A lot of the time these cards won't work at all without a Thunderbolt header on the motherboard (bypassing the header sometimes gives partial functionality). Sometimes they also don't work outside the manufacturer's motherboards as well.
As for form factor none of Gigabyte's ITX boards have TB headers (somewhat makes sense given single slot), but it seems MATX/ATX have decent coverage for top Intel chipsets otherwise it's a bit hit and miss for mid tier.
AMD on the other hand is a whole other thing entirely. It looks like none of the X870/E boards have headers (possibly due to USB4 being mandatory) though several of the 600 series do. Overall it's just a lot more hit and miss on the AMD side of things.
I understand why the header is necessary, but it not being ubiquitous like USB means potential frustration for anyone who didn't plan ahead/expect to need the functionality.
Problem is far too many full size ATX boards only have two slots on them. I think I've even seen some with 1 at all. Gone are the days when you commonly saw 5, 6, or 7 slots on a board. (which is what you should see.) or 3, 4 slots on an mATX board.
These motherboard makers have gone down a bad path by wasting all of that board space with flat M.2 devices when those could either simply be a PCI-e slot of their own, they could be mounted vertical instead of flatly, or mounted on the back of the board. They could eliminate the SATA and make those a bridge insert inside the m.2 to free up more board space as well, there's a lot that could be done to give us back our much needed PCI-e slots. But it doesn't happen.
Gigabyte already has Maple Ridge controllers on their TRX50 boards for their USB 4 interface. While they work beautifully for Thunderbolt devices, Gigabyte did not pay for the certification. So, while all of the hardware works great, some of their new software features don't due to lack of Intel certification (read: pay intel money).
Intel just can’t stop making crappy Thunderbolt controllers… SMH
2 * 80~120 gbps-capable ports, but the bottleneck for PCIe traffic is 64 gbps (PCIe 4 * 4). I would’ve very much liked ASMedia’s approach of not assuming everyone funnels video traffic through the same cable being used for data.
This card should be PCIe Gen 5 * 4 at the very least.
Intel just can’t stop making crappy Thunderbolt controllers… SMH
2 * 80~120 gbps-capable ports, but the bottleneck for PCIe traffic is 64 gbps (PCIe 4 * 4). I would’ve very much liked ASMedia’s approach of not assuming everyone funnels video traffic through the same cable being used for data.
This card should be PCIe Gen 5 * 4 at the very least.
Problem is far too many full size ATX boards only have two slots on them. I think I've even seen some with 1 at all. Gone are the days when you commonly saw 5, 6, or 7 slots on a board. (which is what you should see.) or 3, 4 slots on an mATX board.
These motherboard makers have gone down a bad path by wasting all of that board space with flat M.2 devices when those could either simply be a PCI-e slot of their own, they could be mounted vertical instead of flatly, or mounted on the back of the board. They could eliminate the SATA and make those a bridge insert inside the m.2 to free up more board space as well, there's a lot that could be done to give us back our much needed PCI-e slots. But it doesn't happen.
Too many MoBo makers are designing modern MoBo's via Pure Aesthetics First, Functionality LAST.
It's asinine and Ass-Backwards.
Nobody needs all the fancy looks & colors.
We need Maximum Connectivity, Minimum Color-Fullness & Decorative Heatsinks.
NoBody asked for M.2 to be used outside of "Note Books".
I wanted the old 1.8" HDD Form Factor to be re-purposed to be the new SSD format.
The only difference is for consumer end, we would have a maximum specified thickness of 5 mm.
Otherwise, the shell & the SATA/SAS esque connector would remain largely the same.
The exact same nVME over PCIe connection option would remain the same for a U.2/U.3 connection
No more of that proprietary Heat Sink BS we see with all those new M.2 drives.
Every time, SSD Manufacturers seem to come out with some crazy Heat Sink to cool their M.2 drive.
That's not necessary, standardize on the SSD Physical Form Factor for DeskTop, use U.2 / U.3 or create a U.4 that fits in U.2.
