen GPU for example will saturate x4 gen 4 lanes.

Fox2232:

On AMD side it is pretty severe limitation. Some Disabled SATA slots in case you use M.2.

That problem is due to lack of lanes, not the generation. Regardless - I agree that is a bad limitation, especially considering there are only 4x SATA ports to begin with (I have 8-year-old motherboards with 6x SATA III ports, with no caveats).

And that M.2 only in PCIe 2.0 mode anyway. Sometimes disabled PCIe slot if you use M.2 slot.

I thought the primary M.2 slot was 3.0? Either way, it doesn't appear there are any major performance losses. However, benchmarks show that M.2 SSDs tend to perform a little faster on Intel. Part of me wonders is this is due to the higher clock speeds, though. As for the disabled PCIe slot, that again comes down to number of lanes, not the generation.

Only thing I was disappointed with Ryzen was that AMD did not skip PCIe 3.0.

I personally think 3.0 was perfectly fine - what disappointed me was how few lanes there were, and, how much it depends on PCIe 2.0. I was really hoping the 400 series chipsets would've addressed one of these issues. But... my Ryzen rig is mini ITX and I don't have any M.2 drives, so, these problems never affected me.

Silva:

It's not only about saturating but efficiency too. Current SSD can already max PCIe3.0 4x for example. Business needs this badly to improve performance. As for us consumers, CPUs have a limited number of lanes because they're expensive to make. So, if we need less lanes to power the same things, we get cheaper products or more feature rich products for the same price! Currently some boards with 2 nvme slots have one running at 3.0 and the other at 2.0 through the chipset. An SSD connected through the chipset will have its performance degraded by half.

Install into a intel system with spectre and meltdown patch enabled, problem solved. :P

BLEH!:

Are we even close to saturating PCIe 3.0 yet?

as others have said that part doesn't matter, i think only a few cards actually saturate PCIe 2.0 x16 and even then its within a small margin or error in many cases, its more about allowed PCIe to do more with m.2 and addon cards becoming more of a thing having your GPU only need PCIe x4 and still work 100% leaving the other 12 we used to need to be able to do other things is a dream. Also means if you want to run sli you wont need to worry about bandwidth if you also want a m.2 and a sound card for example

as far as I know there has not been ANY GPU that saturates pci-e spec 2 @ x16 (they are perfectly content being x8 mode unless drivers or cruddy implementation from SLI/CF profile, bad mobo or whatever) pci-e 3.0 has more "available" because of overhead reduction, so a x4 spec 3 is pretty smack on identical in real world performance to spec 2 x8, 3.0 x8 is better than spec 2 x16. 3+ GB/s is well beyond graphics cards needs UNLESS very very specific AI/machine learning/compute monsters that are well beyond "normal" ranges. keep in mind, just because a gpu has say 1024bit memory running "effective 500+ GB/s this does NOT mean it is sending hundreds of gb per second over the pci-e bus (most of this is internalized usage) this is why for example ASICS can get away using such small amounts of actual data compared to the amount of "hashing" they are actually doing which can range in the hundreds or thousands of GB/s. SSD/m.2/NVME and such on the other hand can be severely limited by having "only" 3 odd gb/s available for their usage especially if they drop from spec 3 x4 to spec 2 levels (motherboard limitation/implementation) So, to say AMD/Nv GPU "claiming" they saturate any spec beyond spec 1 @ x8 is nonsense unless you are in a SLI/CF situation where that small "margin of error" becomes no longer imperceptible, but for a single card usages or for 99% of "gaming machines" there is no real current limitation for them to get the speed from spec 2x8 or spec 3x8, the "machine learning" ones such as Volta or whatever I honestly do NOT see spec 3x8 being any more a limitation than any other graphics card out there,

Dragonstongue:

as far as I know there has not been ANY GPU that saturates pci-e spec 2 @ x16 (they are perfectly content being x8 mode unless drivers or cruddy implementation from SLI/CF profile, bad mobo or whatever) pci-e 3.0 has more "available" because of overhead reduction, so a x4 spec 3 is pretty smack on identical in real world performance to spec 2 x8, 3.0 x8 is better than spec 2 x16.

Whether it's because of the overhead reduction or because of the additional bandwidth, there are some significant performance improvements in gen 3.0 vs 2.0. Techpowerup has an article proving this with a GTX 1080.

3+ GB/s is well beyond graphics cards needs UNLESS very very specific AI/machine learning/compute monsters that are well beyond "normal" ranges.

Not necessarily. Some GPUs have plenty of processing power but not enough VRAM, where you need a LOT of bandwidth to access system RAM. As for non-gaming tasks, they might need a lot of bandwidth to fill up the VRAM, but otherwise in a lot of cases, the GPUs are just left churning data on their own, just sipping bandwidth as they go. But, I am aware this isn't true of all compute scenarios - some do demand chronic high bandwidth.

