PCIe 7.0 On Track For a 2025 Release (pcgamer.com) 29
An anonymous reader shares a PC Gamer report: PCI Express 7.0 is coming. But don't feel as though you need to start saving for a new motherboard anytime soon. The PCI-SIG has just released the 0.5 version, with the final version set for release in 2025. That means supporting devices are not likely to land until 2026, with 2027-28 likely to be the years we see a wider rollout. PCIe 7.0 will initially be far more relevant to the enterprise market, where bandwidth-hungry applications like AI and networking will benefit. Anyway, it's not like the PC market is saturated with PCIe 5.0 devices, and PCIe 6.0 is yet to make its way into our gaming PCs.
PCI Express bandwidth doubles every generation, so PCIe 7.0 will deliver a maximum data rate up to 128 GT/s. That's a whopping 8x faster than PCIe 4.0 and 4x faster than PCIe 5.0. This means PCIe 7.0 is capable of delivering up to 512GB/s of bi-directional throughput via a x16 connection and 128GB/s for an x4 connection. More bandwidth will certainly be beneficial for CPU to chipset links, which means multiple integrated devices like 10G networking, WiFi 7, USB 4, and Thunderbolt 4 will all be able to run on a consumer motherboard without compromise. And just imagine what all that bandwidth could mean for PCIe 7.0 SSDs. In the years to come, a PCIe 7.0 x4 SSD could approach sequential transfer rates of up to 60GB/s. We'll need some serious advances in SSD controller and NAND flash technologies to see speeds in that range, but still, it's an attractive proposition. Further reading: PCIe 7.0 first official draft lands, doubling bandwidth yet again.
PCI Express bandwidth doubles every generation, so PCIe 7.0 will deliver a maximum data rate up to 128 GT/s. That's a whopping 8x faster than PCIe 4.0 and 4x faster than PCIe 5.0. This means PCIe 7.0 is capable of delivering up to 512GB/s of bi-directional throughput via a x16 connection and 128GB/s for an x4 connection. More bandwidth will certainly be beneficial for CPU to chipset links, which means multiple integrated devices like 10G networking, WiFi 7, USB 4, and Thunderbolt 4 will all be able to run on a consumer motherboard without compromise. And just imagine what all that bandwidth could mean for PCIe 7.0 SSDs. In the years to come, a PCIe 7.0 x4 SSD could approach sequential transfer rates of up to 60GB/s. We'll need some serious advances in SSD controller and NAND flash technologies to see speeds in that range, but still, it's an attractive proposition. Further reading: PCIe 7.0 first official draft lands, doubling bandwidth yet again.
PCIeX (Score:2)
Re: (Score:3)
I'm way ahead of you. I'm still using PCIe3 (year 2010), so I think I will skip at least 3 generations. I bet no-one has older than that as their main gaming PC.
Re: (Score:1)
Re: PCIeX (Score:2)
There was PCI-X in 1998.
I don't think they will use an acronym so close to it. I predict we won't see PCIeX.
GT/s? (Score:2)
What is that? Giga-Terras per second?
Re: (Score:3)
GigaTransfers/sec
Don't ask me why in the world they went with that terminology.
Re: (Score:1)
DDR RAM being a prime example. It is called that because it could do two RAM things a cycle as pose to one of SD RAM.
Re: (Score:2)
You're absolutely right. PCIe6 uses PAM4 to send 4 bits per clock.
Throughput is an even better metric. PCIe6 isn't 2X the throughput of PCIe5 because of some additional overhead that PAM4 adds.
Re:GT/s? (Score:4, Informative)
GigaTransfers/sec
Don't ask me why in the world they went with that terminology.
Because everything after Gen2 is a little fuzzy with the data rate. Actually even Gen1 and Gen2 have framing overhead that is not accounted for.
Gen3 is not quite 2x Gen2 because they opted reduce the overhead instead of double the line rate. But they did not eliminate the overhead and 8Ghz can only be twice the rate of 5Ghz if the 8B/10B overhead was eliminated. 64/66b is more efficient but it isn't free.
As others have mentioned Gen6 adds further fuzziness. Two bits are transferred on every cycle with PAM4. The overhead equation has changed too. Some is reduced, some is increased, and the fixed sized flits means that the amount of overhead varies.
Re: (Score:3)
Don't ask me why in the world they went with that terminology.