What-ever it takes to standardize on the DeskTop side and to STOP using M.2 in a DeskTop PC form factor.
the M.2 connector is rated ONLY for 50 Mating Cycles.
The SATA / SAS Connector that we all know & love is rated for 10,000 Mating Cycles.
Same with USB Type-C, which we can adapt M.2 into a External USB Drive and still make good use of M.2 drives.
But for Gods-Sake, stop using the M.2 drive form factor in DeskTop as a internal drive.
NoBody asked for it, stop pushing it on us.
Let's use a updated 1.8" HDD Case Form Factor.
If you're worried about PCB Area, it has as much PCB area as the Largest 30x110 mm, if not slightly more, including factoring in Screw Holes & Casing.
TL;DR
I want ALL of my PCIe Rear Add-In Card Slots back in my ATX MoBo's.
Stop shoving M.2 into the DeskTop World.
Re-Purpose old 1.8" HDD Spec into new 5mm Thick 1.8" SSD Drive Spec.
Easy Peasy.
Seperate M.2 for LapTop/Mobile and 1.8" SSD Drive for DeskTop.
That's 100% Beneficial to segregate the NAND Flash supply into multiple SSD Drive Markets.
Heck let's push CFexpress Type-A & Type-B as the new replacements for the Old School A-Drive & B-Drive that was used by Floppy Disks back in the day.
Then you can segment the NAND Flash supply into 4x Major Markets.
1) LapTop / Mobile M.2 SSD
2) Client/Consumer DeskTop 1.8" SSD
3) Memory Cards (CFx Type-A/B) + SD Card support
4) Server Memory SSD's and the various competing SSD form factors.
Also, give me U.2 / U.3 / U.4 connectors for my Storage Drives.
I want to decide what type of Drives I want to connect, stop wasting MoBo Vendor's SKU's with minor variations based on Connector Options.
"Less is More", we need to move to a world based off Mini-SAS HD connectors using U.3 / U.4 that allows 100% backwards compatibility with SATA/SAS/nVME over PCIe/PCIe connections.
Let’s assume your last claim is true. Then here’s the reason:
There are two ports, each capable of minimum 64 gbps PCIe traffic bidirectionally, making the combination capable of 128 gbps. The connection on the host side being PCIe Gen 4 × 4 caps the bandwidth of both combined ports to 64 gbps.
So the only way to utilize the full Thunderbolt bandwidth is to dedicate one port to video or to use both ports for both video and data.
Intel’s implementation has a two-to-one host-to-port PCIe bandwidth ratio.
ASMedia’s implementation of USB4 (off-brand Thunderbolt 4) also has two ports, but at half the bandwidth, and yet the host side is PCIe Gen 4 × 4. This design choice makes it possible to use at least 32 gbps of PCIe bandwidth for each port without any video traffic, thereby affording users a choice of how to allocate their port bandwidth.
ASMedia is still developing their USB4 v2.0 implementation, so it’s unfortunate that Intel’s video-first implementation is the one and only game in town. But by analogy, they would have to be using a PCIe Gen 5 interface on the host side and possibly also on the device side.
Now, this second part is more speculative, but:
Circumstantial evidence points to USB4 being actually capable of 40 gbps PCIe bandwidth if both the host and device side use PCIe Gen 4 × 4. Because this is just what ASMedia did for their USB4 chips for devices (e.g., the chip that goes into the external SSD enclosure). How would PCIe Gen 4 × 4 make sense unless the designers knew it was possible to surpass 32 gbps of PCIe bandwidth per Thunderbolt connection?
