PCIe 4.0 Specs Revealed: 16GTps Rate and Not Just For Graphics Cards Anymore (tomshardware.com) 62
Freshly Exhumed writes: PCI-SIG has released the specifications for version 4.0 of the PCIe (Peripheral Component Interconnect Express) bus, which, according to Chairman Al Yanes, promises data transfer rates of 16GTps, extended tags and credits for service devices, reduced system latency, lane margining, superior RAS capabilities, scalability for added lanes and bandwidth, improved I/O virtualization and platform integration. Tom's Hardware has posted a slide deck of the new version's specifications.
Re: (Score:3, Informative)
Giga Transfers, double version 3
Re: (Score:3)
The press release used this too, but the second slide indicates that this is equivalent to 65 GB/s
Re: (Score:2)
Gah, 64 of course.
Re: (Score:3)
The second slide clearly shows a capital B, and the figures match those on wikipedia e.;g 133 MB/s for the original PCI spec.
Re: (Score:2)
It is definitely GBytes, not Gbits. I have benchmarked CPU Memory->GPU transfers of > 100 Gbps with PCI-E 3.0
Re:16GTps? (Score:5, Informative)
GigaTransfers per second, essentially the clock speed of the bus. The bandwidth can then be found by multiplying the number of transfers per second with the bus width, and adjust for the encoding overhead (8/10 for PCIe 2.0 and earlier, 128/130 for PCIe 3.0 and above).
So, 4x (lane) PCIe 4.0 can do 16 * 4 * 128 / 130 = ~63Gb/s or ~7.9GB/s.
Wikipedia got a nice table: https://en.wikipedia.org/wiki/PCI_Express#History_and_revisions [wikipedia.org].
Re: (Score:3)
16,000,000,000 transfers per second * 1 bit per lane (in each direction) = 16,000,000,000 bits per second per lane (in each direction).
16,000,000,000 bits per second per lane (in each direction) / 130 bits per datagram =123,076,923.07692307692307692307692 datagrams per second per lane (in each direction).
123,076,923.07692307692307692307692 datagrams per second per lane (in each direction) * 128 bites payload data per datagram = 15,753,846,153.846153846153846153846 payload bits per second per lane (in each d
Re: (Score:2)
... anyone got a Tylenol?
Re: (Score:3)
I have a bottle of TylenoI capsules I bought in Chicago in 1982. I'm sure it'll be fine.
Re: 16GTps? (Score:2)
Re: (Score:2)
I'm not a fan of fucking the existing definition of kilobit/kilobyte/megabit/megabyte/etc., or approximating discrete things unnecessarily.
Re: (Score:2)
Re: (Score:1)
a solid state drive connected directly to a 16x PCIe 4.0 would be sweet, assuming that the SSD could transfer up to 7.9 GB/s x4 =31,6 GB/s
that would be faster than DDR4 SDRAM that transfers at 17066.67 MB/s or approx 17.06667 GB/s
DDR5
https://en.wikipedia.org/wiki/DDR5_SDRAM
is estimated to arrive in 2020
even 8x PCIe 4.0 or perhaps as low as 4x PCIe 4.0, might be sufficient for HP's "the machine" dream to really take off
Re: (Score:1)
16 giga-tera-phaps per second. Nearing my average too.
Need for speed... (Score:2, Funny)
You can now buy motherboards with 19 PCIe slots [amazon.com] for crypto currency mining.
Re: (Score:2)
But you have to dangle your GPU's out in a rack rather than being in a single case.
Re: (Score:3)
You can now buy motherboards with 19 x1 PCIe slots for crypto currency mining.
Need for slots w/o need for speed...
Re: (Score:1)
I only buy every 10++ yrs.
Last I bought was 2014
I'll have one HECK of a machine in 2019
Re: I love hearing about improvements! apk (Score:2)
Re:I love hearing about improvements! apk (Score:5, Insightful)
Don't worry, Microsoft will make sure new AMD CPU's aren't compatible with Windows 7, just like Intel CPU's
AMD officially don't support Windows 7 for Ryzen or Threadripper CPU's.
It's going to get harder and harder to get drivers for Windows 7 for either platform.
