Intel announces next-gen Thunderbolt with 20 Gbps throughput, 4K support | Hacker News Reader

Intel announces next-gen Thunderbolt with 20 Gbps throughput, 4K support

engadget.com

76 pointsLightning61 comments

USB 3.0 is 'fast enough' for most people and it's way, way cheaper. Only the super-rich or professional market (who've abandoned Apple, haven't they?) can afford things like daisy-chained Thunderbolt monitors.

Firewire never really caught on (with consumers) like Apple would have liked-and it had huge advantages over USB (especially USB 1.0). I bought a Sony laptop specifically because it had a i.Link port and I could daisy chain both my CD-RW and 80GB HD.

Apple's decision to add not just a second Thunderbolt port but also to add an HDMI port to their Retina MacBooks actually reduces demand for Thunderbolt products since users now can connect multiple legacy monitors to their machine. There's good and bad to cannibalizing the same port as video output.

External GPU? Nice idea, but between the cost of the GPU card, the housing, and dealing with all the quirks that go with it, it makes sense to just buy another laptop/desktop.

That said, a $99 Thunderbolt hub that gives 4-6 USB ports, gigabit LAN, video, Firewire, headphone/mic would be a huge seller

Notice that they announced this at NAB - the professional market is definitely a big target, as it should be. We've already seen the trend away from computers and towards iPads and tablets in the consumer space, the market for peripherals at all is increasingly going to become a professional market as this trend continues.

I agree with this, but there's definitely a market for it. I currently have dual monitors with my laptop on a docking station at work, and I'd love to get an ultrabook. I think right now you have to use the laptop monitor + an external, but it does it all via a USB3 docking station. I'm never going to get any kind of render performance out of a USB3 docking station that's handling everything else for me in addition to video.

  > External GPU? Nice idea, but between the cost of the GPU card, the housing, and dealing with all the quirks that go with it, it makes sense to just buy another laptop/desktop.

It would be pretty neat to have an external monitor with a gpu inside it. Then the laptop only needs a gpu powerful enough to drive its own display, instead of also driving external displays.

This.

Given that Apple can control their hardware and software, they should absolutely be doing taking advantage of their opportunity to do so more easily than others.

There are a host of issues with that, unfortunately. I think multi-gpu support is still pretty crummy, and the cost of a good gpu plus a monitor fancy enough to do something so progressive? That's going to be crazy expensive.

I see Thunderbolt not as a USB 3 competitor, but as something in a different category. It can drive external monitors. It can pass data along with that, so you can have external display + audio + USB 3.0 ports all accessible via one cable.

USB 3 can't do that.

Can't it? There are USB 3.0 docking stations that connect to your computer's USB 3.0 port and can handle two external displays via DisplayLink, and provide additional USB 3.0 ports and audio jacks.

I kind of like the external GPU idea, but they are still so expensive. In essence i could carry around an Ultrabook with a decent CPU and at home/work plug it into a a Thunderbolt Dock that gives tons of ports an has an external GPU + Dual Screens. That way i can be truely mobile on the go and have a decent workstation and gaming rig at home, without the Laptop GPU Fan going at 100% and also much faster.

$299 http://www.belkin.com/us/p/P-F4U055

You can't actually buy that, though. The Belkin dock is the Duke Nukem Forever of the peripheral industry.

Ah, sorry, I didn't realize that.

I suppose this is inevitable. The speed advantage is pretty impressive. And the ability to run everything on your "desk" over one cable to the "computer" is pretty handy too. Seem's like the NeXT cube [1] was only a bit more than 20 years before its time :-)

[1] http://www.vectronicsappleworld.com/2006/0611.html

Wake me when someone creates an TB expansion worth having at costs that aren't prohibitive.

The only thing that TB can do that USB3 can't right now is either running multiple very high-res displays, or putting the graphics card off-motherboard.

I'd love to pair a MB Air with a "desktop dock" that had it's own graphics system and replacing all my connectors other than power and TB.

And more RAM. It's be awesome to be able to dock to 128-256GB of RAM with my Macbook Air.

Unfortunately I doubt that would work very well, even thunderbolt 2.0 is no where near the bandwidth of internal RAM and more importantly for many applications, the latency would be terrible.

This is no where near fast enough for RAM, for DDR3 you need at least 50Gbps for the slowest PC3-6400, while for much more common PC3-12800 your looking at 100Gbps.

You can use it for an external ram drive though.

Yes but you could just use SSDs since the extra speed RAM would afford is lost anyways

> putting the graphics card off-motherboard.