Because it's a definition of the speed of a parallel bus. It's commonly used in inter chip communication terms because the the bus width becomes arbitrary as does the point at which data is decided to be sent, but we are interested in performance of the actual underlying bus itself. If you need more bandwidth you can just increase the size of the bus, or change how you transfer data on the bus.
This is important when comparing the bus of changing systems, e.g.
PCIExpress 1.0 had a bus speed of 2.5GT/s but was
Re: (Score:2)
Re: (Score:2)
Roughly it means bits per second, including overhead bits.
It means nothing of the sort. The number of bits of data you transfer is dependent on a lot of things. You already mentioned overhead but neglected to talk about bus width, which is one of the defining features of PCIExpress. That's why a 16x bus has 16x the number of bits despite having the same transfers per second.
Re: (Score:2)
Transactions per second. Its a measure of how fast it is based on a transaction, which is generally a single cycle of the bus
This is because the amount of data transferred per transaction can vary depending on whether you're using a x1 or x16 link. Sine the latter can do 16 times more data in a cycle than the former, using raw bandwidth is a bit useless.
So basically it's saying it can do 128GT/s. Each lane can do 1 bit in each direction per transaction, so a x1 will do 16GB/s (8GB/s up and 8GB/s down) in to
All that speed.. (Score:4, Funny)
..yet here I am, still typing with two fingers.
Re: (Score:2)
Or one finger/antenna on smartphones. :(
Re: (Score:1)
I wonder what is neuralink's bandwidth..
It's an interesting phenomenon.. (Score:2)
PCIe 3 stood for 7 years, then we started getting doubling every 2 or 3 years after PCIe 4 happened.
However, PCIe Gen 5 is still waiting for a lot of devices to come along, most devices decided PCIe4 was plenty fast to keep up with them, and haven't bothered targeting Gen5.
which means multiple integrated devices like 10G networking, WiFi 7, USB 4, and Thunderbolt 4
PCIe limitations were not standing in the way of any of that. You could have saturated concurrent use of two maxed out USB4 devices, unachievable
Re: (Score:2)
The lane count reduction is already starting, it's just slower than I would have expected. Low/mid range video cards already have reduced lane counts. The XBox Series X/S have a Gen4x2 SSD. Samsung released a drive recently that was either Gen4x4 or Gen5x2.
PCIe Not Keeping Up (Score:2)
It's time to put the GPU local to the system memory, using a DMA block authorization scheme. It doesn't make any sense for all the textures/information to be coming from the drive across the box--and then go again--across the bus to the video car.
Re: (Score:3)
Does anyone think that it's going to run a few RTX 4090s?
It's time to put the GPU local to the system memory, using a DMA block authorization scheme.
Yes I do, it's no where near enough to run an RTX4090. The memory on your graphics card is an order of magnitude faster still using different signalling with bandwidth now in the Terrabytes per second.
It doesn't make any sense for all the textures/information to be coming from the drive across the box--and then go again--across the bus to the video car.
That I agree with. And so does everyone else, which is precisely why the industry is moving away from using any system RAM at all and providing APIs to developers to allow the GPU to request the data directly from the drive without messing around with the system RAM or CPU at all. Look up GPU Direct Storage.
1-lane nvme (Score:2)
Basically pcie7 gives the same bandwidth through one lane that pcie5 gives through 4. It would be nice to see the ability to connect multiple nvmes at pcie5 speeds through one lane each, rather than having 4 pcie7 lanes dedicated to a single nvme that likely can't make use of the bandwidth (seeing how current pcie5 nvmes need massive heatsinks).
Generations ahead (Score:2)
Have PCIe 5 and the only thing on the market I know of are some kind of strange RAM add-on cards. If you need memory wouldn't it be better to buy what you need and have way higher bandwidth and lower latency access I guess if you need more than 2TB of RAM the cards might be a way to go but this seems fairly niche. ... and 7 is two generations still ahead of that...
Other than some kind of futuristic AI processor or GPU with little to no discrete VRAM of its own I have no clue what would even use that kind
Re: (Score:2)
Is it a replacement for (discontinued) Intel Optane nvram cards?
Re: Generations ahead (Score:2)
I remember RAM add-on cards from the 1980s. On ISA bus. You could add as much as 128KB of RAM to extend RAM from 512KB to 640KB. And all it would cost you was about $1000, due to shortages of RAM.