Thunderbolt 3 Controller on M1 Macs

[Yebubbleman]

Am curious as, the more I think about it, this is seeming nebulous. On M1 Macs, is there an actual integrated Thunderbolt 3 controller embedded on the M1 as one of its many sub-components? Or is Apple merely just using USB4 and claiming Thunderbolt 3 support by virtue of that being something that USB4 inherently supports?

I guess it doesn't really matter ultimately, as the end result is the same regardless. But I'd still be curious if anyone out there knows for sure. (If you have sources, please share them too!)

 

[Gnattu]

The TB3 compatibility of USB4 hosts are "optional", not "required", per USB-IF. In other words, not all USB4 ports can work with Thunderbolt3 devices.
TB3 is easier to be verified than USB4 at the moment, because we have tons of Thunderbolt devices out there, but we have no USB4 device now.

 

[Pressure]

You can't find an Intel ThunderBolt controller on the PCB. There are 2 small retimers for ThunderBolt 4 (JHL8040R) that cost about $2.40 a piece.

 

[Hexley]

You can't find an Intel ThunderBolt controller on the PCB. There are some small retimers for ThunderBolt 4 (JHL8040R) that cost about $2.40 a piece.

[Teardown] Late 2020 Mac mini: Apple Silicon M1, Thunderbolt...

As soon as the M1 Mac mini arrived, I tore it down to confirm a few things Apple didn't make clear. All M1 Macs come with Thunderbolt 4 ports. I prepa...

egpu.io

Teardown confirms the SKU number

 

[Yebubbleman]

The TB3 compatibility of USB4 hosts are "optional", not "required", per USB-IF. In other words, not all USB4 ports can work with Thunderbolt3 devices.
TB3 is easier to be verified than USB4 at the moment, because we have tons of Thunderbolt devices out there, but we have no USB4 device now.

You can't find an Intel ThunderBolt controller on the PCB. There are 2 small retimers for ThunderBolt 4 (JHL8040R) that cost about $2.40 a piece.

Is there any way to confirm whether Apple has their own Thunderbolt controller in M1; or whether it's just USB 4; or whether it's something borrowed/licensed from Intel that they've merely baked into M1? I know the end result is the same, but I'd definitely still be curious.

 

[chabig]

Is there any way to confirm whether Apple has their own Thunderbolt controller in M1

Apple's controller is in the M1. Apple says so:

"An Apple-designed Thunderbolt controller with support for USB 4, transfer speeds up to 40Gbps, and compatibility with more peripherals than ever."

Apple unleashes M1

Apple today announced the biggest leap forward for the Mac with M1, the first system on a chip designed specifically for the Mac.

www.apple.com

 

[Yebubbleman]

Apple's controller is in the M1. Apple says so:

"An Apple-designed Thunderbolt controller with support for USB 4, transfer speeds up to 40Gbps, and compatibility with more peripherals than ever."

Apple unleashes M1

Apple today announced the biggest leap forward for the Mac with M1, the first system on a chip designed specifically for the Mac.

www.apple.com

I wasn't sure if they used an ACTUAL controller or if it was just marketing nonsense and they were leveraging USB4 features to be backwards compatible or if Intel had some kind of hand in what they used. I suppose the most logical and likely answer is that, now that Intel has opened up the standard, Apple probably just designed their own without them.

 

[chabig]

I don't know with certainty that Thunderbolt is open, but it is royalty-free.

 

[tdar]

The difference between the Apple designed TB controllers and the intel ones used before is that Apple has only one TB port per controller ,while intel have two. Apple has them on the die of the SoC and just includes the number of ports that they want.

I see so many people who are worried that there will never be more than 2 ports. Not true.,

 

[Krevnik]

I wasn't sure if they used an ACTUAL controller or if it was just marketing nonsense and they were leveraging USB4 features to be backwards compatible or if Intel had some kind of hand in what they used. I suppose the most logical and likely answer is that, now that Intel has opened up the standard, Apple probably just designed their own without them.

As already stated, USB4 doesn’t mandate Thunderbolt support. You can get 20Gbps over USB-C using dual 10Gbps USB 3 streams. USB 3.2 Gen 2x2 or whatever word salad they are calling it. Really USB4 is: Mandated USB-C + Optional TB3.

This is a good example of “distinction without a difference” though. An Alpine Ridge TB3 controller for example, is practically a USB4 controller in all but name. Apple would have likely started from some reference point and then made customizations, rather than designing from scratch.