Honestly its still not very good at that. Its 'doable' but for intensive applications, it can't come even close to internal connections. For comparison a PCI Express 3.0 x16 link provides 16GB/sec in each direction to main to memory. Thunderbolt '1.0' only offers 1.25GB/sec in each direction per channel, use both channels and you get to 2.5GB/sec putting you in the neighborhood of the old AGP x8 protocol in terms of total bandwidth. Double it for 'thunderbolt 2.0' and its still only 5.0GB/sec. And then there is latency....

Agree with this.

While the steps Thunderbolt is going is great, it's still too much money for what it doesn't deliver. I'd love to have a TB attached SAN for my Macbook Pro, but what I need is usually sufficed by FireWire 800* or USB 2.0 external drive.

* Even a FW800 external drive / enclosure is many times cost prohibitive.

Too much money is a function of how bad you want the performance. I have a Promise RAID attached to my MBP using TB, and it's really fast -- it makes firewire 800 look unacceptably poky. It's been a long time since I measured throughput, but it's like a factor of 10. If you work with large datasets, it's a clear win.

How does that compare, though to a raided SSD tied directly to the PCI board?

I've been led to believe that 300k IOPS is possible using sub-$1000 gear from OCZ [1]. Combined with an agressive backup to spinning disk, this sounds like it'd scream like a monster and be stable to boot.

[1] http://ocz.com/consumer/revodrive-3-x2-pcie-ssd

These two devices (external TB RAID vs. internal PCI SSD) operate in somewhat different parts of the (high end) of the bandwidth/capacity plane.

Anand (http://www.anandtech.com/show/4489/promise-pegasus-r6-mac-th...) found ~700MB/s sequential read/write performance of the TB RAID. He didn't test IOPS directly, although he did put SSDs into the TB RAID and test that (~1000 MB/s). You have to give him credit for that.

In short, the sequential performance is pretty close to the real-world sequential performance of the OCZ unit. Which is quite impressive.

The Promise RAID-6 I use has 8TB usable (and I got it 2 years ago). The units you mention have just 1TB usable (and no redundancy, which is a problem for big data sets).

All this is somewhat academic for my use case, because I dropped my desktop two years ago when TB came out.

If Thunderbolt is cost prohibitive, I suggest moving to USB3. It's cheaper and much more widely deployed, and it's much much faster than FW800 and USB2. In fact, for most external drives, it's just as fast as Thunderbolt as the limiting factor is often the speed of the drive(s) and/or controller and not the IO bus.

Thunderbolt can currently drive only one 30" monitor. This update will allow it to drive two 30" monitors, or one 4k monitor. After this update, it will be possible to drive 4 monitors at 1920x1200, or 5 at 1920x1080, but neither of those really counts as very high resolution.

Thunderbolt can currently drive only one 30" monitor.

Not exactly. It depends on the controller and GPU. The MacBook Air (probably the most common Thunderbolt device) can drive one. The MacBook Pro can drive two.

A current Thunderbolt port provides a DisplayPort 1.1a connection, which has 8.64Gbps. A 30" monitor currently means 2560x1600 at either 24bpp or (more commonly) 30bpp. Even at 24bpp and using CVT-RB, that requires 6.45 Gbps per display. The MacBook Pro only supports two 30" monitors by having them connected to two separate Thunderbolt ports, rather than daisy-chained as is possible with lower resolutions.

EDIT: It looks like the provisioning of bandwidth for a Thunderbolt connection is more complicated, and it's hard to tell from Wikipedia exactly how it works. Thunderbolt and DisplayPort both support 4 high-speed lanes. In Thunderbolt, the lanes are currently 10Gbps, configured as two pairs of input and output lanes carrying PCIe traffic. In DP, they're 4 outbound lanes at up to 5.4 Gbps per lane (for DP 1.2). It appears that Thunderbolt, despite not yet supporting DP 1.2, can operate in DP mode at DP 1.2 speeds of 5.4Gbps per lane, but it's not clear if Thunderbolt supports the multiple stream transport (MST) feature of DP 1.2 that is necessary to drive more than one display over a single DP connection. If a TB cable is configured to provide two DP connections of two lanes each, then it would work, but then it seems that further daisy-chaining of TB devices wouldn't work. If Thunderbolt can drive a monitor with all lanes configured to carry PCIe traffic, that's news to me, and I want to know how Intel and Apple pulled that off.

A MacBook air with 1 current generation thunderbolt port can drive two 2560x1440 thunderbolt displays via daisy chaining.

I can't find any indication on Apple's site that the Air can drive more than one of Apple's Thunderbolt displays, and the math doesn't support your assertion that it can: two such displays would require a minimum of 11.6 Gbps, which can only be provided by DisplayPort version 1.2 (or later). EDIT: If the video data is being sent with the cable operating in pure Thunderbolt mode with no DisplayPort lanes configured for the connection, then there's enough bandwidth, but I don't know how that would work.

