I believe existing TB3 controllers also have a problem with uneven performance when the 2nd TB3 port in a controller is used to connect regular USB devices. The 22 Gbps limit, the USB connection stability issue, USB 3.2 2x2 compatibility, there are definitely useful improvements that TB4 can hopefully address even if the theoretical bandwidth isn't increasing.
I guess we'll probably have to wait for Intel PCH to adopt PCIe 4.0 before Thunderbolt gets more raw bandwidth, which could take a while seeing their CPU's don't even support PCIe 4.0 yet.
PCIe 4.0 has much stricter signaling requirements, so a hypothetical 80 Gbps Thunderbolt based on PCIe 4.0 would require super expensive cables. Their maximum length would also need to be shorter, so it might make more sense to use a fiber optics cable rather than copper. The problem with fiber is that it cannot provide power though.
Nothing prevents you from having a hybrid cable (optical for data and copper for power). Even on TB3 power is carried by seperate wires. The big issue with optical is cost. Assumming you are using standard usb-c connector it means an electrical to optional converter in the cable at both ends. That is going to mean really expensive cables. However to go to 80+ Gbps you might not have much choice.
Active cables during TB2 ended up being a sore point for the suppliers, AFAIK. The photonics were only provided by Intel, and had a high enough failure rate that nobody went on to develop TB3 optical cables.
I somewhat wonder who supplied the original joint venture (Light Peak) with Sony. Did Sony have photonics back then? Did Intel?
Light Peak / Thunderbolt never used silicon photonics. The optical versions were based on conventional 850nm VCSELs from Oclaro, PIN photodiode receivers from Enablence, embedded optical engine from Avago, optical module by SAE, driver and receiver silicon from IPtronics, transceiver IC from Ensphere, and multi-mode fiber and waveguides from FOCI.
Thunderbolt optical cables were developed using more or less off-the-shelf optical transceivers and fiber because they only needed to support 10.3125 GT/s NRZ signaling, which is rather common. Thunderbolt 3 would require transceivers capable of at least 20.625 and 10.3125 GT/s modes that also fit within the footprint and power budget of a Type-C cable connector. I'd be very surprised if anyone bothered to develop optical transceivers specifically for Thunderbolt 3 applications because the addressable market is tiny (1000's of units).
There is zero chance of that happening. One reason TB3 is at least marginally adopted is it is a usb-c port. So it is usb AND TB3. You can plug a TB3 device in or a usb device in. A special optical port only used by Thunderbolt isn't happening. Hell it is hard enough to get OEMs to keep widely used ports like HDMI and ethernet on laptops they aren't putting a special TB only optical port.
As much as I like TB, it's a dead standard. Even if Intel increments it it's going nowhere.
If the USB forum wasn't so brain dead with their naming conventions and inability to force the manufacturers to a clear naming scheme TB would have died quicker than it did. Honestly anything with a licensing fee as big as Intel wanted for TB is a dead standard when USB3 is good enough. TB never saw anything other than niche use, and with it's guts contributed to USB-IF it might see broad adoption in 10 years after USB fully integrates it in the next version but otherwise it's not going anywhere. It just can't compete against USB when USB chips can be had for pennies.
Intel's foolish if they keep spending money developing TB hoping it will go somewhere.
Apple alone buys ~$150M / yr. in Thunderbolt controllers, which is probably more than Intel makes off of client Wi-Fi solutions, so there's that. If you can make decent money serving the niche applications that USB can't, why wouldn't you?
At some point soon, the ports themselves will just have to become optical (at least for the data-carrying component, apart from any "hybridized" power delivery.) That'd make for cheaper cables...
Post-PCIe4, the bus itself might have to switch to an optical implementation. Hopefully, silicon photonics will have finally reached a sufficient level of maturity, power efficiency, and manufacturing readiness a few years from now (there's not much time left before we start hitting serious roadblocks in power dissipation and performance, with current tech...)
You don't need silicon photonics for 80gbps. For 100 gbps Ethernet we already have very cheap VCSEL (4x25 as cwdm or multiple fiber cable). 100gbps transceiver are available for US$ 99.00 now with all the overkill that Ethernet transceivers have (like management I2C).
Thunderbolt is 2x20 though, and can't easily go wider. 40 Gbit/s over copper is totally doable, at a lower cost and using less power than optical transceivers for the sub 5m lengths that are required, but PAM4 and FEC are almost certainly going to be necessary.