The PS2251-21 USB4 NAND flash controller from Phison claims a 4,000 MB/s rate. That would be impossible if PCIe data were capped to 32 gbps out of a total 40 gbps becuase of PCIe and USB/Thunderbolt overhead. I’m no stranger to <Mod Edit> marketing claims―especially product listings which advertise theoretically impossible speeds because they are also counting the protocol overhead as part of the useful payload of data being transmitted. Thus, I was surprised to find this Kickstarter which shows a USB4 external storage device breaking past even 4,000 MB/s:
That Kickstarter project is the one and only example I could find, so I’m taking it with a grain of salt. But taken together, it’s compelling evidence that 32 gbps is theoretically not the PCIe bandwidth cap for USB4/Thunderbolt 4 and 64 gbps won’t be the cap for USB4 v2.0/Thunderbolt 5.
Finally, I’m going to admit I am a bit uncertain about the technical capabilities of USB4 v2.0/Thunderbolt 5’s assymetrical mode of operation. All the marketing materials thus far use one-way video transmission as the application for this capability. But that doesn’t necessarily mean it can only be used for video.
This had been my (any many others’) mistake in understanding USB4/Thunderbolt 4’s capabilities years ago. I thought PCIe bandwidth was capped at 22 gbps or so after subtracting video’s exclusive share of the 40 gbps. Many tests were able to experimentally verify and no counterexamples existed. But that was also because Intel’s implementation was the only game in town. Only when ASMedia’s USB4 implementation became available were we proven wrong.
I hope we’re all wrong thinking that video traffic gets dibs on 120 gbps of asymmetrical bandwidth and that the 120 gbps is only outgoing (from the host). Because why would such a feature not be useful for reading or writing a lot of data to an external SSD or loading tensors to eGPU VRAM?
Anyhow, my whole point has been: I hate Intel’s implementation, which continues a long tradition of crappy implementations, and I hope ASMedia fully fulfills what is technically possible for USB4 v2.0.
You wrote a lot of words to say: you don't know what the Thunderbolt spec is and can't be bothered to look it up... Intel sets the specification for TB period. USB4 is based on TB, but is it's own specification so I would expect devices using USB to work at a higher rate since it doesn't use PCIe tunneling for data.
edit: Also without a PCIe switch you're not going to be able to convert PCIe 5.0 x4 into 2x PCIe 4.0 x4 (TB uses PCIe for data) which would probably double if not triple the cost of the card if there's even a switch that small available for PCIe 5.0. This is of course assuming that would even be possible with a single TB controller and if not you'd need two of those as well.
Also without a PCIe switch you're not going to be able to convert PCIe 5.0 x4 into 2x PCIe 4.0 x4 (TB uses PCIe for data) which would probably double if not triple the cost of the card if there's even a switch that small available for PCIe 5.0. This is of course assuming that would even be possible with a single TB controller and if not you'd need two of those as well.
ASMedia’s controller does just fine using a PCIe Gen 4 × 4 host interface to support USB4/Thunderbolt 4. What doesn’t translate to a PCIe Gen 5 × 4 host interface supporting USB4 v2.0/Thunderbolt 5?
I laid out my points with information I could gather. It’s empirical/circumstantial, but all of it points to Intel’s Thunderbolt chips being deficient in PCIe bandwidth capabilities.
All of this implies that the Thunderbolt 4 spec has nothing to do with such limitations.
ASMedia’s controller does just fine using a PCIe Gen 4 × 4 host interface to support USB4/Thunderbolt 4. What doesn’t translate to a PCIe Gen 5 × 4 host interface supporting USB4 v2.0/Thunderbolt 5?
Having a higher speed interface doesn't change how the controller works, perhaps in your original post you meant the controller should be PCIe 5.0 not the card? If so then we were both talking about two different things.
Controller PCIe interface wise they can use whatever they want to, but maximum speeds are limited by the design of the protocols themselves. In other words if your goal was getting maximum speed across two ports then yes PCIe 4.0 based controller would be able to do that.
Except that Kickstarter page I linked also shows a 12th Gen Intel mobile CPU being capable of over 32 gbps PCIe bandwidth:
So Intel’s CPUs with Thunderbolt built in can blow past 32 gbps. Since it’s Intel’s baby, it surely is Thunderbolt.