All of Apple's 2012 models (save the Mac Pro), and a few of its 2011 models, can drive 2 Thunderbolt displays:

http://support.apple.com/kb/HT5219?viewlocale=en_US#dispnum

I've done this myself with both the MacBook Air (Mid 2012) and the Mac Mini (Late 2012).

A single current generation Thunderbolt port runs two separate 10Gbit channels. Each channel is a mux of PCI-E data and a DisplayPort 1.1a stream. Two displays (each up to DisplayPort 1.1a's 8.64Gb maximum bandwidth) can be driven on one chain, one off of each channel. This can either be two Thunderbolt displays or one Thunderbolt display and one pass through raw DisplayPort display at the end.

Some restrictions apply due to the need for one Thunderbolt controller chip per display to do the DisplayPort demuxing. A Thunderbolt monitor will typically have one Thunderbolt chip and will use its DisplayPort output itself. A DisplayPort display can't be daisy chained directly after a Thunderbolt display (but can after a non-display Thunderbolt peripheral) for this reason.

It sounds like (according to the OP article) that the 2013 Thunderbolt chips (Redwood Ridge) will allow DisplayPort 1.2 pass through only. In this configuration the port could either be run as a Thunderbolt chain (with the above limitations to 2xDP 1.1) or as a DisplayPort 1.2 port. In 1.2 mode Thunderbolt peripherals couldn't be used, but DisplayPort 1.2 features (in particular 4k support and MST) could be used. Presumably the 2014 20Gbit channel Thunderbolt could allow 2xDP 1.2 as that tops out at 17.28Gbit.

Edit: yes, this means that devices sharing the chain with two Thunderbolt displays will have constrained incoming bandwidth. Probably even more so as the only Thunderbolt displays in existence also sport USB, speakers, a webcam, FW800, and Gigabit Ethernet, which together could easily eat up the remaining bandwidth on the TB channel. This may also be part of the reason why Apple hasn't released an updated Thunderbolt display with USB3, 2x10Gbit doesn't go very far with a pair of 27 inch displays hogging most of it.

Ok. That clears things up a bit. So a PCIe device and a DP device can share the same 10Gb channel at the same time? Does this mean that the presence of a DP device will mean that writes to the PCIe device are much slower than the corresponding reads (since the DP device is only using bandwidth in one direction)?

A video of an MBP (not an air) doing this is here:

http://www.youtube.com/watch?v=sP1thwUcO9c

Skip to 30s in to see the single TB cable driving the two displays. You really can drive two of these monitors thru a single TB connection.

> Thunderbolt can currently drive only one 30" monitor.

The 27" monitors are not 30" monitors.

The 30" quoted by the GP was 2560x1600 pixels

The TB 27" monitor has 2560x1440 pixels (10% less)

That's close enough for the ballpark calculation of the GP.

(I'm not saying I 100% follow the GP's understanding of the numbers and systems issues, because I don't. I responded because I've seen the two-monitor, one-cable setup in practice.)

You're right. I was thinking of the similar-but-crucially-different situation of using two Apple Thunderbolt Displays chained together. Multiple sources confirm that works, but obviously that doesn't help if you're trying to use a pair of non-Thunderbolt DisplayPort monitors. The combination of slightly lower resolution (1440 vs 1600) and the built-in controller seems to make all the difference.

Thunderbolt was introduced in 2011 with support for 10Gbps PCIe transport and DisplayPort 1.1 (8.64 Gbps). DisplayPort 1.2 (which allows for 17.28 Gbps) was finalized at the end of 2009. It's a bit surprising that it took Intel this long to announce this update.

That's exactly what's I feel is so wrong in the industry. Intel isn't pushing out new technologies and therefore "standards" into the market. Development is stalled to a minimum innovation, maximum profit situation.

And more generally, Intel just isn't interested in equipping anything with much bandwidth except in the high-end server market. Their current mainstream consumer/low-end server/workstation platform has 20 PCIe 3.0 lanes direct from the CPU (but only 16 are enabled for the consumer chips), and 4 PCIe 2.0 lanes (20Gbps total) for the chipset and all the other peripherals to share. A single one of these new Thunderbolt ports could monopolize all of the peripheral device bandwidth available through Intel's current CPU socket. Copying a file from one SSD to another SSD uses more than half the available bandwidth. It used to be that a desktop system had the advantage of being upgradeable and expandable, but now you often need to get a new CPU+motherboard in order for new peripherals to run at full speed.

It seems obvious that Broadwell will probably have more PCIe bandwidth so that it can feed Thunderbolt 2.0 plus some other stuff.

PCI Express 3.0 is 1GB/s aka 8Gb/s per lane, you seem to have got confused between GB/s and Gb/s. Three lanes would give you more than you need for a new thunderbolt port.

Also if you take x79 which is a consumer platform it has 40 PCI Express 3.0 lanes which would be enough for 16 of the new thunderbolt ports.