SSD/m.2/NVME and such on the other hand can be severely limited by having "only" 3 odd gb/s available for their usage especially if they drop from spec 3 x4 to spec 2 levels (motherboard limitation/implementation)

As stated earlier in this thread, it is devices like these (among others) which are very demanding of more bandwidth in fewer lanes. These are what push the demand of newer gens of PCIe.

So, to say AMD/Nv GPU "claiming" they saturate any spec beyond spec 1 @ x8 is nonsense unless you are in a SLI/CF situation where that small "margin of error" becomes no longer imperceptible, but for a single card usages or for 99% of "gaming machines" there is no real current limitation for them to get the speed from spec 2x8 or spec 3x8, the "machine learning" ones such as Volta or whatever I honestly do NOT see spec 3x8 being any more a limitation than any other graphics card out there,

I don't think either AMD or Nvidia claim anything about their products' bandwidth needs. But it's important to keep in mind why GPUs support all x16 lanes, despite the fact that most motherboards of their time don't need all of those lanes. The reason ultimately comes down to future-proofing. An overclocked 1080Ti might run perfectly fine on a 3.0 @ x8 slot, but will probably be bottlenecked on a 2.0 @ x8 slot. However, a 2.0 @ x16 slot should offer plenty of bandwidth. It's also important to keep in mind that the inverse does NOT apply. Hypothetically speaking, let's say you have a PCIe 2.0 GPU that saturates more than x8 lanes. If you put that GPU in a gen 3.0 slot (regardless of lane count), you aren't going to get more bandwidth. Remember, eGPUs are just starting to become a viable option. If you have a gen 4.0 or even 5.0 slot, you can operate a modestly powerful GPU with no major performance loss on x2 lanes, as long as you are running a display through the GPU itself. So, getting GPUs on newer gens of PCIe is important, not for x16 slots, but for everything else. We're reaching a point where x16 slots might become totally obsolete.

Dragonstongue:

to say AMD/Nv GPU "claiming" they saturate any spec beyond spec 1 @ x8 is nonsense

TPU has shown a GTX 1080 on PCIe 1.1x8 is 13% slower than same card on PCIe 3.0x8, so I wouldn't say it's nonsense. As was stated, the immediate benefit will be for PCIe storage devices. A friend recently built a server, and ran into PCIe 3.0 limitations when it came to creating an array of U.2 drives. The options are not so great right now. Multiple RAID cards is one option, but many servers don't even have one x16 electrical slot. VROC, PCIe switches and retimers is another, but it has issues too. PCIe 4.0 boards are available, so I say bring on the cards!

BLEH!:

Are we even close to saturating PCIe 3.0 yet?

depends on the workload, Support for pcie gen4 will mainly benefit compute type workloads, graphics workloads, will not benefit very much, maybe 1-2% in most cases, maybe if you were pushing an 8k monitor and using more than 8gb of vram on an rx vega card, you might see a difference

I'm pretty sure modern PCs still use bellow architecture. https://en.wikipedia.org/wiki/Video_card#/media/File:Desktop_computer_bus_bandwidths.svg With RAM hitting easily 40GB/s and VRAM being about 10 times faster on high end cards, I can see how PCIe 3.0 x16 running at 16GB/s can be bottleneck for coping data from RAM to VRAM. This is a typical workload for computing applications designed to off-load work to GPU e.g.: using CUDA. I'm not sure how modern game engines work, but I think they load textures first to RAM and then copy them to VRAM. So at the very least it could reduce that operation cost / time. VRAM speed even many times faster than RAM is still a limiting factor in computing operation for GPU, so the sooner you can get data to/from there the better.

I’m actually more interested in the storage speed advancements with PCIe 4.0. Double the current limit means even faster boot times and game loading times. I can’t wait for that though!

RooiKreef:

I’m actually more interested in the storage speed advancements with PCIe 4.0. Double the current limit means even faster boot times and game loading times. I can’t wait for that though!

Not being rude but do boot times really matter anymore? I run HP TUNERS software for my 1999 LS1 Camaro and the BESTBUY brand tablet/laptop has an ATOM CPU & 2GB RAM and Windows 10 loads within 10 secs. My office PC has an AMD PHENOM 965BE @4.2GHZ and 8GB ddr2 RAM & 1TB Mechanical hard drive and WINDOWS 7 Loads in 15 secs.

Boot times still depend a lot on CPU speed as well. The main reason SSDs are so much faster is because of the accesstimes which are hugely better than mechanical drives. I had a cheap laptop for a few weeks through work and it had a 128GB SSD and it did not boot so fast because of the cheap Celeron CPU. I'm not really excited about anything atm, I hope for 7nm GPUs next year.