Re: I love hearing about improvements! apk (Score:2)
Re: M$ wants to destroy themself? Fine... apk (Score:2)
"not just for graphics cards anymore" (Score:5, Insightful)
It was never just for graphics cards. It's a replacement for the PCI bus. It just happened to also replace AGP, which was a dedicated graphics port.
It's commonly used for network cards, audio cards, storage (NVMe is PCIe), etc.
Thunderbolt is PCIe + DisplayPort
Re: (Score:3)
NVMe is not PCIe.
NVMe is implemented over PCIe. It doesn't have to be. PCIe is a physical and electrical bus with logical controllers. NVMe is a logical interface that runs over PCIe. It could run on whatever.
Re: (Score:2)
I suppose so...
NVM Express (NVMe) is an interface that allows host software to communicate with a non-volatile memory
subsystem. This interface is optimized for Enterprise and Client solid state drives, typically attached as a
register level interface to the PCI Express interface.
Re:"not just for graphics cards anymore" (Score:4, Informative)
After actually reading the NVMe spec, it's specifically designed to be used over PCI Express. Other transports look to be an afterthought. NVMe over Fabrics is a separate specification for using NVMe over non-PCI* interfaces.
NVMe 1.2 was only intended to be used over PCI type interfaces, primarily PCI Express, but also PCI and PCI-X.
Re: (Score:1)
It was never just for graphics cards. It's a replacement for the PCI bus. It just happened to also replace AGP, which was a dedicated graphics port.
It's commonly used for network cards, audio cards, storage (NVMe is PCIe), etc.
Thunderbolt is PCIe + DisplayPort
Amen to that viperidaenz... And that's on yer punny desktops and Laptops.
Please do not get me started on what can be done with PCIe on my servers, including but not limited to 40G and 100G Eth cards, 32G Fibrechannel cards, 50Gbps infiniband cards, SAS RAID Cards for booting, and mezzanine cards Galore in Blade servers....
Re: (Score:2)
Where did you get 50Gbps Infiniband cards? Valid Infiniband speeds are SDR at 10Gbps, DDR at 20Gbps, QDR at 40Gbps, FDR at 56Gbps, EDR at 100Gbps and HDR at 200Gbps.
Re: (Score:2)
It wasan FDR card, and I got the speed wrong. Since the error is close (but not exactly) 10%, is withing acceptable range ;-) :-P
Re: (Score:2)
AGP is a modified form of PCI - the only real difference is it's optimized for data transfers from the computer to the card (since video cards are output devices, there's not as much need to have a truly bidirectional, link). The other difference is it
more lanes are needed as 3.0 hardware will not (Score:2)
more lanes are needed as 3.0 hardware will not run at the same speed in a half as wide 4.0 slot then a 3.0 slot. Unless they add switchers to take 4.0 and split out 3.0 lanes.
still need cpus to add 4.0 and that will take time.
will amd and intel skip the 4.0 and go 5.0 (Score:3)
see how amd will not have an new socket till 2019-2020 any ways AMD is the king of PCI-e right now and they really don't need 4.0 or 5.0 right away.
Now intel is low on lanes and if there idea of 4.0 is just the same number of lanes then it will not really do much.
Re: will amd and intel skip the 4.0 and go 5.0 (Score:2)
Was it ever ? (Score:2)
Was it ever "just for graphics cards" though ?
Re: (Score:2)
No, but there's a whole universe of people involved with computers who got started on the gaming track and for whom the x86 platform is defined by the parts involved in gaming. Their reference point for PCIe slots is graphic cards because on most desktops in the last 10 years every other interface was integrated into the motherboard. For those people, the only apparent purpose of PCIe slots was graphics cards.
They don't remember the old days of ISA/EISA based boards where there was literally nothing integ
Re: (Score:2)
I fail to see how it still isn't just for graphic cards.
Technically you are right, it is used by just about everything, not just graphic cards. However only GPU really have any limitations on any existing bus systems in the modern past. Everything else that is attached to it is too slow to care, and is more influenced by timings and scheduling anyway. I mean I/O on even a high end SSD isn't going to really improve no matter how fast a bus you buckle it to... Although the summary does mention some other thin
Re: Was it ever ? (Score:2)
PCIe was never just for Graphics Cards (Score:2)