News PCI-SIG announces PCIe 8.0 spec with twice the bandwidth — 1TB/s of peak bandwidth, 256 GT/s per lane, and a possible new connector

[Admin]

PCI-SIG has officially begun developing PCIe 8.0, doubling bandwidth to 256 GT/s while tackling signal integrity limits over copper with protocol upgrades.

PCI-SIG announces PCIe 8.0 spec with twice the bandwidth — 1TB/s of peak bandwidth, 256 GT/s per lane, and a possible new connector : Read more

 

[Jame5]

I know it's an extra cost, but what about optical between the controller and the slot? It's only a few inches replaced, but it would help.

 

[logainofhades]

Not sure we should trust them with a new connector

 

[chaz_music]

Even with PAM4, this is solidly in high frequency RF signaling. This might be the end of backwards compatible PCIe slots. I could see a connector scheme with shielded coaxial interconnects, or as others said using optical instead of RF over copper.

 

[S58_is_the_goat]

Gigabyte: best we can do is pcie4...

 

[jp7189]

Im pretty sure the use of pam4 rules out optical and strongly indicates copper.

 

[bit_user]

The upcoming PCIe 8.0 specification will double the raw bit rate of PCIe 7.0 to 256.0 GT/s, enabling up to 1 TB/s of bi-directional bandwidth across a x16 configuration.

For the love of... !

Please stop parroting this BS marketing number. The far more useful number is 512 GB/s. If someone has a PCIe 8.0 SSD on a x4 link, the bandwidth of the other direction will not help increase their read or write speeds! Okay, sure there are mixed workloads, but most are either predominantly reading or writing, and the numbers drive makers tout are certainly uni-directional numbers.

Most of the time, you're bottlenecked on one direction or the other. This is why the bandwidth usually quoted is the uni-directional speed, even in a full-duplex link. In gaming, nearly all the data transfer volume is from host -> GPU. The fact that the link is bidir and symmetrical is almost inconsequential.

For example, we say that Ethernet is 1 Gigabit or 10 Gigabit. Just because it happens to (normally) be full-duplex, we don't then say it's 2 Gigabit or 20 Gigabit!

@Paul Alcorn Can the new Toms Hardware please add this to your style guide?
: )

 

[bit_user]

I know it's an extra cost, but what about optical between the controller and the slot? It's only a few inches replaced, but it would help.

I can't find the articles, but IIRC there's both a working group for routing existing PCIe signals over optical links and another working group exploring post-PCIe optical interconnects.

Here's a proprietary effort at routing standard PCIe 7.0 over optical:

https://www.tomshardware.com/tech-industry/optical-pcie-70-connection-hits-a-blazing-128-gts
​

 

[bit_user]

Not sure we should trust them with a new connector

Even if it's the only option? There might be no way to meet the electrical requirements of PCIe 8.0 with the existing connector.

Just because it's PCIe doesn't mean we need to have the same edge connector we know and love(?). M.2 and U.2 are two examples of PCIe signals using different connectors. The fact that the signal is exactly the same means you can do weird hacks, like mounting M.2 SSDs on a cheap PCIe add-in card or even having a M.2 board that can host a PCIe add-in card.

 

[bit_user]

Gigabyte: best we can do is pcie4...

To be fair, that was only on B650 boards. Many other B650 boards never even gave you the option to enable PCIe 5.0, such as an ASUS board I have.

 

[logainofhades]

Gigabyte: best we can do is pcie4...

99% of all PC users don't need more than that, and many of those don't even need Gen 4.

Even a SATA ssd is fast enough for general PC usage.

 

[logainofhades]

Even if it's the only option? There might be no way to meet the electrical requirements of PCIe 8.0 with the existing connector.

They approved the melty connector, so I can't trust them to get this right either.

 

[hotaru251]

Not sure we should trust them with a new connector

don't want you pcie catching fire and melting?

and seriously only benefit to new pcie gen anymore is it allows you to bifurcation the slow and add more fast prior gen storage w/o losing performance.

 

[bit_user]

They approved the melty connector, so I can't trust them to get this right either.

Oh, I guess I missed that you were referring to the ill-fated 12vhpwr connector.

 

[DS426]

For the love of... !

Please stop parroting this BS marketing number. The far more useful number is 512 GB/s. If someone has a PCIe 8.0 SSD on a x4 link, the bandwidth of the other direction will not help increase their read or write speeds! Okay, sure there are mixed workloads, but most are either predominantly reading or writing, and the numbers drive makers tout are certainly uni-directional numbers.

Most of the time, you're bottlenecked on one direction or the other. This is why the bandwidth usually quoted is the uni-directional speed, even in a full-duplex link. In gaming, nearly all the data transfer volume is from host -> GPU. The fact that the link is bidir and symmetrical is almost inconsequential.

For example, we say that Ethernet is 1 Gigabit or 10 Gigabit. Just because it happens to (normally) be full-duplex, we don't then say it's 2 Gigabit or 20 Gigabit!