Those results, if they're even to be believed in the first place, are on laptops which have integrated controllers that aren't connected by PCIe. In the case of Intel the links are separated so they get maximum bandwidth possible with the only limit being the protocol not the interface which should mean exactly 32Gbps (4GB/s) for data. I'm not as familiar with AMD's design, but as far as I'm aware they integrate their USB controllers into the silicon for laptop SoCs so they should act identically to Intel's.
I laid out my points with information I could gather. It’s empirical/circumstantial, but all of it points to Intel’s Thunderbolt chips being deficient in PCIe bandwidth capabilities.
TB3 controllers shared everything which is what caused the low transfer rates which was somewhat fixed with the TB4 controllers, but Intel's initial ones were PCIe 3.0 so that still limited the maximum bandwidth to the controller. ASMedia's host solution came out a couple of years after Intel's first TB4 host controller and used PCIe 4.0 which eliminates that bottleneck. Due to the extra bandwidth this controller could act a lot more like the integrated implementations in laptop SoCs.
Intel didn't bother releasing another TB4 controller until the TB5 controllers (~4 years apart), but this one does use PCIe 4.0 so I imagine it too will perform much like integrated.
I assume the reason they didn't go PCIe 5.0 for the new controller is that there aren't extra PCIe 5.0 lanes on any Intel client platforms. There's no reason to design a controller for use with an interface that simply doesn't exist.
What are thunderbolt headers for? Is it only compatible with specific Gigabyte motherboards, like for the USB4 ASM4242 (Asus/MSI) add-in card, or with any motherboard with generic pcie 4.0 slots?
Yeah they said the card though which threw me. The controller is PCIe 4.0 maximum 4 lanes so it'd require two controllers to get full bandwidth over two ports at the same time. There are no client boards with 8 lanes open that aren't off the CPU so that design wouldn't make sense. As I've already explained going PCIe 5.0 with the controller itself doesn't make sense either with the current client landscape as there are no Intel boards with extra PCIe 5.0 slots and I believe only one or two AMD so the vast majority would again require splitting CPU PCIe. Down the road this might make sense, but there's also no sign of increasing PCIe on client systems.
I think the only way we'd see a refresh controller any time soon that supports PCIe 5.0 is if they're unable to integrate TB5 into the CPUs. Currently there are no implementations and the best guess is that it simply uses too much power so adding bandwidth right now certainly wouldn't help that.
The PS2251-21 USB4 NAND flash controller from Phison claims a 4,000 MB/s rate. That would be impossible if PCIe data were capped to 32 gbps out of a total 40 gbps becuase of PCIe and USB/Thunderbolt overhead. I’m no stranger to <Mod Edit> marketing claims―especially product listings which advertise theoretically impossible speeds because they are also counting the protocol overhead as part of the useful payload of data being transmitted. Thus, I was surprised to find this Kickstarter which shows a USB4 external storage device breaking past even 4,000 MB/s:
edit: Also without a PCIe switch you're not going to be able to convert PCIe 5.0 x4 into 2x PCIe 4.0 x4 (TB uses PCIe for data) which would probably double if not triple the cost of the card if there's even a switch that small available for PCIe 5.0. This is of course assuming that would even be possible with a single TB controller and if not you'd need two of those as well.
What are thunderbolt headers for? Is it only compatible with specific Gigabyte motherboards, like for the USB4 ASM4242 (Asus/MSI) add-in card, or with any motherboard with generic pcie 4.0 slots?
They're for integrating into the PCIe lanes for hot swap as well as GPIO, UART and sleep etc. Intel hasn't made a header specification as far as I'm aware so all of the motherboard manufacturers use proprietary which means the cards only work fully with their respective motherboard vendor. It's something Intel should probably fix, but I doubt they really care as add on TB has always seemed like a low priority.
