Will Ryzen 4000 CPUs have more PCIE lanes?

[jigger]

It shouldn't be hard to understand. I need more lanes, without compromise, than is available from the current 'enthusiast' AM4 platform. I've said this more than once.

In reality it doesn't matter what i'm using and it's not me forcing the artificial limitations....the motherboards are. The PCIe bandwidths are not a substitute for lanes in all circumstances although people seem set on constantly conflating them. Your SM/enterprise-class example clearly shows that you gained benefit from PCIe4 because it unlocked the bandwidth you required. This isn't my use case. I require lanes and not bandwidth and PCIe4.0 will not change this.

I mentioned to you i have other >x1 cards and the reasons why it is not optimal to use. I'll humour you by letting you know that, among other things, i use a Perc H710 and as i alluded to i compromise my GPU by using it.



You had a bandwidth deficit here. I don't. This is an apples/oranges comparison so while it was interesting for me to read, offered nothing to my situation.



AMD designed a new IO chip, of course they could have added another 8 lanes while they were at it. Unless you have first-hand knowledge of what it took to design the IO die and why 8 lanes couldn't be added i'll file this statement under 'speculation'. What still remains is the fact they missed an opportunity to drive differentiation.



Harsh but true. Again, total bandwidth is not a substitute for available lanes (and vice versa). I also didn't lament having 990FXA for AM4, i lamented the flexibility of the large lane allocation. AMD could easily have done that while implementing PCIe3.0. It's a bit of a nonsense to bring PCIe2.0 up as if that's what i wanted.

At this point, PCIe4.0 is not a reason to buy for most people. Should AMD have brought out PCIE4.0 and more lanes than X370/X470 then i'd have bought, even at PCIe3.0. Not for the bandwidth, for the lanes. If AMD had brought out X470 but with an uplift from X370 in PCIe lanes then i'd have bought (where i didn't). See where i'm going with this? I don't think i'm unique in this regard.

You could use bifurcation and chop up a 16x slot.

 

[Journey]

It shouldn't be hard to understand. I need more lanes, without compromise, than is available from the current 'enthusiast' AM4 platform. I've said this more than once.

So AM4 isn't for you then. If you really need the lanes then they come with a premium price, just like multi-socket boards do, or anything else that takes you out of the 'normal' user, and pushes you up to the actual 'enthusiast' or professional market products.

AMD designed a new IO chip, of course they could have added another 8 lanes while they were at it. What still remains is the fact they missed an opportunity to drive differentiation.

I agree, they made a 3-4 year plan when they implemented the Zen lineup on the AM4 socket, aiming to add compatibility where they could along the way, even when that meant having imperfect solutions in place for adding newer CPU's to current boards, but at least trying to off the end user something rather than nothing at all. Maybe if they hadn't made TR4 they would have made AMX4 or similar, and added something but this would be HEDT, rather than SHEDT which is what people are seeing with the current TRX40 platform.

Silly thing is the Ryzen TR 1950X was ~£899 and you could get a nice board for £250, that had 16c/32t etc. and loads of I/O available, and if you wanted less CPU, but the same I/O they even had the 8c/16t 1900X at £499 retail (or less). I'm not sure how much cheaper they could have made the platform to attract users like yourself to it. Funny thing is you can bag a 1900X for ~£130 brand new today, and there are TR4 boards floating around for a reasonable price still.

AMD could easily have done that while implementing PCIe3.0. It's a bit of a nonsense to bring PCIe2.0 up as if that's what i wanted.

I brought it up since you mention your GPU's being held back by losing 8x lanes to another device. I assume you are running RTX 2080 Ti's or equivalent professional grade cards, but losing 3-5% performance by using 8x lanes rather than 16x isn't ideal, but is hardly a deal breaker unless you are using it in a mission critical or time sensitive application.

 

[Quartz]

It doesn't work like that. The CPU would have to be designed for 36 or however many lanes.

Intel CPUs have managed it.

 

[Quartz]

Ti's or equivalent professional grade cards, but losing 3-5% performance by using 8x lanes rather than 16x isn't ideal, but is hardly a deal breaker unless you are using it in a mission critical or time sensitive application.

And how much performance do you think the RTX 3000 GPUs will lose? And Sony have shown us the benefits of full PCIE v4 bandwidth with the PS5.

 

[Journey]

And how much performance do you think the RTX 3000 GPUs will lose?

Well I'd hope that Nvidia will take the sensible approach and offer cards that work either 8x PCI-E 4.0 or 16x PCI-E 3.0 depending on the board they are in, especially since the next generation Intel platform is PCI-E 4.0 ready as well. So either nothing, or more than now but it depends on the performance.

And Sony have shown us the benefits of full PCIE v4 bandwidth with the PS5

What have they shown us that we didn't already know?

We already knew that a 4x 4.0 = 8x 3.0, and we already know that is the same 7.88GB/s maximum theoretical throughput. 'Full' PCI-E 4.0 bandwidth isn't achieved with a 4x slot btw.