@Paul Alcorn Can the new Toms Hardware please add this to your style guide?
: )

I strongly agree with you on this, @bit_user. Aggregrating bi-directional bandwidth goes against the standard of measurement that our industry uses for bus performance.

Moreover, the fact that PCIe operates in full duplex (or really dual simplex) means lanes are dedicated to one direction and thus can't be repurposed to aggregate more bandwidth in a single direction like USB 4.0 can.

 

[DS426]

99% of all PC users don't need more than that, and many of those don't even need Gen 4.

Even a SATA ssd is fast enough for general PC usage.

Exactly. Gen 5 is notably more expense to manufacture than Gen 4, so unless it's absolutely needed (professional workstations, maybe the very best E-ATX gaming and creator boards, etc.), it's relatively cost-prohibitive for your average consumer hardware. In the server world, it's all the rage and sorely needed, which is where this contract will only grow more stark as new PCIe standards really only apply to servers and pro workstations. This means consumer hardware will be sitting on the same PCIe standards for longer. PCIe 4.0 is already fast enough that AMD and nVidia are using x8 bifurcation on x16 slot GPUs to actually take advantage of the available bandwidth, e.g. for a SSD or two that's mounted on the GPU.

I think PCI-SIG will need to fork PCIe at some point -- probably when they go optical -- to segment an economical consumer side of PCIe while allowing the bleeding-edge side of PCIe to continue moving forward for servers. Anyone remember PCI-X? It wasn't typically available on consumer motherboards as it wasn't practical while making sense in the server world.

 

[ExiIe]

Exactly. Gen 5 is notably more expense to manufacture than Gen 4, so unless it's absolutely needed (professional workstations, maybe the very best E-ATX gaming and creator boards, etc.), it's relatively cost-prohibitive for your average consumer hardware. In the server world, it's all the rage and sorely needed, which is where this contract will only grow more stark as new PCIe standards really only apply to servers and pro workstations. This means consumer hardware will be sitting on the same PCIe standards for longer. PCIe 4.0 is already fast enough that AMD and nVidia are using x8 bifurcation on x16 slot GPUs to actually take advantage of the available bandwidth, e.g. for a SSD or two that's mounted on the GPU.

I think PCI-SIG will need to fork PCIe at some point -- probably when they go optical -- to segment an economical consumer side of PCIe while allowing the bleeding-edge side of PCIe to continue moving forward for servers. Anyone remember PCI-X? It wasn't typically available on consumer motherboards as it wasn't practical while making sense in the server world.

PCIe 1.0 finalized 2003, first chipset 2004 (915P)
PCIe 2.0 finalized 2007, first chipset 2007 (X38)
PCIe 3.0 finalized 2010, first chipset 2011 (Z68)
PCIe 4.0 finalized 2017, first chipset 2019 (X570)
PCIe 5.0 finalized 2019, first chipset 2021 (Z690)
PCIe 6.0 finalized 2022, first chipset 2028?
PCIe 7.0 finalized 2025, first chipset 2033?
PCIe 8.0 finalized 2028, first chipset 2038?

The first five PCIe versions reached mainstream products within 0–2 years of release, but PCIe 6.0 is lagging behind. An RTX 5090 shows a 1–4% performance drop when limited to PCIe 4.0, while the RTX 5060 can lose up to 10%. We are already benefitting from 5.0. Storage devices have maxed out PCIe 5.0 bandwidth, PCIe 6.0 could bring meaningful gains there as well.

 

[DS426]

...

The first five PCIe versions reached mainstream products within 0–2 years of release, but PCIe 6.0 is lagging behind. An RTX 5090 shows a 1–4% performance drop when limited to PCIe 4.0, while the RTX 5060 can lose up to 10%. We are already benefitting from 5.0. Storage devices have maxed out PCIe 5.0 bandwidth, PCIe 6.0 could bring meaningful gains there as well.

Noted.

That's fine and dandy, I'm just saying that PCIe is getting more expensive to develop and manufacture. Folks assuming that they are going to get gobs of the latest PCIe lanes on their motherboards well into the sunset of the future are probably going to be mistaken, again, unless one is willing to pay the price premium. Except for PCIe 5.0 x4 for an SSD, even the latest AMD B850 chipset doesn't have 5.0 x16 slots. That's a mainstream chipset.

 

[bit_user]

An RTX 5090 shows a 1–4% performance drop when limited to PCIe 4.0,

Where is that 4% even coming from? TechPowerUp tested 25 games at 3 different resolutions and here's what they found:

"We are happy to report that performance loss in this mode is well contained, and you lose about 1% performance, across all three resolutions. There are barely any outliers to report about from our set of game tests."

Source: https://www.techpowerup.com/review/nvidia-geforce-rtx-5090-pci-express-scaling/33.html
​

while the RTX 5060 can lose up to 10%.

More nonsense.

"Averaged across all game tests, we see a 2% drop in FPS at 1080p, and an even smaller 1% drop at 1440p. While not exactly relevant to this GPU, even 4K Ultra HD posts only a 1% drop in performance with PCIe Gen 4 x8."

