News PCIe 6.0 and 7.0 standards hit a roadblock — compliance slowdown could lead to broader delays

[Admin]

PCI-SIG delays conformity tests for PCIe 6.0 and PCIe 7.0, which could slowdown adoption of these technologies.

PCIe 6.0 and 7.0 standards hit a roadblock — compliance slowdown could lead to broader delays : Read more

 

[boju]

Pcie 5 is barely a thing, no graphics cards yet. I think we got time for them to sort it out.

 

[thisisaname]

Pcie 5 is barely a thing, no graphics cards yet. I think we got time for them to sort it out.

With a delay to 6 and 7, version 5 will have a chance to become more main stream before the upgrade cycle comes along. Maybe a better option would be to skip what would be 6 and release version 7 as 6. A longer cycle but bigger upgrades

 

[Eximo]

So I guess I won't be running my 800Gbps network card any time soon, darn.

 

[G.A.D]

PCI-SIG delays conformity tests for PCIe 6.0 and PCIe 7.0, which could slowdown adoption of these technologies.

PCIe 6.0 and 7.0 standards hit a roadblock — compliance slowdown could lead to broader delays : Read more

Pcie 5 is barely a thing, no graphics cards yet. I think we got time for them to sort it out.

Past PCIe 5 isn’t really for consumers. Graphic cards can’t really saturate the PCIe 5 interface with a few lanes and for some bizarre reason they continued with the 16 lane allocation which is now pointless. Data centres that require 100Gb or faster networking is where this is important.

 

[Eximo]

And yet here we are using more than 10MB of hard drive space and 640KB of memory.

Who knows how much bandwidth the average GPU will need in the future. How else are we going to run those personal holodecks?

 

[usertests]

Past PCIe 5 isn’t really for consumers. Graphic cards can’t really saturate the PCIe 5 interface with a few lanes and for some bizarre reason they continued with the 16 lane allocation which is now pointless. Data centres that require 100Gb or faster networking is where this is important.

I think PCIe 6 might not be much harder to implement than PCIe 5, with similar trace lengths and signal integrity, so it will probably come to consumers within a few years. Benefiting SSDs more than anything else since those are already close to maxing PCIe 5 (not that it could be sustained for very long or anybody actually needs it).

PCIe 7 is more questionable.

 

[Notton]

Why not use Gen6/7 as the interconnect between CPU and mobo chipset?
from there it can break out into Gen4/5 for more SSDs, or 80Gb/s USB-C

 

[DougMcC]

I think PCIe 6 might not be much harder to implement than PCIe 5, with similar trace lengths and signal integrity, so it will probably come to consumers within a few years. Benefiting SSDs more than anything else since those are already close to maxing PCIe 5 (not that it could be sustained for very long or anybody actually needs it).

PCIe 7 is more questionable.

We are far, far from having enough bandwidth to persistent storage. Desktop storage is currently 2 orders of magnitude behind being 'fast enough' for things people currently want to do, and can't keep up with TB5. PCIe 6 will catch up with TB5 just in time for TB6. So I absolutely guarantee lots of consumers are gonna want PCIe 7 and 8 to move their 8k/120 uncompressed video around.

 

[mac_angel]

It sounds like they need to come up with a universal optical interface to be able to handle beyond the boards capabilities. It's going to go that way sooner or later, it's best to come up with it sooner.
As for people thinking that GPUs don't need that bandwidth, it's not just GPUs that use PCIe; nor is the industry catering to consumers with these standards. The hot ones right now are storage, and inter-communication (Industry level ethernet, NVidia's "SLI" (yes, it's still being used, in a different capacity), etc)

 

[jackt]

good. this thing of the pcie X was out of control!
Active cooling for the ssd is madness...

 

[bit_user]

Pcie 5 is barely a thing, no graphics cards yet. I think we got time for them to sort it out.

These standards are aimed primarily at servers. They're already further along the adoption path, with AMD's MI300X, Nvidia's H200, and Intel's Datacenter GPU Max all supporting PCIe 5.0:

https://www.amd.com/en/products/accelerators/instinct/mi300/platform.html

NVIDIA H200 Tensor Core GPU

For supercharging AI and HPC workloads.

www.nvidia.com

https://www.intel.com/content/www/u...-data-center-gpu-max-1550/specifications.html
​

That's not to mention 100+ Gigabit networking cards.

https://www.colfaxdirect.com/store/pc/viewPrd.asp?idproduct=4064&idcategory=6
​

Maybe a better option would be to skip what would be 6 and release version 7 as 6. A longer cycle but bigger upgrades

Disagree. The technical hurdles and cost increases are probably better handled in the steps they defined. 6.0 adds PAM4 and other signalling changes, while 7.0 is another clock frequency doubling. You probably don't want to try and tackle PAM4 and flits while also trying to double the clock speed again.

it's not just GPUs that use PCIe; nor is the industry catering to consumers with these standards. The hot ones right now are storage, and inter-communication (Industry level ethernet, NVidia's "SLI" (yes, it's still being used, in a different capacity), etc)

Nvidia has NVLink, which is not based on PCIe (or CXL).