I agree that Apple likely did customize this controller though, since they’ve been deep enough in Thunderbolt up to this point to be rolling custom firmware on the Intel TB controllers they have been using. And since Intel is interested in more tight integration with their stuff, there’s a decent amount of stuff that Apple wouldn’t want or couldn’t use from a licensed Intel logic block.

The difference between the Apple designed TB controllers and the intel ones used before is that Apple has only one TB port per controller ,while intel have two. Apple has them on the die of the SoC and just includes the number of ports that they want.

Which at least is in part because using on-die controllers means they need the retimer chips near the physical ports themselves. Thunderbolt is very sensitive to trace distances.

AFAIK, there’s not yet a chip that can be both a port mux and a retimer for Thunderbolt. I’m curious if Apple will add more on-die controllers, which eats die space, requires more I/O pins, and more retimer chips. Or if they are either waiting or working on a chip that can be a retimer and port mux.

 

[tdar]

As already stated, USB4 doesn’t mandate Thunderbolt support. You can get 20Gbps over USB-C using dual 10Gbps USB 3 streams. USB 3.2 Gen 2x2 or whatever word salad they are calling it. Really USB4 is: Mandated USB-C + Optional TB3.

This is a good example of “distinction without a difference” though. An Alpine Ridge TB3 controller for example, is practically a USB4 controller in all but name. Apple would have likely started from some reference point and then made customizations, rather than designing from scratch.

I agree that Apple likely did customize this controller though, since they’ve been deep enough in Thunderbolt up to this point to be rolling custom firmware on the Intel TB controllers they have been using. And since Intel is interested in more tight integration with their stuff, there’s a decent amount of stuff that Apple wouldn’t want or couldn’t use from a licensed Intel logic block.



Which at least is in part because using on-die controllers means they need the retimer chips near the physical ports themselves. Thunderbolt is very sensitive to trace distances.

AFAIK, there’s not yet a chip that can be both a port mux and a retimer for Thunderbolt. I’m curious if Apple will add more on-die controllers, which eats die space, requires more I/O pins, and more retimer chips. Or if they are either waiting or working on a chip that can be a retimer and port mux.

I’d bet for the current design in the short run and the in-house retimers in the long term.
Possibly even on package.

 

[Pressure]

Remember Apple co-developed LightPeak (ThunderBolt) with Intel.

 

[Krevnik]

I’d bet for the current design in the short run and the in-house retimers in the long term.
Possibly even on package.

Unless I'm mis-reading this, on-package retimers would be worthless.

 

[Yebubbleman]

I don't know with certainty that Thunderbolt is open, but it is royalty-free.

Wouldn't it need to be open for Apple to be using their own custom controller?

 

[leman]

Which at least is in part because using on-die controllers means they need the retimer chips near the physical ports themselves. Thunderbolt is very sensitive to trace distances.

AFAIK, there’s not yet a chip that can be both a port mux and a retimer for Thunderbolt. I’m curious if Apple will add more on-die controllers, which eats die space, requires more I/O pins, and more retimer chips. Or if they are either waiting or working on a chip that can be a retimer and port mux.

You seem to know your stuff, so please correct me if I am wrong. If I understand it correctly, the Thunderbolt 3 implementation on Intel Macs uses shared controllers, where two ports are connected to a single controller. This would mean that you don't actually get the full bandwidth, as there is only one 4-lane PCIe link for each pair of ports. I would guess that Apple aims to go for "full" TB3 implementation, where each port can actually offer the full bandwidth.

 

[Captain Trips]

Here is a blog article from Otherworld Computing by one of their people who writes and debugs storage drivers:

THUNDERBOLT ON THE M1 MAC MINI – WHEN 2 ACTUALLY DOES EQUAL 4

 

[leman]

Here is a blog article from Otherworld Computing by one of their people who writes and debugs storage drivers:

THUNDERBOLT ON THE M1 MAC MINI – WHEN 2 ACTUALLY DOES EQUAL 4

Thanks for this, it’s exactly what I thought. Seems that Apple has really undersold the I/O capabilities of the M1 Macs...

 

[Krevnik]

You seem to know your stuff, so please correct me if I am wrong. If I understand it correctly, the Thunderbolt 3 implementation on Intel Macs uses shared controllers, where two ports are connected to a single controller. This would mean that you don't actually get the full bandwidth, as there is only one 4-lane PCIe link for each pair of ports. I would guess that Apple aims to go for "full" TB3 implementation, where each port can actually offer the full bandwidth.

It’s always about the trade-offs. Intel offered two controllers up until recently. A one-port bus controller, and a two-port bus controller. Since the controllers needed to be *very* close to the ports, and aren’t exactly small, Apple opted for the two-port bus controllers. It saves space on the logic board, and only requires 4x PCIe lanes to feed it, so you aren’t fighting for I/O from the CPU that is also needed for the 5300M in the 16” MBP.

