Intel is halting development of the networking chip it got from Barefoot Networks

[erek]

The end of an era

“Intel is getting out of the business of making chips for networking switches as it seeks ways to cut costs after a dismal fourth quarter.”



Source: https://www.bizjournals.com/sanjose...halts-development-of-tofino-switch-chips.html

 

[Mega6]

It's amazing how much money Intel can squander away. No wonder their chips cost so much.

 

[1337Goat]

There is something so hilariously short-term about this decision that it's simply dark to ponder. Such is life in the world of business execs and marketers.

Don't they understand the possibilities and potential of 100+ gigabit networking? On the affordable consumer level, it won't be here for another 10 or 20 years probably, but this is the time to be developing it, ain't it? On the other hand, perhaps the Barefoot chip schematics were just too optimized for a specific configuration, and wasn't the sort of chip design that could be expanded upon or improved easily.

For example, imagine four pixels, like this ::
There's no open space inside for detail. It's just a square.

But now imagine four pixels on each side, such that there are four empty spaces in the middle. This might be a great design for a "circuit," because it's small, but has room for complexity. An electron can complete one circuit in exactly 12 nanoseconds; one nanosecond per pixel, and we can call that a hertz in this example.

However, if you were to use eight pixels on each side, such that there is now a very large middle area, things will become more tricky. Well, that's just how exponents work, and the reason why both the scaling of software and the Martingale betting strategy become very difficult to sanely utilize at larger magnitudes. With a bigger set of pixels, that 12 nano seconds is now 28 nanoseconds, and to get the same end result, we would now need to clock that chip at 2.34 hertz to reach a 12 nanosecond loop.

And of course, if we needed this to be very fast, and very precise, and also very powerful, then the complexity, the size, the clock speed, must all increase in such a way that a chip that is good at a very specific scale might completely fail at even the smallest attempt at an increase in its overall performance.

And this same scale is why overclocking used to be fun, and now it barely does anything. The chips are big and complex. The vivid difference between a nice Pentium 1 overclock from 75 MHz to 133 MHz nearly doubles the performance, right? Bingo. But take Intel's newest little chip, and how they brag about it reaching 6 GHz.... well, the old one got to 5.8 GHz, so the difference is not very impressive. That's not even close to doubling anything. (Although it's still impressive.)


And it's the same with networking, except for the troublesome fact that networking requires extremely precise clocks, making "overclocking" completely nutty when it comes to a NIC or something. It needs to run at a very specific speed based on the chip design, and if the chip changes, then it can only change to a precise standard. It MUST be exactly the same as the networking chip on the router side, such that 10 gigabit must communicate with 10 gigabit, and 2.5 gigabit must communicate with 2.5 gigabit. Trying to negotiate a networking speed requires this type of precision, and without it, the entire thing falls apart.


But who in this dark world has the comprehension and understanding of these concepts, but also a boring thing called an expensive university degree so they get noticed by one of the big companies? Well, not me hahah.

 

[GotNoRice]

It makes me sad, but on the other-hand, in the past, Intel was the option where you couldn't go wrong when it came to networking. Their gigabit adapters were #1 by far, and I still have computers using 20 year old Pro/1000 XT server adapters. In more recent years, there have been quite a few issues. They had a cablemodem chipset that had issues, making Broadcom-based modems the go-to choice instead. More recently, their 2.5G networking chipsets have been having issues, causing a situation that I never thought I would ever see, where Realtek is actually preferred over Intel.

 

[Lakados]

Intel makes great NIC’s but terrible switches. They don’t have the software stack to support them. Broadcom may have a few proprietary things that makes them not work well with others, but Aruba, Cisco, DLink, HPe, Juniper, … All use Broadcom at this point, so even if Intel makes the better NIC you get better results more than not with Broadcom because of better product stack compatibility. Intel bought good hardware and puts out solid drivers but they didn’t innovate the software and features, and that’s what’s killing them there. They could maybe catch up but it’s going to cost a lot and take too long to do it with no guarantees, sad but it’s better for them to cut their losses here.

