2.5/5 GBps USB Ethernet Adapter Announced

[RMerlin]

While people are glossing over 802.11ax routers, this is something that particularly caught my attention yesterday:

https://www.anandtech.com/show/1285...ngle-chip-usb-30-to-5g25g-dongles-coming-soon

A USB 2.5/5 GBps Ethernet adapter. Pricing will be the important thing here, but if I remember correctly, that manufacturer was the one who opened the door to inexpensive PCI-Express 5 Gbps adapters a year or two ago.

 

[joegreat]

Sounds to good to be true: Why they use USB 3.0 interface (which is max. 5 Gbit throughput) and not USB 3.1/C which would allow double the throughput?
I would be surprised to see the full 5 Gbit speed via this device a it would saturate the USB port/hub on the host...

But if the price is low enough it could be a very good upgrade for 2.5 or near 5 Gbit connectivity for mobile devices!

 

[Internet Man]

Sounds to good to be true: Why they use USB 3.0 interface (which is max. 5 Gbit throughput) and not USB 3.1/C which would allow double the throughput?
I would be surprised to see the full 5 Gbit speed via this device a it would saturate the USB port/hub on the host...

But if the price is low enough it could be a very good upgrade for 2.5 or near 5 Gbit connectivity for mobile devices!

The Aquantia chips do appear to support USB 3.1

"The Aquantia AQtion AQC111U and AQC112U are Universal Serial Bus (USB) to Multi-Gig Ethernet controllers that integrate a USB v3.1 Gen1 PHY and 5GBASE-T/2.5GBASE-T MAC/PHY in single-chip packages." - Source

 

[RMerlin]

Sounds to good to be true: Why they use USB 3.0 interface (which is max. 5 Gbit throughput) and not USB 3.1/C which would allow double the throughput?
I would be surprised to see the full 5 Gbit speed via this device a it would saturate the USB port/hub on the host...

But if the price is low enough it could be a very good upgrade for 2.5 or near 5 Gbit connectivity for mobile devices!

Market share, I assume.

Sent from my P027 using Tapatalk

 

[RMerlin]

The Aquantia chips do appear to support USB 3.1

"The Aquantia AQtion AQC111U and AQC112U are Universal Serial Bus (USB) to Multi-Gig Ethernet controllers that integrate a USB v3.1 Gen1 PHY and 5GBASE-T/2.5GBASE-T MAC/PHY in single-chip packages." - Source

USB 3.1 Gen 1 is actually USB 3.0. Gen 2 is the new faster standard.

Sent from my P027 using Tapatalk

 

[System Error Message]

since many usb3 ports are either 2.5Gb/s or 5Gb/s, this'll be perfect.

 

[scooter2765]

I worry what cpu utilization would be and how much it would tie up the USB bus as I have backup disks hanging off USB. Would rather have pcie. Good for temp situations though imaging etc..

 

[RMerlin]

I worry what cpu utilization would be and how much it would tie up the USB bus as I have backup disks hanging off USB. Would rather have pcie. Good for temp situations though imaging etc..

Make sure to use a USB port from a different hub/controller than the disk.

 

[sfx2000]

Aquantia is hitting on all cylinders these days - the chipset here is a big deal - previously for laptops with TB3, 10G was still a bit spendy...

Now if we can get switch prices down on a per-port basis...

 

[Razor512]

While hard to do with the standard ping measurements, has anyone done testing on the response times of USB Ethernet controllers VS ones that are connected to the PCI express bus?

 

[sfx2000]

While hard to do with the standard ping measurements, has anyone done testing on the response times of USB Ethernet controllers VS ones that are connected to the PCI express bus?

Not 10G, but having access to Thunderbolt and USB3 on MacOS with 1G devices, Thunderbolt by a long shot....

 

[System Error Message]

Not 10G, but having access to Thunderbolt and USB3 on MacOS with 1G devices, Thunderbolt by a long shot....

I worry what cpu utilization would be and how much it would tie up the USB bus as I have backup disks hanging off USB. Would rather have pcie. Good for temp situations though imaging etc..

Make sure to use a USB port from a different hub/controller than the disk.

Thunderbolt 3 can do up to 40Gb/s or 4 PCIe 3.0 lanes alongside 100Watts of power and separate display channels too. Thunderbolt 2 is half of that. USB 3.1 or gen-C is at most 10Gb/s at best.

In most chipsets like with x86, the USB ports and storage are attached to the same southbridge/northbridge sharing the same bus to the CPU. That bus has a lot of bandwidth and in SoCs the USB ports are connected to the controller on the SoC. The northbridge/controllers all do the talking so the CPU or CPU memory controller does the passing of data around so unless the SoC is very lacking in hardware processing, it will not tie up the CPU. What does tie up the CPU however are some drivers, for instance some USB to ethernet controller drivers use the CPU to perform the tasks rather than just passing off data and instructions to the chips as many usb3 to ethernet controllers boast of their own processing capabilities.