PCIe Gen3 x4 is four lanes at 8 GT/s with 128b/130b encoding leaving ~31.5 Gbit/s for the upper layers. Like all packet based protocols, there is also protocol overhead. With a TLP maximum payload size of 128 B (which is what Thunderbolt uses), the upper limit for throughput once you remove all the framing, packet headers, and error correction code is ~27.25 Gbit/s or 3406 MB/s. That's for 100% maximum payload size packets without any retransmits or intervening DLLPs, so real-world throughput will necessarily be lower.
Thunderbolt storage benchmarks top out around 22.5 Gbit/s or 2812 MB/s which is 17.4% lower than the theoretical maximum. Looking through the USB4 specification, my best guess is that the observed discrepancy is due to the considerable latency involved in Thunderbolt links combined with I/O stalls resulting from the inability of the PCIe lane adapters to keep a 40 Gbit/s channel full short of stuffing it with dummy packets.
Also, the total number of bytes being transmitted over the Thunderbolt link is slightly higher, as each tunneled PCIe packet includes an additional 4 byte header.
Thunderbolt 3 is 40 Gbit/s per link provided to the upper layers, full-stop. It doesn't care what protocol the packets are. The discrete controllers to date have only had at most a PCIe Gen3 x4 back end to the host. That is what limits PCIe bandwidth (along with a bit of additional overhead due to the Thunderbolt meta-protocol and channel-bonding). There are also up to 2x DisplayPort 1.4 sinks, which can be used by far more devices than the PCIe functions alone, especially on a controller with two 40 Gbit/s ports.
Thunderbolt signaling rates have nothing to do with the signaling rates of the PCIe back end.
Tiger Lake does have integrated Thunderbolt as well as PCIe Gen4 support, but usually Intel does not increment the Thunderbolt version number unless the front-end capabilities change. So this could mean USB 3.2 Gen 2 x 2 support and full USB4 certification.
STOP IT. We have USB 4.0. Both in the consumer and enterprise world this has been an utter disaster. We have Dell D6000 USB-C docks that handle DUAL 4K60 monitors no problem on PC but refuse to work on Mac because it isn't Thunderbolt. Then we have the new Caldigit USB-C Pro Dock (which is also Thunderbolt 3) that works fine on Macs and PCs with 4 lanes PCIe but not PCs with a Thunderbolt port limited to 2 lane PCIe (even though those same computers work just fine with 10 Gbps on a USB-C dock). And the USB-C mode on the Caldigit dock only allows 4k60. This confusion branding nonsense was supposed to end with Thunderbolt 3 finally dying and we move on to USB 4. We don't need to bring Thunderbolt 4 back from the dead.
Thunderbolt 3 is not dead; it's inclusion w/ USB4 IS what makes USB4 interesting + its side-effect of making it royalty free. It brings features USB simply does not have like daisy chaining.
Thunderbolt 4 is much needed since the core features of Thunderbolt are in need of a speed upgrade that many pros + Apple will strongly appreciate.
"Intel did confirm that that they were referencing USB 3.2 Gen 2 - the 10 Gbps version - in the keynote presentation. This means that the peak speeds (40 Gbps) are not changing relative to Thunderbolt 3." In other words there is no point for Thunderbolt 4. If it's intended for "certification of certain USB-C ports" why do they need to change TB3 to TB4?
As the TB3 spec is open, other silicon vendors (like JMicron and ASMedia) can also implement it and be the host for certain Type-C ports. TB4 certification will only be for ports enabled by an Intel controller - now, what the extra features offered by the TB4 ports are - we don't know yet. Only thing we know as of now is that peak bandwidth - 40 Gbps - is not changing.
And note that USB4 uses slightly different signaling rates and encoding schemes than Thunderbolt 3 does.
USB4 specifies Gen 2 signaling at 10 GT/s with 64b/66b or RS(198,194) encoding, or Gen 3 signaling at 20 GT/s with 128b/132b or RS(198,194) encoding, with support for Reed-Solomon forward error-correction code (RS-FEC) being optional.
Thunderbolt uses Thunderbolt/Thunderbolt 2 signaling at 10.3125 GT/s with 64b/66b encoding, or Thunderbolt 3 signaling at 20.625 GT/s with 64b/66b encoding, with channel-bonding being optional.
Wait, what? If so, I stand corrected: Thunderbolt 4 is pointless. It needs to be significantly faster than Thunderbolt 3 to align w/ Displayport 2.0 & HDMI 2.1 bandwidth increases
Thunderbolt 3 is 4 lanes at 20.625 Gbit/s, DisplayPort 2.0 is 4 lanes at 20 Gbit/s. Although a Thunderbolt link can only transport up to 40 Gbit/s of a DP 2.0 link at current rates because it is bi-directional, it could still easily support a native DP 2.0 link at 80 Gbit/s. This is another possible feature of Thunderbolt 4, support for native DP 2.0. Although I kinda doubt that made it into Gen12 / Xe graphics.