What puzzles me is that TB certification imposes guaranteed minimum bandwidths (this is the added value over USB). But obviously, only one port is tested on this kind of addin card, never both at the same time. Otherwise, Intel itself would have failed its own certification, with its TB4 and TB5 controlers.
They're for integrating into the PCIe lanes for hot swap as well as GPIO, UART and sleep etc. Intel hasn't made a header specification as far as I'm aware so all of the motherboard manufacturers use proprietary which means the cards only work fully with their respective motherboard vendor. It's something Intel should probably fix, but I doubt they really care as add on TB has always seemed like a low priority.
And without that proprietary header, the TB port doesn't work at all, right?
Most general purpose controllers (sata, nvme, usb...) work over pcie without any additional header. It's really pity that USB4 / TB controllers need that header and aren't autonomous on pcie.
I believe it's possible to get some of them to work, at least partially, but there would likely be bugs. I don't think there's any way they could get it to work through just the PCIe slot alone, but if Intel standardized the header every card would at least work on any board with a header. That would be a big improvement over how it is now even if it's not the best for end users.
4,059.15 MB/s is 32.47 gb/s, and that’s just the goodput―which does not include the overhead that is typically counted in the nice round figures. With 128b/132b overhead, that would come out to 33.48 gb/s, and that’s before accounting for NVMe protocol overhead (which is negligible, but certainly enough to remove doubt that the PCIe traffic exceeds 32 gb/s).
Those results, if they're even to be believed in the first place, are on laptops which have integrated controllers that aren't connected by PCIe. In the case of Intel the links are separated so they get maximum bandwidth possible with the only limit being the protocol not the interface which should mean exactly 32Gbps (4GB/s) for data. I'm not as familiar with AMD's design, but as far as I'm aware they integrate their USB controllers into the silicon for laptop SoCs so they should act identically to Intel's.
It’s really a shame that laptops get the built-in support. I could do with less cable clutter behind the desktop. Every device I plug in is connected to a long cable to a single port on the back or I gotta fumble for the hole to plug it into. 😞
ASMedia's host solution came out a couple of years after Intel's first TB4 host controller and used PCIe 4.0 which eliminates that bottleneck. Due to the extra bandwidth this controller could act a lot more like the integrated implementations in laptop SoCs.
I assume the reason they didn't go PCIe 5.0 for the new controller is that there aren't extra PCIe 5.0 lanes on any Intel client platforms. There's no reason to design a controller for use with an interface that simply doesn't exist.
Some motherboard vendors will use CPU lanes for their Thunderbolt support though. My ASRock X670E Taichi does just this: a PCIe Gen 3 × 4 controller consuming the CPU’s PCIe Gen 5 lanes. There’s those extra PCIe Gen 5 lanes. They’re just being underutilized.
The key is that this isn't the case for Intel client platforms and quite frankly there's no reason for them to design something that only benefits AMD.
I believe this is the case for AMD's mandatory USB4 on X870 based boards as well though I assume they're using ASMedia's PCIe 4.0 host controller. On Intel's ARL TB4 support is provided by the CPU itself rather than using PCIe lanes. I'd love to know how TB4/5 is implemented on Intel client platforms because Asus' ProArt series doesn't use any of the available CPU PCIe lanes for TB4 on Z790 or TB5 on Z890. I'm guessing they're able to route it off of the chipset.
It’s really a shame that laptops get the built-in support. I could do with less cable clutter behind the desktop. Every device I plug in is connected to a long cable to a single port on the back or I gotta fumble for the hole to plug it into. 😞
I believe ARL is the first desktop CPU to come with native Thunderbolt support, and I hope this becomes the standard going forward. I also hope that when AMD replaces the current desktop IOD they integrate USB4 into the new one.
4,059.15 MB/s is 32.47 gb/s, and that’s just the goodput―which does not include the overhead that is typically counted in the nice round figures. With 128b/132b overhead, that would come out to 33.48 gb/s, and that’s before accounting for NVMe protocol overhead (which is negligible, but certainly enough to remove doubt that the PCIe traffic exceeds 32 gb/s).