Even fakeraid or raid using intel chipset or amd realtek also have some of the processing done on the chip level but also utilises the ram shared with the OS and CPU. So performance is highly dependent on drivers and the features the chip has and how many of those processing features are utilised and offloaded. This is how the raspberry pi can handle 100Mb/s of ethernet or even more if you use a gigabit usb ethernet controller despite the slower CPU when its own GPIO is only capable of a few Mb/s as it directly uses the CPU.

Performance mainly depends on the controller, drivers and how fast the bus is.

 

[quadra2030]

Apple Thunderbolt 2 to Gigabit Ethernet adapter has definitely lower CPU usage (and better throughput) on MacOS X than any USB 3.x Gigabit Ethernet adapter (AFAIK it uses Broadcom chip)..

 

[TheLostSwede]

Aquantia is hitting on all cylinders these days - the chipset here is a big deal - previously for laptops with TB3, 10G was still a bit spendy...

Now if we can get switch prices down on a per-port basis...

I visited Realtek today, but stupidly forgot to take a picture.
They have a 10Gbps 8-port "dumb" switch that's going to go for less than $30 in quantity...
But, and it's a big BUT, unless you're using SFP+ for your network, you're still going to need a PHY for each of the ports to get RJ45.
I would be more than possible to do a $100 8-port SFP+ switch with the chip in question though.
Realtek's preferred PHY partner is Aquantia btw.

On the other hand, Marvell was showing off a single chip 8-port PHY for speeds up to 5Gbps. So we might actually get to see some affordable 5Gbps products in the market this year.

 

[TheLostSwede]

since many usb3 ports are either 2.5Gb/s or 5Gb/s, this'll be perfect.

Sorry, what?
USB 3.0/3.1 gen 1 is "5Gbps" or more correct, 4.8Gbps, whereas 3.1 gen 2 is 10Gbps and might actually reach it thanks to 128/132b encoding, rather than 8/10b encoding, as it should reduce overheads quite a bit.

This is obviously assuming that you have a controller that is connected to a fast enough bus. In the case of Intel's "real" 300-series chipsets (not the Z370), it would appear each 10Gbps port only has an 8Gbps interface (one PCIe 3.0 lane equivalent) to the chipset, which is a fairly obvious bottleneck. In other words, any USB 3.1 gen 2 controller that uses anything less than a two lane PCIe 2.0 interface won't deliver the 10Gbps performance.

I have to agree though, that it's odd that they chose to use a USB 3.0/3.1 gen 1 host interface for these, but maybe they tested and figured that the network overheads are high enough that the slight reduction in (theoretical at least) the maximum USB performance wouldn't adversely affect the performance. Or maybe it was a cost consideration and we'll actually see these going for $30 as Anandtech seems to dream about.

 

[System Error Message]

Sorry, what?
USB 3.0/3.1 gen 1 is "5Gbps" or more correct, 4.8Gbps, whereas 3.1 gen 2 is 10Gbps and might actually reach it thanks to 128/132b encoding, rather than 8/10b encoding, as it should reduce overheads quite a bit.

This is obviously assuming that you have a controller that is connected to a fast enough bus. In the case of Intel's "real" 300-series chipsets (not the Z370), it would appear each 10Gbps port only has an 8Gbps interface (one PCIe 3.0 lane equivalent) to the chipset, which is a fairly obvious bottleneck. In other words, any USB 3.1 gen 2 controller that uses anything less than a two lane PCIe 2.0 interface won't deliver the 10Gbps performance.

I have to agree though, that it's odd that they chose to use a USB 3.0/3.1 gen 1 host interface for these, but maybe they tested and figured that the network overheads are high enough that the slight reduction in (theoretical at least) the maximum USB performance wouldn't adversely affect the performance. Or maybe it was a cost consideration and we'll actually see these going for $30 as Anandtech seems to dream about.

the first USB 3 ports that came out were 2.5Gb/s , and many USB-C and usb-3 ports are 5Gb/s despite the fact that there are usb c and 3 ports that are 10Gb/s. My laptop has both separate thunderbolt and usb-c ports but the usb-c port is a regular usb3 port at 5Gb/s just in a different form.

A lot of USB ports are directly connected to the chipset and do not use lanes, therefore the bottleneck is either the SoC (for stuff like your phones) or the bus between the northbridge and CPU which for laptops and PCs nowadays is shared with nvme and other things.

The problem is that very few manufacturers actually put a 10Gb/s usb3 on and some are lazy by using PCIe lanes rather than a fast bus like was in the past between the CPU and chipset.

 

[sfx2000]

Apple Thunderbolt 2 to Gigabit Ethernet adapter has definitely lower CPU usage (and better throughput) on MacOS X than any USB 3.x Gigabit Ethernet adapter (AFAIK it uses Broadcom chip)..