 

[erek]

Intel makes great NIC’s

Not anymore as per

Intel 1226-V 2.5GbE Ethernet Chipset Showing Connection Drop Issues (Chipset Used on most Raptor Lake Motherboards)

https://hardforum.com/threads/intel...-most-raptor-lake-motherboards.2025009/unread

 

[Lakados]

Not anymore as per

Intel 1226-V 2.5GbE Ethernet Chipset Showing Connection Drop Issues (Chipset Used on most Raptor Lake Motherboards)

https://hardforum.com/threads/intel...-most-raptor-lake-motherboards.2025009/unread

As I explained there, 2.5Gb was released by manufacturers before there was an agreed standard and just as it was with wireless N it has resulted in a slurry of incompatible products. There are 3 different competing standards for 2.5Gb Ethernet and you can’t mix and match because of how they deal with signalling any interference causes them to fail at renegotiation it’s a PITA. Intels approach with the NBase-T standard is better and maintains better forward and backward compatibility with other protocols for PoE delivery and Adaptive vLAN assignment, but Broadcom and their MGBase-T standard has better market penetration because of how many devices use Broadcom for their controllers.

 

[Brackle]

As I explained there, 2.5Gb was released by manufacturers before there was an agreed standard and just as it was with wireless N it has resulted in a slurry of incompatible products. There are 3 different competing standards for 2.5Gb Ethernet and you can’t mix and match because of how they deal with signalling any interference causes them to fail at renegotiation it’s a PITA. Intels approach with the NBase-T standard is better and maintains better forward and backward compatibility with other protocols for PoE delivery and Adaptive vLAN assignment, but Broadcom and their MGBase-T standard has better market penetration because of how many devices use Broadcom for their controllers.

So basically you are making excuses for Intel for releasing a faulty product? No way around it. Intel fucked up on this one.

 

[Ranulfo]

Don't they understand the possibilities and potential of 100+ gigabit networking? On the affordable consumer level, it won't be here for another 10 or 20 years probably, but this is the time to be developing it, ain't it? On the other hand, perhaps the Barefoot chip schematics were just too optimized for a specific configuration, and wasn't the sort of chip design that could be expanded upon or improved easily.

They don't care or will wait until the speed is truly needed because of storage sizes and data drive speeds. They've been talking about 2.5-5-10g coming to SOHO for 8 years or so. Its been what 5, since Aquantia's stuff that seemed to work well. There are consumer mobos with 2.5Gb ports that are 2-3 years old now, the cheaper $100-150 models. Only in the last year have 2.5Gb switches dropped to sub $200 for an 8 port switch.

 

[jardows]

This makes me wonder if Barefoot would still be around and profitably making networking chips if Intel hadn't bought them. Seems to be the case often where a big company buys the smaller company, then ends up shuttering that business when it doesn't work out for them. Would be nicer if they would spin the company off and make it independent again. Alas, such is the way of the corporate world.

 

[erek]

As I explained there, 2.5Gb was released by manufacturers before there was an agreed standard and just as it was with wireless N it has resulted in a slurry of incompatible products. There are 3 different competing standards for 2.5Gb Ethernet and you can’t mix and match because of how they deal with signalling any interference causes them to fail at renegotiation it’s a PITA. Intels approach with the NBase-T standard is better and maintains better forward and backward compatibility with other protocols for PoE delivery and Adaptive vLAN assignment, but Broadcom and their MGBase-T standard has better market penetration because of how many devices use Broadcom for their controllers.

can't they gloss over the problem with software engineering? fix all the incompatibilities through the driver and offloading to the CPU? at least it will be compatible

 

[Mega6]

Hardware is always faster than software, and in networking - latency is king. Software can only go so far here. Everyone knows betamax was better than VHS, but look what won that war.