4,059.15 MB/s is 31.71Gbps (I assume it's 4,059 MiB/s, not MB/s, the tool wrongly reports this), just below the bandwidth of 4 PCIe Gen 3 lines, and well below the bandwidth of USB4 40Gbps.
Not sure how to take account the overhead and which values include or excludes them (40Gbps? 32Gpbs? 4059MBs/?)
On the X670E, the ASM4242 is connected directly to the CPU, while on the X870E, it is connected throught the chipset (itsel connected to the CPU with 4 lines Gen 4), and share its bandwidth with all the others links (ethernet, wifi, m2...)
I believe ARL is the first desktop CPU to come with native Thunderbolt support, and I hope this becomes the standard going forward. I also hope that when AMD replaces the current desktop IOD they integrate USB4 into the new one.
4,059.15 MB/s is 31.71Gbps (I assume it's 4,059 MiB/s, not MB/s, the tool wrongly reports this), just below the bandwidth of 4 PCIe Gen 3 lines, and well below the bandwidth of USB4 40Gbps.
I don’t see any way to squeeze your numbers out of any interpretation of the units.
Binary
Decimal
4,059.15 MiB/s
4,059.15 MB/s
4,156,569.60 kiB/s
4,059,150.00 kB/s
4,256,327,270.40 B/s
4,059,150,000.00 B/s
34,050,618,163.20 b/s
32,473,200,000.00 b/s
34,050,618.16 kb/s
32,473,200.00 kb/s
34,050.61 Mb/s
32,473.20 Mb/s
34.05 Gb/s
32.47 Gb/s
Either way you convert the units, you do not get anything lower than 32 Gb/s. And that is the goodput. We can omit negligible overhead from the NVMe protocol.
There is a minimum 24 bytes of overhead [1], and the packets are limited to a maximum payload size (MPS) of 128 bytes [2]. That pads the goodput with additional bits which yields a throughput that is at least 118.75% the size of its payload in the best case scenario.
Binary
Decimal
40.44 Gb/s
38.56 Gb/s
The binary interpretation here yields a figure greater than the nominal capacity now, but let’s carry on to the end.
The physical throughput (actual bits pushed through the wires) is after you perform line encoding. For USB4 carrying 40 gb/s PCIe traffic, every 128 bits are encoded to 130 bits (128b/132b encoding) [1], so:
Binary
Decimal
41.70 Gb/s
39.77 Gb/s
And now I’m actually glad I went beyond the napkin math for this reply, because it’s crystal clear that CrystalDiskMark does not use binary units. 🙂
I would like to be fact-checked on this though. There is a nagging feeling from being so used to Windows’ units that I can’t shake.
I have two X670E boards. One is CPU-connected, and the other is chipset-connected. Both use Intel’s controller. ASMedia hadn’t even come out with its USB4 controller until recently. [3]
That makes sense, to me, that the speed stays below 32Gbps. And with MiB/s instead of MB/s, that works (many people are confused with them, even professionnal or company, sometimes).
But I'll give you the benefit of the doubt. Maybe it's really MB/s and so, above 32Gbps. In any case, it's well below 40Gbps. Enough room for the protocol's stuff.
I have two X670E boards. One is CPU-connected, and the other is chipset-connected. Both use Intel’s controller. ASMedia hadn’t even come out with its USB4 controller until recently. [3]
Okay. I was referring to the leaks of Gigabyte documents (2021), prior to the launch of AM5 motherboards, where this design (with discret USB4 controler) was recommended by AMD. Since the Promontory chipset is also from ASMedia, It seemed logical that they should work together to propose this solution and to promote ASMedia's chips.
But after that, each motherboard makers were free to do as they want (especially if ASMedia was late with its chip).
On X870, both M.2 from the CPU are free (I think it's mandatory). So the USB4 chip is connected to the chipset.
Facebook
Twitter
Instagram