They could have done 1-port-per-bus under Intel. They just decided not to. But the 16” MBP logic board looks pretty crammed as it is, so I’m not terribly surprised here.

Moving to Apple Silicon gives them more options. They don’t have to ration PCIe lanes if they don’t want to. Their GPU design may give them the ability to free up lanes and I/O that would normally go to a dGPU in the 16” MBP. But there’s still the trade offs. 4 integrated controllers means 4 retimers and more I/O pins dedicated to it. There are advantages to this, but my use of TB doesn’t really benefit from it. About the only devices capable of saturating TB are PCIe SSDs, some RAIDs, and eGPUs. With DSC in the mix now, high resolution displays going forward will likely use less bandwidth than the LG 5K Ultrafine. It depends a bit on how common SSDs and RAIDs are in pro workflows, I bet.

The question really winds up being: What does Apple think Pros want/need? What trade offs in the “M1X” design are they willing to make to cater to that vision? I’m honestly not very sure how this will play out. Apple certainly has data on how the four port MacBook Pros get used, and I kinda wish I had access to it.

Thanks for this, it’s exactly what I thought. Seems that Apple has really undersold the I/O capabilities of the M1 Macs...

To be honest, that article is also doing a lot of lifting as marketing copy for their new Thunderbolt hub. Those sorts of hubs weren’t feasible until recently, as it requires the new JHL8440 controller released in the last few months. This new controller is meant for docks, but is also the first controller that supports 4 TB ports.

This is also why docks up to this point only ever had 1 or 2 TB ports, depending on which controller they used.

 

[ght56]

I'd be very curious to hear what speeds people are getting with USB 3.2 gen 2x2 NVMe SSDs. AFAIK, the only super common consumer device with this ATM is the SanDisk Extreme Pro V2 SSD that got updated two months ago. Theoretically, sustained reads and writes should both be upwards of 2,000 MB/s.

 

[MikhailT]

Watch out for marketing stunts on trying to sell TB/USB4 as if it is TB4 + USB4.

TB4 and USB4 are not the same thing; Intel still mandates and control the certifications of any Thunderbolt devices (except for Apple most likely). The difference between USB4 and TB4 is that USB4 optionally supports TB3 but TB3 itself does not have to have the same requirements that would make it TB4. In other words, TB4 made most of the optional spec in TB3 mandated in TB4.

USB4 = USB C + USB 3.2 Gen 2 + TB3 (optional and with optional specs not required)
TB4 = USB C + TB3 mandated spec by Intel here: https://newsroom.intel.com/news/int...versal-cable-connectivity-everyone/#gs.aaack6

Intel allows companies to build TB3/TB4 controllers but only if they meet Intel's certifications and requirements. I haven't heard anyone done this yet because Intel is known for high fees for this, thus why TB didn't take off so fast.

Wouldn't it need to be open for Apple to be using their own custom controller?

No, as noted before, Apple helped developed Thunderbolt with Intel. They most likely have an exclusive contract to allow them to use it in any shape they want. This is traditionally how they work with partners, they do not give tech/patents out freely without something in return.

 

[Sarpanch]

I'd be very curious to hear what speeds people are getting with USB 3.2 gen 2x2 NVMe SSDs. AFAIK, the only super common consumer device with this ATM is the SanDisk Extreme Pro V2 SSD that got updated two months ago. Theoretically, sustained reads and writes should both be upwards of 2,000 MB/s.

As of now, the M1 chips are only supporting 10Gb/s, even on 3.2 Gen 2x2 devices. More details here

Is it possible that Apple’s USB 4 implementation doesn’t support 20Gb/s speeds over USB?

 

[ght56]

As of now, the M1 chips are only supporting 10Gb/s, even on 3.2 Gen 2x2 devices. More details here

Is it possible that Apple’s USB 4 implementation doesn’t support 20Gb/s speeds over USB?

Bummer. I really hoped these Macs supported 2x2. I suppose the line of thinking is that USB 4 drives will rapidly become low cost and the standard without having to pay the TB royalties, but who knows when they will hit the market in quantity. The 2x2 SanDisk EP v2 is so neat because it is super fast storage for a pretty reasonable price. The 2TB 2x2 V2 is only 20 bucks more than the V1.

 

[Yebubbleman]

It’s always about the trade-offs. Intel offered two controllers up until recently. A one-port bus controller, and a two-port bus controller. Since the controllers needed to be *very* close to the ports, and aren’t exactly small, Apple opted for the two-port bus controllers. It saves space on the logic board, and only requires 4x PCIe lanes to feed it, so you aren’t fighting for I/O from the CPU that is also needed for the 5300M in the 16” MBP.