X79/LGA2011 is way too expensive to be considered as a mainstream option in the consumer marketplace, and isn't even an automatic choice when shopping for a workstation (since the Xeon E3s are so cheap).

On LGA1155, the connection between the CPU and GPU is now PCIe 3.0, but the connection to the chipset and other peripherals is still running at PCIe 2.0 speeds and is only 4 lanes wide. Unless you have a motherboard that is specifically designed to rob PCIe lanes from the GPU, all USB 3.0 and SATA 3 ports and gigE ports are sharing that 20Gbps link from chipset to CPU.

And, over two years after hardware first shipped, there's still no Linux support for first-gen Thunderbolt from Intel. This is, to put it mildly, disappointing.

20Gbps is only 2.5GB/s, sounds like I can't even saturate x16 PCI-Express lane running at 252.08 Gbps (=31.51 GB/s) so much to "Thunderbolt"!

Correct me if I'm wrong.

The only reason that would make me buy a Macbook Pro is that I can attach eGPUs to it (and hopefully eCPUs in the future aswell).

Sources: Google, http://en.wikipedia.org/wiki/Usb3.0#Thunderbolt

https://www.google.de/#q=20gigabit++to+gigabyte

EDIT: Added sources.

There's your interconnect for the upcoming Mac Pro with a Retina desktop display.

Or with a Retina TV

... in 2014.

Note that this post is talking about two different things: Redwood Ridge (2Q13) and Falcon Ridge (2014).

Good point. I was hasty in my read.

Is there any reason a networking switch couldn't be built with Thunderbolt ports (20Gb/s)? Or for that matter, USB 3.0 (5Gb/s)? Seems a cheaper way to get higher speed interconnects for server clusters than using Ethernet. Sorry if its a naive question, just curious.

EDIT: This would be for new installs, of course. You'd use a motherboard with only a Thunderbolt or USB 3.0 port.

EDIT: I'd love to hear the "more complicated reasons", as your other objections dont apply. In a compute cluster, you have short (within-rack) cable lengths, and the owner of the switch is the owner of the cluster so security is not an issue. Cost is a super big deal in a cluster. More expensive interconnect means buying fewer servers, and that's just dumb.

The obvious response is that you'd have to swap out all the Ethernet hardware in order to get the speed advantage, which gets a bit expensive.

The other half is software, maybe, but I honestly doubt that is actually the problem here. We've been encapsulating protocols in other protocols for so long that that problem is essentially solved in all but the weirdest, most obscure cases, and this wouldn't be that weird or obscure.

The more complicated reasons aside for the moment, thunderbolt is more or less like an external PCI express connector (with some added goodness). As a result you can access a lot of things via thunderbolt you probably wouldn't want outwardly accessible (imagine a network where anybody could do a DMA on demand). Then there are a number of other things, for example the protocol and phy not being designed for 100 meter cable runs like Ethernet, no low level networking support, the fact we have a reasonable and ubiquitous interconnect for networks already.

There are alternatives for super high speed interconnects within a data center though (think infiniband and cousins) or high bandwidth options for longer distances (like fiber) etc.

Infiniband and 10Gb ethernet are both ludicrously expensive steps up from gigabit ethernet which is what makes people look for alternatives everywhere.

Back in the day there were people doing IP over (P)ATA and SCSI for similar reasons... Cable length is not necessarily a big deal - for a lot of applications the appeal would be to use it as an interconnect between servers in the same rack.

10GE and Infiniband are actually not that expensive for HBAs now. Just expensive for switches, but not that expensive, and would be just as expensive for any other non-blocking switch with that level of bandwidth. Juniper EX4500 is probably the cheapest right now, at around $20k for 40 port.

$500/HBA and $500/switch port is cheaper than GE was at the same point in its adoption cycle, I think. Might even be half that with discounts. Aggregated GE is still cheaper and makes more sense for servers in most cases, though.

Hmm, and a rumor came out yesterday on Mac sites, from a source that "has been correct in the past", that the new Mac Pro will be announced this month.

If that's so, could it have this? Check this comment again on April 30th!

From TFA:

…production is set to ramp up in 2014.

"Ramp up" is not "start".

It's when they will be producing it in a streamlined way in large volumes.

Nothing stops them from producing a small run for something like the Map Pro.

And the high cost of the Mac Pro, plus its target market's relative price insensitivity, can help finance the early volume until the price comes down.

Remember, Apple announced products with Thunderbolt on the same day Intel announced it existed... so they had a 6 month lead time on anyone else.

The same could very well be true here.

Does anyone know how OCuLink (similar to Thunderbolt, but by the PCI SIG) is coming along? There doesn't seem to be much public information.

Reply on news.ycombinator.com