As for HDMI 2.1, it's only 42.66 Gbit/s max, so unless you have a device that requires more than 40 Gbit/s of display data but less than 42.66, you'd be fine with Thunderbolt 3. And Intel is finally touting native HDMI 2.0b with Tiger Lake, so also probably not an issue as far as they're concerned.
Because the modus operandi of new Intel is to announce products ever further in the future...
“Will there be a TB4 one day, say maybe in 2028?” “Perhaps.” “OK then, let’s announce it as CES along with all our other nonsense, like Foveros 2 (before v1 has even shipped) and EMIB 2, and our ten year fab “roadmap” “.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
29 Comments
Back to Article
jeremyshaw - Tuesday, January 7, 2020 - link
TB3 is currently around 22Gbps data, with the rest wasted on a DP allocation that cannot be repurposed, even when there is no video stream whatsoever.Does TB4 change this relationship?
ltcommanderdata - Tuesday, January 7, 2020 - link
I believe existing TB3 controllers also have a problem with uneven performance when the 2nd TB3 port in a controller is used to connect regular USB devices. The 22 Gbps limit, the USB connection stability issue, USB 3.2 2x2 compatibility, there are definitely useful improvements that TB4 can hopefully address even if the theoretical bandwidth isn't increasing.I guess we'll probably have to wait for Intel PCH to adopt PCIe 4.0 before Thunderbolt gets more raw bandwidth, which could take a while seeing their CPU's don't even support PCIe 4.0 yet.
Santoval - Tuesday, January 7, 2020 - link
PCIe 4.0 has much stricter signaling requirements, so a hypothetical 80 Gbps Thunderbolt based on PCIe 4.0 would require super expensive cables. Their maximum length would also need to be shorter, so it might make more sense to use a fiber optics cable rather than copper. The problem with fiber is that it cannot provide power though.TheUnhandledException - Tuesday, January 7, 2020 - link
Nothing prevents you from having a hybrid cable (optical for data and copper for power). Even on TB3 power is carried by seperate wires. The big issue with optical is cost. Assumming you are using standard usb-c connector it means an electrical to optional converter in the cable at both ends. That is going to mean really expensive cables. However to go to 80+ Gbps you might not have much choice.brakdoo - Tuesday, January 7, 2020 - link
I don't think he meant active cables. The optical transceiver should be in your laptop and the end device (like monitor).jeremyshaw - Tuesday, January 7, 2020 - link
Active cables during TB2 ended up being a sore point for the suppliers, AFAIK. The photonics were only provided by Intel, and had a high enough failure rate that nobody went on to develop TB3 optical cables.I somewhat wonder who supplied the original joint venture (Light Peak) with Sony. Did Sony have photonics back then? Did Intel?
repoman27 - Tuesday, January 7, 2020 - link
Light Peak / Thunderbolt never used silicon photonics. The optical versions were based on conventional 850nm VCSELs from Oclaro, PIN photodiode receivers from Enablence, embedded optical engine from Avago, optical module by SAE, driver and receiver silicon from IPtronics, transceiver IC from Ensphere, and multi-mode fiber and waveguides from FOCI.Thunderbolt optical cables were developed using more or less off-the-shelf optical transceivers and fiber because they only needed to support 10.3125 GT/s NRZ signaling, which is rather common. Thunderbolt 3 would require transceivers capable of at least 20.625 and 10.3125 GT/s modes that also fit within the footprint and power budget of a Type-C cable connector. I'd be very surprised if anyone bothered to develop optical transceivers specifically for Thunderbolt 3 applications because the addressable market is tiny (1000's of units).
TheUnhandledException - Tuesday, January 7, 2020 - link
There is zero chance of that happening. One reason TB3 is at least marginally adopted is it is a usb-c port. So it is usb AND TB3. You can plug a TB3 device in or a usb device in. A special optical port only used by Thunderbolt isn't happening. Hell it is hard enough to get OEMs to keep widely used ports like HDMI and ethernet on laptops they aren't putting a special TB only optical port.rahvin - Tuesday, January 7, 2020 - link
As much as I like TB, it's a dead standard. Even if Intel increments it it's going nowhere.If the USB forum wasn't so brain dead with their naming conventions and inability to force the manufacturers to a clear naming scheme TB would have died quicker than it did. Honestly anything with a licensing fee as big as Intel wanted for TB is a dead standard when USB3 is good enough. TB never saw anything other than niche use, and with it's guts contributed to USB-IF it might see broad adoption in 10 years after USB fully integrates it in the next version but otherwise it's not going anywhere. It just can't compete against USB when USB chips can be had for pennies.