 

[Quartz]

We already knew that a 4x 4.0 = 8x 3.0, and we already know that is the same 7.88GB/s maximum theoretical throughput. 'Full' PCI-E 4.0 bandwidth isn't achieved with a 4x slot btw.

It can be after you process the data. You're not streaming straight from the NVME drive to the GPU; you're doing it via the CPU. Processing is involved. Note that Sony did a comparison with PCIe v3 and that fell way behind.

 

[Journey]

It can be after you process the data. You're not streaming straight from the NVME drive to the GPU; you're doing it via the CPU. Processing is involved. Note that Sony did a comparison with PCIe v3 and that fell way behind.

So they showed that PCI-E 3.0 is slower than 4.0 using 4 lanes, well that would be as it is half the bandwidth at 3.94 vs 7.88 GB/s. I am not sure where you are going with this for use on the PC.

 

[jigger]

Intel CPUs have managed it.

Not really. Most are 16 lanes.

 

[Aretak]

It's the 5500XT that gets bottlenecked if its 4GB of VRAM gets maxed out. I'm presuming PCIE 4.0 allows it to swap out data faster. Hardware Unboxed tested it. It's all a bit academic but it does look interesting in view of what we know about the PS5 and the use of fast storage. Is that why AMD went PCIE 4.0 so early?

It's not even PCIe 3.0 that's the issue with the 5500 XT. It's the fact that the card is only physically wired for an x8 connection as a cost-saving measure. With a PCIe 4.0 board, an x8 link provides the same relative bandwidth as an x16 3.0 one, but that doesn't mean that if you stick it in a motherboard that only supports 3.0 that you can still get that same x16-equivalent bandwidth, even in a PCIe slot wired for x16 3.0. You'll only be getting x8 3.0, because the card only has enough physical connections to address eight PCIe lanes, no matter what speed those lanes are running at. PCIe lanes aren't an abstract concept that can simply be reallocated at will. The physical components have to be there.

Intel CPUs have managed it.

None of Intel's mainstream CPUs offer more than 16 PCIe lanes, up to and including the 10900K. Their HEDT products do, but then so do AMD's. For Intel, all of their HEDT CPUs are simply repurposed Xeon silicon with bits lasered off to fit the product stack, so unlocking extra PCIe lanes for a product refresh as they did with Skylake-X is simply a matter of lasering less bits off. You can't restore stuff that was never there to begin with though, and it wasn't ever there on the mainstream dies, because those aren't repurposed Xeons.

 

[Augustus]

So AM4 isn't for you then. If you really need the lanes then they come with a premium price, just like multi-socket boards do, or anything else that takes you out of the 'normal' user, and pushes you up to the actual 'enthusiast' or professional market products.

Correct. AMD regressed their enthusiast desktop platform in terms of flexible PCIe lanes and i've been working around it ever since. The gap to Threadripper is too large to offer value to me so really AMD created a segment gap that i fell into. I'm sure others fell into it and have been 'managing' like me but in their own way with their own use cases.

I disagree that my need should come with a price. I don't need that much and all i'm asking is the flexibility i had with 990FXA. As i said before, if AMD wanted to add NVME after retaining PCIe lane parity then yes i'd pay for it. I wouldn't be paying over the odds for it though. As it is, i don't have the option and i buy the premium option motherboards already.

I agree, they made a 3-4 year plan when they implemented the Zen lineup on the AM4 socket, aiming to add compatibility where they could along the way, even when that meant having imperfect solutions in place for adding newer CPU's to current boards, but at least trying to off the end user something rather than nothing at all. Maybe if they hadn't made TR4 they would have made AMX4 or similar, and added something but this would be HEDT, rather than SHEDT which is what people are seeing with the current TRX40 platform.

Silly thing is the Ryzen TR 1950X was ~£899 and you could get a nice board for £250, that had 16c/32t etc. and loads of I/O available, and if you wanted less CPU, but the same I/O they even had the 8c/16t 1900X at £499 retail (or less). I'm not sure how much cheaper they could have made the platform to attract users like yourself to it. Funny thing is you can bag a 1900X for ~£130 brand new today, and there are TR4 boards floating around for a reasonable price still.

I almost went TR4. I was glad i didn't when the platform didn't last long. Unfortunately TR4 CPUs proved to be just that bit compromised in their operation to justify the price. sTRx is a whole other situation though and their price reflects it. For very specific workloads TR4 CPU looks immense value at the moment. I have too general needs to make it a value buy.

I don't think AMD even twigged they created a segment gap and so i think what we got (AM4,TR4/sTRX4) what they planned and wouldn't have gotten something in the middle. Really, you've nailed what AMD should have done but didn't. TR is already SHEDT but the fact they call TR HEDT shows they don't think they have a gap. They could have closed that gap from X470 onwards but show no signs of doing so. AMD could/should be giving people tangible reasons for people to upgrade but when it comes to the motherboards they aren't really and it all feels forced and awkward.