 

[Lakados]

can't they gloss over the problem with software engineering? fix all the incompatibilities through the driver and offloading to the CPU? at least it will be compatible

They sort of did. The 3'rd standard is 802.3bz which is the standard for 2.5gb and 5gb and it is compatible with both MGBase-T and NBase-T, but because of how it does signaling in actual buildings with air handlers, lighting ballasts, and junction boxes, etc ... unless you have your cable runs properly gapped, or are running solid cat5e it's a hot mess so for reliability you really need a minimum of Cat6a for it but switches that have the 802.bz are generally expensive. So then you find yourself in a price bracket where you are also getting 10Gb, and that standard is well-defined so why fiddle around with standards riddled with incompatibilities and technical difficulties when you then have the option to bypass them all and greatly increase your network speed? Ultimately for Enterprise right now the only place we even bother with 2.5Gb and 5Gb is for our new Wireless APs as PoE works really well there and they are brand matched to their switches so there aren't any compatibility problems there. Desktops can happily sit at 1Gb, with the needed workstations at 10Gb, because the cost of the manpower to troubleshoot and diagnose 2.5Gb greatly exceeds the costs of just bypassing it and going to something faster and far more reliable where needed.

 

[Lakados]

So basically you are making excuses for Intel for releasing a faulty product? No way around it. Intel fucked up on this one.

It's not faulty, it just doesn't work with Broadcom network chips unless that Broadcom device also supports 802.3bz draft 1.1 at the very least, which Broadcom doesn't include in their consumer or unmanaged small business switches. The catch here is Broadcom is the supplier for most of the cheaper networking switches, so unless you are dropping $3K+ on a Broadcom-powered switch they don't include that standard, so if you then run into a situation where there is any signal interference how the MGBase-T standard used by Broadcom and the NBase-T standard used by Intel negotiates differently as they use a different signal timing, because the timing doesn't match they will just fight it out and drop packets in the mean time, they then also fail to auto-negotiate speeds so it won't even drop down it usually pops over to "Network Cable Disconnected" instead. The 2.5Gb spec is just problematic garbage that should be avoided, and we are seeing it now because 2.5G is a bigger number that sounds more impressive than 1 and there is a huge surplus of 3+ year-old chips that didn't sell so Intel, Broadcom, and the others have lots of old stock they just want gone so they are basically giving them away to the OEM's because Enterprise doesn't want them because they are too problematic and the hidden costs of supporting them far outweigh the costs of bypassing them.

 

[erek]

They sort of did. The 3'rd standard is 802.3bz which is the standard for 2.5gb and 5gb and it is compatible with both MGBase-T and NBase-T, but because of how it does signaling in actual buildings with air handlers, lighting ballasts, and junction boxes, etc ... unless you have your cable runs properly gapped, or are running solid cat5e it's a hot mess so for reliability you really need a minimum of Cat6a for it but switches that have the 802.bz are generally expensive. So then you find yourself in a price bracket where you are also getting 10Gb, and that standard is well-defined so why fiddle around with standards riddled with incompatibilities and technical difficulties when you then have the option to bypass them all and greatly increase your network speed? Ultimately for Enterprise right now the only place we even bother with 2.5Gb and 5Gb is for our new Wireless APs as PoE works really well there and they are brand matched to their switches so there aren't any compatibility problems there. Desktops can happily sit at 1Gb, with the needed workstations at 10Gb, because the cost of the manpower to troubleshoot and diagnose 2.5Gb greatly exceeds the costs of just bypassing it and going to something faster and far more reliable where needed.

how bad is the performance hit on throughput and latency for doing this?

do the drop outs stop now?

 

[Lakados]

how bad is the performance hit on throughput and latency for doing this?

do the drop outs stop now?

We gave up, we now only use 2.5 for our Wireless AP’s and they are all direct run with Cat6A and as they are brand matched to the switches so we have no compatibility issues and no packet drops.

We use Intel NIC’s for all our 10Gb stuff which we use internally as Multi-Mode fibre for the building runs or DAC cables for anything within the same rack or room.

2.5 was just too problematic and we don’t trust it. So I don’t really have any data more recent than 2020 because we tore that stuff out while we were shut down for COVID because I needed to find jobs for the CUPE staff or risk loosing them and nobody wanted that.