Intel's foolish if they keep spending money developing TB hoping it will go somewhere.
repoman27 - Tuesday, January 7, 2020 - link
Apple alone buys ~$150M / yr. in Thunderbolt controllers, which is probably more than Intel makes off of client Wi-Fi solutions, so there's that. If you can make decent money serving the niche applications that USB can't, why wouldn't you?boeush - Tuesday, January 7, 2020 - link
At some point soon, the ports themselves will just have to become optical (at least for the data-carrying component, apart from any "hybridized" power delivery.) That'd make for cheaper cables...Post-PCIe4, the bus itself might have to switch to an optical implementation. Hopefully, silicon photonics will have finally reached a sufficient level of maturity, power efficiency, and manufacturing readiness a few years from now (there's not much time left before we start hitting serious roadblocks in power dissipation and performance, with current tech...)
brakdoo - Tuesday, January 7, 2020 - link
You don't need silicon photonics for 80gbps. For 100 gbps Ethernet we already have very cheap VCSEL (4x25 as cwdm or multiple fiber cable). 100gbps transceiver are available for US$ 99.00now with all the overkill that Ethernet transceivers have (like management I2C).
repoman27 - Tuesday, January 7, 2020 - link
Thunderbolt is 2x20 though, and can't easily go wider. 40 Gbit/s over copper is totally doable, at a lower cost and using less power than optical transceivers for the sub 5m lengths that are required, but PAM4 and FEC are almost certainly going to be necessary.repoman27 - Tuesday, January 7, 2020 - link
Thunderbolt 3 is *already* operating 29% faster than PCIe Gen4 (20.625 GT/s vs 16 GT/s). 80 Gbit/s Thunderbolt will require PAM4 signaling.crimsonson - Tuesday, January 7, 2020 - link
I think you mean "32 Gb" and not 22 Gb. 8Gb is reserved for video signal, inaccessible to data transport.repoman27 - Thursday, January 9, 2020 - link
Nothing is reserved for anything.PCIe Gen3 x4 is four lanes at 8 GT/s with 128b/130b encoding leaving ~31.5 Gbit/s for the upper layers. Like all packet based protocols, there is also protocol overhead. With a TLP maximum payload size of 128 B (which is what Thunderbolt uses), the upper limit for throughput once you remove all the framing, packet headers, and error correction code is ~27.25 Gbit/s or 3406 MB/s. That's for 100% maximum payload size packets without any retransmits or intervening DLLPs, so real-world throughput will necessarily be lower.
Thunderbolt storage benchmarks top out around 22.5 Gbit/s or 2812 MB/s which is 17.4% lower than the theoretical maximum. Looking through the USB4 specification, my best guess is that the observed discrepancy is due to the considerable latency involved in Thunderbolt links combined with I/O stalls resulting from the inability of the PCIe lane adapters to keep a 40 Gbit/s channel full short of stuffing it with dummy packets.
Also, the total number of bytes being transmitted over the Thunderbolt link is slightly higher, as each tunneled PCIe packet includes an additional 4 byte header.
repoman27 - Tuesday, January 7, 2020 - link
Thunderbolt 3 is 40 Gbit/s per link provided to the upper layers, full-stop. It doesn't care what protocol the packets are. The discrete controllers to date have only had at most a PCIe Gen3 x4 back end to the host. That is what limits PCIe bandwidth (along with a bit of additional overhead due to the Thunderbolt meta-protocol and channel-bonding). There are also up to 2x DisplayPort 1.4 sinks, which can be used by far more devices than the PCIe functions alone, especially on a controller with two 40 Gbit/s ports.Thunderbolt signaling rates have nothing to do with the signaling rates of the PCIe back end.