 

[bit_user]

I think PCIe 6 might not be much harder to implement than PCIe 5, with similar trace lengths and signal integrity,

PAM4 requires a better SNR. Multiplexing and demultiplexing also burns more power, not to mention FEC.

so it will probably come to consumers within a few years.

Only if it's to enable CXL 3.0 on consumer platforms, IMO. And the main reason for that would be to enable CXL-based memory expansion.

Benefiting SSDs more than anything else

LOL, no. Consumer SSDs don't really even need PCIe 5.0!

... or anybody actually needs it).

Yeah, it's hard to call it a benefit when you can barely even feel the difference!

 

[bit_user]

Why not use Gen6/7 as the interconnect between CPU and mobo chipset?
from there it can break out into Gen4/5 for more SSDs, or 80Gb/s USB-C

I'm with you on this general thinking, but neither Intel nor AMD has even used PCIe 5.0 for that! Intel still has PCIe 4.0 x8 and AMD has just PCIe 4.0 x4.

 

[thestryker]

PCIe 5.0 isn't particularly for consumer level products, but 6.0 and 7.0 absolutely aren't. If I had to hazard a guess the delays behind 7.0 are driven by optical interconnect mass production feasibility.

Due to Broadcom's effective murder of the high performance PCIe switch for consumers I'd like to see PCIe 5.0/6.0 x8/x4 to the chipset to enable more connectivity. Outside of that I don't see any real world advantages to PCIe 5.0+ on consumer platforms since everyone is adverse to bifurcation and there's only one PCIe 5.0 x2 SSD on the market as far as I'm aware (not that there are any motherboards providing x2 slots again see bifurcation).

 

[DavidLejdar]

Yeah, it's hard to call it a benefit when you can barely even feel the difference!

Which comes down to almost all software not making full use of it (yet). PCIe 5.0 still has double the capacity per lane, compared to PCIe 4.0, though, and apparently also a bit better latency and power efficiency.

Currently just numbers to most users, of course. Even myself, I am in no rush to get me a Gen5 NVMe, in particular as there isn't any game, which would want to load 20 GB of data every second into RAM and/or VRAM. And no noticeable difference, whether I run Win11 from a Gen4 x4 M.2 socket, or from a Gen3 x4 M.2 socket (as I currently do).

But in terms of expecting the MB to last me at least 5 years, likely upgrading in the near future to 4K (with larger size textures), I went straight with AM5 (from a DDR3 rig), to have plenty of "infrastructure capacity" for whichever next-gen GPU and possibly even 8K.

Upgrading from an older rig with classical SSD, to PCIe 4.0 with Gen4 NVMe, that still is a huge upgrade nevertheless. Up to almost like 20-times more storage read speed. Like, when a game wants to load 5 GB of data, SSD takes even more than 10 seconds, whie NVMe Gen4 can deliver in under one second.

 

[bit_user]

Upgrading from an older rig with classical SSD, to PCIe 4.0 with Gen4 NVMe, that still is a huge upgrade nevertheless. Up to almost like 20-times more storage read speed. Like, when a game wants to load 5 GB of data, SSD takes even more than 10 seconds, whie NVMe Gen4 can deliver in under one second.

Yes, SATA -> NVMe was a big jump, not to mention if you went straight to PCIe 4.0! However, the end user is going to feel less of a difference going from PCIe 4.0 -> 5.0 than they did going from 3.0 -> 4.0 (assuming a different hypothetical user, who went to PCIe 3.0, first). In both cases, there's a doubling of speed. However, the faster you make storage, the more you'd tend to become bottlenecked by CPU or other things.

 

[thestryker]

Upgrading from an older rig with classical SSD, to PCIe 4.0 with Gen4 NVMe, that still is a huge upgrade nevertheless. Up to almost like 20-times more storage read speed. Like, when a game wants to load 5 GB of data, SSD takes even more than 10 seconds, whie NVMe Gen4 can deliver in under one second.

It really doesn't in reality though so long as your SATA SSDs are good. I've had two SATA SSDs along with NVMe in my primary system for years and in practice there's no noticeable difference between games installed on SATA or NVMe.

As long as we're using NAND there won't be a notable improvement for anything other than peak transfer rates. The only NAND implementations that really change things are SLC/XL-Flash and those really only exist to fill the Optane void as the cost is high enough they'd never sell volume in consumer.

 

[bit_user]

As long as we're using NAND there won't be a notable improvement for anything other than peak transfer rates.

If you mean improvement vs. SATA, that would be incorrect. IOPS and latency are both massively improved by NVMe.

Here's an IOPS graph from one of the best SATA drives vs. QD=1 and QD=256 IOPS from one of the fastest NVMe drives:

Source: https://www.tomshardware.com/reviews/samsung-860-pro-ssd-review,5434-2.html

Source: https://www.tomshardware.com/pc-components/ssds/crucial-t705-2tb-ssd-review/2
​

So, 4k IOPS approximately doubled at QD=1 and improved 15x at max QD!