Indeed it is - Broadcom BCM57762 is used in the Apple Thunderbolt Gigiabit Ethernet adapter... I also have a TP-Link USB3.0 adapter, which uses an Asix AX88179 chipset (which works with Win/Linux/Mac/ChromeOS) for when I hook up a displayport monitor to my MacBook Air...

CPU utilization on the Thunderbolt adapter is roughly the same as the onboard PCI-e NIC in my older MacBook Pro (2012 non-retina) - the USB 3.0 is higher in utilization - performance for both is about the same.

 

[sfx2000]

USB 3.0/3.1 gen 1 is "5Gbps" or more correct, 4.8Gbps, whereas 3.1 gen 2 is 10Gbps and might actually reach it thanks to 128/132b encoding, rather than 8/10b encoding, as it should reduce overheads quite a bit.

Cool - if this helps...

I've got one machine that has TB3 and USB3.1 Gen2 - and they're on the same port - so the choice is a 40Gbps (TB3) or a 10Gbps (USB3.1g2) - but I can't have both. That being said, the box also has USB3.0 ports - key difference being that the USB3.1g2 and TB3 are run from the Alpine Ridge chip, and that USB is not shared with the USB3.0 ports which are provided by the PCH in the i5-7260U chipset.

My preference, given the choice, would be Thunderbolt - it's just more efficient... that being said, FireWire was also more efficient than USB, but USB won there...

 

[sfx2000]

On the other hand, Marvell was showing off a single chip 8-port PHY for speeds up to 5Gbps. So we might actually get to see some affordable 5Gbps products in the market this year

Given a choice between Realtek and Marvell, I'll take Marvell any time...

The conversation here is really about the Aquantia adapter and chipset - and this is a price/performance breakthru...

 

[TheLostSwede]

the first USB 3 ports that came out were 2.5Gb/s , and many USB-C and usb-3 ports are 5Gb/s despite the fact that there are usb c and 3 ports that are 10Gb/s. My laptop has both separate thunderbolt and usb-c ports but the usb-c port is a regular usb3 port at 5Gb/s just in a different form.

A lot of USB ports are directly connected to the chipset and do not use lanes, therefore the bottleneck is either the SoC (for stuff like your phones) or the bus between the northbridge and CPU which for laptops and PCs nowadays is shared with nvme and other things.

The problem is that very few manufacturers actually put a 10Gb/s usb3 on and some are lazy by using PCIe lanes rather than a fast bus like was in the past between the CPU and chipset.

You seem to live in a parallel universe... There has never ever been a 2.5Gbps USB 3.0 standard. I have met all of the USB 3.0 chip makers over the years and I was the first and I think only journalist that did a comparison of all the different USB 3.0 host controller available at the point when USB 3.0 was still new-ish. Unfortunately the publication I worked for back then has managed to lose all the pictures, so not much point in sharing it. I tested the the AMD A75 chipset, the ASMedia ASM1042, the Etron EJ168A, the Fresco Logic FL1009, the NEC/Renesas µD720200, the Renesas µD720201 and the VLI VL800 in 2011 and none of them had any other rating than 5Gbps. Back then it wasn't possible to reach the speeds of today, as the device controllers weren't fast enough. Fujitsu had a solution that could do 3.2Gbps/400MB/s. However, today it's not so hard with a nice fast SSD and a moder device controller.

I also don't think that you understand the technical aspects of how a motherboard chipset works, but let me see if I can help with that.
Below is a very rough layout of the available lanes/ports inside the Intel Skylake chipset (more recent chipsets from Intel adds USB 3.1 gen 2).

Image courtesy of Anandtech.

So as you can see, each of the USB ports that comes from a chipset are allocated one of 10 possible lanes. These lanes have the same width as a PCIe lane. So yes, the physical USB ports might connect directly to the chipset, but in the case of Intel, a USB 3.1 gen 2 port would only have one lane of PCIe 3.0 worth of bandwidth available to it, as it wouldn't make sense for Intel to make a custom interface for USB that's faster than PCIe 3.0. This means that if you have an Intel chipset, you'll never see the maximum performance from any chipset connected USB 3.1 gen 2 ports. I would hazard a guess and say that it's likely that AMD has the same limitation, but as they haven't provided any documentation, I can't say for certain. This means that only a couple of third part PCIe 3.0 x2 host controllers can offer near full USB 3.1 gen 2 performance.
More about Intel's 300-series chipsets and lane integration here - https://videocardz.com/76275/intel-releases-full-specifications-of-z390-chipset

Oh, one last thing, the interconnect between the chipset and the CPU today is simply four PCIe 3.0 lanes with a fancy name, nothing more, nothing less. So saying this is a fast interface is simply not true any more. Btw, a Northbrdige hasn't been used in PC architecture for around 8-9 years now, so I think you need to catch up on your tech know how.