Tiger Lake does have integrated Thunderbolt as well as PCIe Gen4 support, but usually Intel does not increment the Thunderbolt version number unless the front-end capabilities change. So this could mean USB 3.2 Gen 2 x 2 support and full USB4 certification.
quiksilvr - Tuesday, January 7, 2020 - link
STOP IT. We have USB 4.0. Both in the consumer and enterprise world this has been an utter disaster. We have Dell D6000 USB-C docks that handle DUAL 4K60 monitors no problem on PC but refuse to work on Mac because it isn't Thunderbolt. Then we have the new Caldigit USB-C Pro Dock (which is also Thunderbolt 3) that works fine on Macs and PCs with 4 lanes PCIe but not PCs with a Thunderbolt port limited to 2 lane PCIe (even though those same computers work just fine with 10 Gbps on a USB-C dock). And the USB-C mode on the Caldigit dock only allows 4k60. This confusion branding nonsense was supposed to end with Thunderbolt 3 finally dying and we move on to USB 4. We don't need to bring Thunderbolt 4 back from the dead.quiksilvr - Tuesday, January 7, 2020 - link
And the USB-C mode on the Caldigit dock only allows 4k60 on a SINGLE monitor, not dual.lilkwarrior - Tuesday, January 7, 2020 - link
Thunderbolt 3 is not dead; it's inclusion w/ USB4 IS what makes USB4 interesting + its side-effect of making it royalty free. It brings features USB simply does not have like daisy chaining.Thunderbolt 4 is much needed since the core features of Thunderbolt are in need of a speed upgrade that many pros + Apple will strongly appreciate.
timecop1818 - Tuesday, January 7, 2020 - link
Yeah but nobody cares what doesn't work on a fucking mac, lol.Santoval - Tuesday, January 7, 2020 - link
"Intel did confirm that that they were referencing USB 3.2 Gen 2 - the 10 Gbps version - in the keynote presentation. This means that the peak speeds (40 Gbps) are not changing relative to Thunderbolt 3."In other words there is no point for Thunderbolt 4. If it's intended for "certification of certain USB-C ports" why do they need to change TB3 to TB4?
ganeshts - Tuesday, January 7, 2020 - link
As the TB3 spec is open, other silicon vendors (like JMicron and ASMedia) can also implement it and be the host for certain Type-C ports. TB4 certification will only be for ports enabled by an Intel controller - now, what the extra features offered by the TB4 ports are - we don't know yet. Only thing we know as of now is that peak bandwidth - 40 Gbps - is not changing.repoman27 - Tuesday, January 7, 2020 - link
The USB4 spec is available from the USB-IF, however, interoperability with Thunderbolt 3 is an optional feature and may not be widely supported.repoman27 - Tuesday, January 7, 2020 - link
And note that USB4 uses slightly different signaling rates and encoding schemes than Thunderbolt 3 does.USB4 specifies Gen 2 signaling at 10 GT/s with 64b/66b or RS(198,194) encoding, or Gen 3 signaling at 20 GT/s with 128b/132b or RS(198,194) encoding, with support for Reed-Solomon forward error-correction code (RS-FEC) being optional.
Thunderbolt uses Thunderbolt/Thunderbolt 2 signaling at 10.3125 GT/s with 64b/66b encoding, or Thunderbolt 3 signaling at 20.625 GT/s with 64b/66b encoding, with channel-bonding being optional.
lilkwarrior - Tuesday, January 7, 2020 - link
Wait, what? If so, I stand corrected: Thunderbolt 4 is pointless. It needs to be significantly faster than Thunderbolt 3 to align w/ Displayport 2.0 & HDMI 2.1 bandwidth increasesrepoman27 - Tuesday, January 7, 2020 - link
Thunderbolt 3 is 4 lanes at 20.625 Gbit/s, DisplayPort 2.0 is 4 lanes at 20 Gbit/s. Although a Thunderbolt link can only transport up to 40 Gbit/s of a DP 2.0 link at current rates because it is bi-directional, it could still easily support a native DP 2.0 link at 80 Gbit/s. This is another possible feature of Thunderbolt 4, support for native DP 2.0. Although I kinda doubt that made it into Gen12 / Xe graphics.As for HDMI 2.1, it's only 42.66 Gbit/s max, so unless you have a device that requires more than 40 Gbit/s of display data but less than 42.66, you'd be fine with Thunderbolt 3. And Intel is finally touting native HDMI 2.0b with Tiger Lake, so also probably not an issue as far as they're concerned.
name99 - Tuesday, January 7, 2020 - link
Because the modus operandi of new Intel is to announce products ever further in the future...“Will there be a TB4 one day, say maybe in 2028?”
“Perhaps.”
“OK then, let’s announce it as CES along with all our other nonsense, like Foveros 2 (before v1 has even shipped) and EMIB 2, and our ten year fab “roadmap” “.
Spunjji - Wednesday, January 8, 2020 - link
Seriously? "It's 4 times the speed of something that Thunderbolt 3 is already 4 times the speed of"Screw this nonsense.