I brought it up since you mention your GPU's being held back by losing 8x lanes to another device. I assume you are running RTX 2080 Ti's or equivalent professional grade cards, but losing 3-5% performance by using 8x lanes rather than 16x isn't ideal, but is hardly a deal breaker unless you are using it in a mission critical or time sensitive application.

So i should be ok with buying an expensive graphics card and consciously giving up on performance? I thought we were supposed to be enthusiasts here.

 

[Augustus]

You could use bifurcation and chop up a 16x slot.

Which x16 slot do you suggest i chop up?

 

[Rroff]

None of Intel's mainstream CPUs offer more than 16 PCIe lanes, up to and including the 10900K. Their HEDT products do, but then so do AMD's. For Intel, all of their HEDT CPUs are simply repurposed Xeon silicon with bits lasered off to fit the product stack, so unlocking extra PCIe lanes for a product refresh as they did with Skylake-X is simply a matter of lasering less bits off. You can't restore stuff that was never there to begin with though, and it wasn't ever there on the mainstream dies, because those aren't repurposed Xeons.

I haven't checked all AMD options but with Intel you can get a foot in the door of the HEDT platform for around the same price as a higher end mainstream setup and get 40+ PCIE lanes.

(This is basically what I've done with most of my recent setups - with the 4820K instead of 4770K, etc.).

 

[BongoHunter]

When Ryzen first came out Threadripper was the obvious solution to this. Not so obvious what to do now the new 3rd gen TRs are here at at much higher price point (though they are much better chips than 1st or 2nd gen)

X570 is the only real option it appears - it might take a lot of thought on the best way to arrange the devices though - and running GFX at x8 would probably be sane if the requirement for expansion is super important.

 

[sandys]

I haven't checked all AMD options but with Intel you can get a foot in the door of the HEDT platform for around the same price as a higher end mainstream setup and get 40+ PCIE lanes.

I considered that but its performance was barely any better than my 2nd gen TR and i'd be relying on a healthy overclock to be a stronger CPU, reviews suggested they were not the best clockers, so I left it.

X570 is the only real option it appears - it might take a lot of thought on the best way to arrange the devices though - and running GFX at x8 would probably be sane if the requirement for expansion is super important.

Certainly what I'll be considering once some PCIe4 GPUs drop, though I was hoping 4Tb NVMe might be cheaper by now so I would need less lanes.

 

[Grim5]

It is something I find annoying with CPUs these days - hence why I went with X79/4820K on my last build for the 40 PCI-e lanes.

(I'm currently using 28 for the record).

EDIT: Don't forget in many cases you get an extra 16-20 lanes provided by the PCH or PLX, etc. on many boards though that isn't necessarily a good solution depending on what you want.

does Threadripper 3000 not have more lanes than rtzen?

 

[BongoHunter]

does Threadripper 3000 not have more lanes than rtzen?

Yeah it does - loads more

 

[Aretak]

I haven't checked all AMD options but with Intel you can get a foot in the door of the HEDT platform for around the same price as a higher end mainstream setup and get 40+ PCIE lanes.

(This is basically what I've done with most of my recent setups - with the 4820K instead of 4770K, etc.).

The problem is, X299 just isn't very good. It would be madness to buy it over Z490 for most people. There's a lot more compromise involved than there was with X79 or X99, which were excellent platforms that just plain outperformed the mainstream equivalents in many scenarios (and pretty much all scenarios by now, given the rise in multi-core support even for gaming). The cheapest CPU you can buy for X299 with 44 PCIe lanes is the 9800X, which is only an 8-core and comes in at around £400, or slightly more than the faster 10700K. If you want a 10-core then it's the same price as a 10900K, yet again markedly inferior for most applications. Unless you have a very specific workload that wants AVX-512 or the bandwidth of quad-channel memory, it's just not a platform that makes any sense. It's a budget option for getting more PCIe lanes relative to a Zen 2-based Threadripper, but Threadripper also has a lot more PCIe lanes and CPU power, even with the base 24-core. You have to have very specific needs indeed for X299 to make sense. And a complete disregard for your electricity bill.

 

[Quartz]

Linus Tech Tips has a video giving an example of why more PCIe lanes are needed:

 

[arconz]

Thanks for providing info on this - its a board I'm considering for my new post-production rig. I will be getting an RTX 3080 when its released, and the rig will have 2 Gen4 NVME drives, maybe 3TB total. I may want to add a third Gen4 NVME at some point. In my situation does the Asus X570 WS Pro-Ace have any advantages over a standard board? I am also considering the Gigabyte X570 Auros Extreme as I really like the passively cooled chipset.

 

[Journey]

Linus Tech Tips has a video giving an example of why more PCIe lanes are needed:

Yes, more PCI-E lanes are needed if you want to run 4x NVMe M.2 drives, we've already covered that. That isn't a standard desktop requirement though.