They could have done 1-port-per-bus under Intel. They just decided not to. But the 16” MBP logic board looks pretty crammed as it is, so I’m not terribly surprised here.

Moving to Apple Silicon gives them more options. They don’t have to ration PCIe lanes if they don’t want to. Their GPU design may give them the ability to free up lanes and I/O that would normally go to a dGPU in the 16” MBP. But there’s still the trade offs. 4 integrated controllers means 4 retimers and more I/O pins dedicated to it. There are advantages to this, but my use of TB doesn’t really benefit from it. About the only devices capable of saturating TB are PCIe SSDs, some RAIDs, and eGPUs. With DSC in the mix now, high resolution displays going forward will likely use less bandwidth than the LG 5K Ultrafine. It depends a bit on how common SSDs and RAIDs are in pro workflows, I bet.

The question really winds up being: What does Apple think Pros want/need? What trade offs in the “M1X” design are they willing to make to cater to that vision? I’m honestly not very sure how this will play out. Apple certainly has data on how the four port MacBook Pros get used, and I kinda wish I had access to it.



To be honest, that article is also doing a lot of lifting as marketing copy for their new Thunderbolt hub. Those sorts of hubs weren’t feasible until recently, as it requires the new JHL8440 controller released in the last few months. This new controller is meant for docks, but is also the first controller that supports 4 TB ports.

This is also why docks up to this point only ever had 1 or 2 TB ports, depending on which controller they used.

You compare the 16" MacBook Pro to the M1 Macs with regards to Thunderbolt controllers (and do so awesomely, might I add). I have a question though. The 4-port 2020 Intel 13" MacBook Pro has the distinct honor of being the only Mac Apple ever shipped with the Thunderbolt controller on the Intel processor itself (unless the 2020 Air's 10th Gen Y-series chips also did this, but I'm not 100% sure they did). How is your comparison different factoring in that machine's Thunderbolt 3 setup?

 

[Krevnik]

You compare the 16" MacBook Pro to the M1 Macs with regards to Thunderbolt controllers (and do so awesomely, might I add). I have a question though. The 4-port 2020 Intel 13" MacBook Pro has the distinct honor of being the only Mac Apple ever shipped with the Thunderbolt controller on the Intel processor itself (unless the 2020 Air's 10th Gen Y-series chips also did this, but I'm not 100% sure they did). How is your comparison different factoring in that machine's Thunderbolt 3 setup?

Things get pretty messy here, which is why I didn't say anything about the Ice Lake 13" MBP. There's no teardown floating around of the 2020, nor can I find any good images of the logic board to confirm what chips Apple has used. I haven't seen a system report from one of these machines yet either describing the number of TB buses available.

Unfortunately, Intel doesn't publicly share tech docs so I can't go delving there either. So really the best source I can find is Wikichip: https://en.wikichip.org/wiki/intel/microarchitectures/ice_lake_(client)#Thunderbolt_IO_subsystem

It looks like the subsystem does support four physical ports, provided you include the retimers for each port. The logic for two ports is consolidated into a single logic block that is duplicated. Since this replaces some of the PCIe lanes that would normally be external I/O, and can also share lines with what would normally be USB-C ports, they probably are more trading off PCIe pins for TB3/USB-C pins.

That said, this source is also clearly flawed and having to make some guesses. It makes the following wrong claim: "Like Titan Ridge, each retimer supports two ports. The retimers themselves are still only sold by Intel but they are a fraction of the size, so there is also a modest board space saving advantage as well." Intel produces no retimer currently that supports two ports, and two retimers look an awful lot like a Titan Ridge controller in terms of foot print when looking at them on M1 systems. Whoops.

So, taking that with a grain of salt, the answer is: I don't know for sure, but I suspect we'd see 4 retimers on the logic board.

Either way, how Ice Lake does it doesn't really mean Apple has to do it the same way. Again, using their own designs means they have full control. But we also don't exactly have precedent from Apple to speculate too much with. They could follow Intel and just expose 4 ports of physical I/O off the chip.

 

[ght56]

Thanks for this, it’s exactly what I thought. Seems that Apple has really undersold the I/O capabilities of the M1 Macs...

Arguably, Apple has undersold the IO capabilities of their computers for the past half decade. And it took the industry a long time to even remotely begin catching up...and even in 2021 the industry is still catching up. You could really even say they went pretty far ahead of the curve starting in 2011 when 10 Gbps was unheard of!