Source: https://www.techpowerup.com/review/nvidia-geforce-rtx-5060-ti-pci-express-x8-scaling/31.html
​

We are already benefitting from 5.0. Storage devices have maxed out PCIe 5.0 bandwidth,

They barely max it out and it took until the first half of 2023 for the first PCI 5.0 to even reach the market! That's almost 2 years for SSDs, and 4 years for GPUs.

That should tell you just how premature it was to bring PCIe 5.0 to the consumer market. Intel didn't do it because it made sense, they did it to leap-frog AMD, after being embarrassed by AMD getting to 4.0 first.

Most people do not benefit from such speeds, either. Because Operating Systems do caching and read-ahead, you rarely ever see a perceivable difference between PCIe 5.0 SSDs and decent 4.0 models.

 

[thisisaname]

Seeing it widely adopted around 2040 or maybe later?

 

[ExiIe]

Where is that 4% even coming from? TechPowerUp tested 25 games at 3 different resolutions and here's what they found:

​

"We are happy to report that performance loss in this mode is well contained, and you lose about 1% performance, across all three resolutions. There are barely any outliers to report about from our set of game tests."​

​

Source: https://www.techpowerup.com/review/nvidia-geforce-rtx-5090-pci-express-scaling/33.html​

​

More nonsense.

​

"Averaged across all game tests, we see a 2% drop in FPS at 1080p, and an even smaller 1% drop at 1440p. While not exactly relevant to this GPU, even 4K Ultra HD posts only a 1% drop in performance with PCIe Gen 4 x8."​

​

Source: https://www.techpowerup.com/review/nvidia-geforce-rtx-5060-ti-pci-express-x8-scaling/31.html​

​

They barely max it out and it took until the first half of 2023 for the first PCI 5.0 to even reach the market! That's almost 2 years for SSDs, and 4 years for GPUs.

That should tell you just how premature it was to bring PCIe 5.0 to the consumer market. Intel didn't do it because it made sense, they did it to leap-frog AMD, after being embarrassed by AMD getting to 4.0 first.

Most people do not benefit from such speeds, either. Because Operating Systems do caching and read-ahead, you rarely ever see a perceivable difference between PCIe 5.0 SSDs and decent 4.0 models.

I posted the google ai results with no research.
The 1-4% was from https://gamersnexus.net/gpus/nvidia-rtx-5090-pcie-50-vs-40-vs-30-x16-scaling-benchmarks
It looks like the "up to 10%" is for the 8GB 5060Ti. https://www.tomshardware.com/pc-com...to-10-percent-performance-when-using-pcie-4-0

 

[thestryker]

More nonsense.

"Averaged across all game tests, we see a 2% drop in FPS at 1080p, and an even smaller 1% drop at 1440p. While not exactly relevant to this GPU, even 4K Ultra HD posts only a 1% drop in performance with PCIe Gen 4 x8."

Source: https://www.techpowerup.com/review/nvidia-geforce-rtx-5060-ti-pci-express-x8-scaling/31.html

I think this is when you factor in running out of VRAM as well. Being on the edge of VRAM makes PCIe bandwidth matter significantly more. That's why of the "bad 8GB VRAM" cards the 9060XT is the best performing in VRAM limited situations.

 

[JamesJones44]

I posted the google ai results with no research.
The 1-4% was from https://gamersnexus.net/gpus/nvidia-rtx-5090-pcie-50-vs-40-vs-30-x16-scaling-benchmarks
It looks like the "up to 10%" is for the 8GB 5060Ti. https://www.tomshardware.com/pc-com...to-10-percent-performance-when-using-pcie-4-0

The difference for the 5060 was on average 6% (59.4 vs 55.8 respectively), most of the titles are within a small percentage of each other, a few titles seem to be the outliers, Dragon Age: The Veilguard and Indiana Jones and the Big Circle had very big drops.

Could be more going on there than just bandwidth, but without more info it's hard to say.

 

[thestryker]

I could see a situation where PCI-SIG potentially goes with optical for the data portion of PCIe. Retimers/redrivers are often required already and they can only get you so far. Assuming optical can be produced at scale without requiring full retooling for motherboard manufacturers I wouldn't even be surprised if it appeared in high end client.

PCIe 4.0 showed everyone the mess that can be with riser cables and OCuLink. PCIe 5.0 isn't really any better and from what I understand 6.0s new signaling largely just means is no more difficult to do right than 5.0. The way things have been going it's hard to see 7.0 let alone 8.0 working it's way down to client.

 

[Alvar "Miles" Udell]

Backwards compatibility is great, but at some point you just have to cut it loose and move on for the future. The last time this happened was when AGP was replaced by PCIe, now it's time for PCIe copper to be replaced by PCIe optical. Given the prices of modern GPUs most people are on more than a 1 generation refresh cycle, and given the capabilities of modern systems there's no reason to upgrade every generation, so the actual impact to most people will be negligible,