That said, if you're not running a server application of some sort, you probably don't care about that high QD IOPS number. However, a common pattern we see in software is to have a bunch of IO operations that are interspersed with computation. In such cases, the latency of the individual I/O operations can definitely limit overall performance (i.e. if the OS is unable to hide it via prefetching).

​

I didn't find separate measurement on the Samsung 860 Pro's read & write latency, but you can estimate the value for Read based on the QD=1 IOPS. So, that would be roughly 82.2 usec or just over double that of the faster NVMe drives.

 

[thestryker]

If you mean improvement vs. SATA, that would be incorrect. IOPS and latency are both massively improved by NVMe.

Here's an IOPS graph from one of the best SATA drives vs. QD=1 and QD=256 IOPS from one of the fastest NVMe drives:

​

Source: https://www.tomshardware.com/reviews/samsung-860-pro-ssd-review,5434-2.html​

​

​

Source: https://www.tomshardware.com/pc-components/ssds/crucial-t705-2tb-ssd-review/2​

​

So, 4k IOPS approximately doubled at QD=1 and improved 15x at max QD!

That said, if you're not running a server application of some sort, you probably don't care about that high QD IOPS number. However, a common pattern we see in software is to have a bunch of IO operations that are interspersed with computation. In such cases, the latency of the individual I/O operations can definitely limit overall performance (i.e. if the OS is unable to hide it via prefetching).

​

​

​

I didn't find separate measurement on the Samsung 860 Pro's read & write latency, but you can estimate the value for Read based on the QD=1 IOPS. So, that would be roughly 82.2 usec or just over double that of the faster NVMe drives.

None of this actually means anything unless your workload is dependent upon it. Switching from SATA to NVMe boot drive in the same system had no noticeable difference neither does switching between the two when it comes to games. NAND is an extremely limiting factor when it comes to the performance on most client workloads interface speed is not.

 

[bit_user]

None of this actually means anything unless your workload is dependent upon it.

QD1 latency affects things like boot time and I explained why. I'm sure we can find benchmarks, if you don't believe it.

Switching from SATA to NVMe boot drive in the same system had no noticeable difference

Qualitative assessments are hardly the gold standard.

neither does switching between the two when it comes to games.

Game loading is definitely something people benchmark. It depends on the game. Also, your antivirus scanner might be getting in the way of seeing the true capabilities of the hardware.

NAND is an extremely limiting factor when it comes to the performance on most client workloads interface speed is not.

Believe what you like. NAND certainly establishes the limits on performance, but plenty of benchmarks show a quantifiable difference between NVMe and SATA.

The past couple years, I've been doing a fair amount with docker containers at work. Some of that stuff is heavily I/O bound and I'm grateful for my PCIe 4.0 NVMe drive everytime I hit such a case. The machines we use are standard compact desktop PCs.

 

[thestryker]

QD1 latency affects things like boot time and I explained why. I'm sure we can find benchmarks, if you don't believe it.

Go for it find some benchmarks that show a notable difference not that boot time even remotely matters now that we're not talking HDDs.

Game loading is definitely something people benchmark. It depends on the game. Also, your antivirus scanner might be getting in the way of seeing the true capabilities of the hardware.

Yes they do, and here's an example of it showing no realistic difference in the majority of test cases:

Storage Game Loading Test: PCIe 4.0 SSD vs. PCIe 3.0 vs. SATA vs. HDD

What sort of storage do you need to play today's games and provide the best loading times? Do you need a PCIe drive? Do you need something...

www.techspot.com

Now maybe if DirectStorage became a real standard used across the board you'd be onto something.

Believe what you like. NAND certainly establishes the limits on performance, but plenty of benchmarks show a quantifiable difference between NVMe and SATA.

Show them then.

 

[bit_user]

Go for it find some benchmarks that show a notable difference not that boot time even remotely matters now that we're not talking HDDs.

If there's no notable difference, then you can't blame it on the NAND. I already showed where NVMe SSDs provides a 2-15x speedup over one of the fastest SATA drives, not counting sequential I/O. If that's not making a difference, then you could replace the NAND with DRAM and still not see a difference in your application.

 

[thestryker]

If there's no notable difference, then you can't blame it on the NAND. I already showed where NVMe SSDs provides a 2-15x speedup over one of the fastest SATA drives, not counting sequential I/O. If that's not making a difference, then you could replace the NAND with DRAM and still not see a difference in your application.

You know as well as I do NVMe has not significantly improved low QD performance (due to NAND) so this is a very disingenuous response. You can toss Optane in there and notice the difference because it is a significant improvement even on the slowest drives.

 

[bit_user]

You know as well as I do NVMe has not significantly improved low QD performance (due to NAND)

I literally just showed you a 2x improvement. That's full, 4k random reads, not merely an interface benchmark.

You can toss Optane in there and notice the difference because it is a significant improvement

Again, if you cannot notice the difference between 80 usec and 40 usec, then tell me how you're going to notice a difference going from 40 usec to 9 usec?? For it to be true that shaving 40 usec off read latency is unnoticeable, there must be so much CPU overhead that there's no way you're going to notice that next 31 usec!