LattePanda Announces Sigma, a 'Hackable Single Board Server'

bit_user said:

I think 10 Gigabit Ethernet MACs were hard-hit by pandemic-related supply/demand issues. I've seen a lot of 10 Gigabit products either in short supply or with significantly elevated pricing. The situation finally seems to be recovering, but it's taken a long time to sort out.

It's getting harder and harder to believe that.

The AQC107 has been out for years and availabilty didn't suffer through the pandemic as I got a steady supply of PCIe x4 cards and the Thunderbolt variants before and troughout without price hikes. The ACQ108 (1-5Gbit/s) was a little less popular, because it really required NBase-T capable switches to support 5Gbit/s, when a lot of switches and PHYs only supported 100/1000/10000 Mbit, but not 2500 or 5000.

But again that stopped being an issue eight years ago, mostly because Aquantia co-developed the NIC and switch silicon and supplied Netgear etc. for cheap 8-port NBase-T switches. Admittedly I didn't try to buy switches during the pandemic, I had enough to last me throughout, only added AQC107 and 2.5Gbit clients, but unlike the data center equipment I didn't see them becoming unavailable.

The AQC113 offers 10Gbit using only a single PCIe v4 lane and was very slow ramping up. It might have suffered from the Marvel aquisition of Aquantia, but that doesn't quite explain why the only AQC113 plugin NICs available are PCIe x2 variants, a form factor that seems little more popular than PCIe x12 or 32, which are part of the official standard.

The AQC10x family seems to remain available and the AQC107 covers both PCIe v2 and v3 needs with the far more prevalent x4 factor quite nicely, so why are all available ACQ113 NICs all in the x2 form factor when nearly nobody offers x2 slots and the far "sexier" 10Gbit speed only requires a single v4 lane?

Last time I looked the highest-end Intel and AMD mainboards in the near €1000 range even had ACQ113 chips (using a single lane) on-board as well as one or two (closed end!) x1 slots, but the x1 plug-in NIC variant was nowhere to be had while x4 v5 slots where clearly much too dear and overkill for 10Gbit Ethernet.

Did I mention that I hate closed-end PCIe slots? They were terrible in the old days, when a PCIe x4 slot was often an unsued (and unusable) surplus while using the 2nd x16 slot e.g. for a x8 RAID controller meant bifurcating the GPU slot.

But they aren't better when the ACQ113 even as an x2 card would work perfectly in an open-ended x1 v4 slot, but these don't exist, cutting them open on a brand new mainboard is no fun (done that on those x4 slots) and even the x1 slots are disappearing alltogether (perhaps x1 NICs are finally coming?).

No, it would seem that really nobody wants to offer commodity 10Gbit on desktops at a bargain and it can't be the availability of the chips, because they are sold onboard and as plugin cards, just not in the single most sensible form factor of PCIe v4 x1.

abufrejoval said:

It's getting harder and harder to believe that.

Just look at what's happened to pricing of 10 Gigabit switches, for a datapoint that has almost nothing to do with the PC market. Last I checked (a few months ago), they still cost almost double the pre-pandemic prices.

abufrejoval said:

The AQC107 has been out for years and availabilty didn't suffer through the pandemic as I got a steady supply of PCIe x4 cards and the Thunderbolt variants before and troughout without price hikes.

The same can't be said for Intel and Broadcom. It's not like AQC107 is a drop-in replacement for them, so I think a lot of manufacturers tried to wait it out.

There was a motherboard with 2x 10 Gigabit MACs I was trying to buy that was unavailable for about a year. They even made a variant that swapped out Intel for Broadcom, but availability of that one was little better.

I have a separate datapoint on 10 Gigabit Intel NICs, which has suffered substantial price increases & availability problems.
https://pcpartpicker.com/product/cpndnQ/intel-x550-t2-network-adapter-x550-t2?history_days=730

bit_user said:

Just look at what's happened to pricing of 10 Gigabit switches, for a datapoint that has almost nothing to do with the PC market. Last I checked (a few months ago), they still cost almost double the pre-pandemic prices.


The same can't be said for Intel and Broadcom. It's not like AQC107 is a drop-in replacement for them, so I think a lot of manufacturers tried to wait it out.

There was a motherboard with 2x 10 Gigabit MACs I was trying to buy that was unavailable for about a year. They even made a variant that swapped out Intel for Broadcom, but availability of that one was little better.

I have a separate datapoint on 10 Gigabit Intel NICs, which has suffered substantial price increases & availability problems.
https://pcpartpicker.com/product/cpndnQ/intel-x550-t2-network-adapter-x550-t2?history_days=730

Geizhals.de confirmed no price hikes on Netgear ProSAFE XS500M during the pandemic, if anything it became cheaper (have a look at MAX data, which goes back to 2017). Prices fell below €50/port, which was inbelieveable during the decade before.

Same with the AQC107 in the very popular ASUS XG-C100C incarnation, cheaper during the pandemic than before or after with €80/port: again the Geizhals price data goes back to 2017.

10Gbit Ethernet was strictly considered enterprise and became a battle zone during the first two decades after its specification, it threatened to replace nice fat FC revenues with FCoE and vendors were going for fat offload ASICs that could also do really cool VM and network virtualization things. Of course everybody wanted to sell fiber links and GBICs so 10GBase-T was treated like a paria. It didn't help that it cost nearly 10 Watts per port originally to do the modulations made necessary by a medium that didn't couldn't just physically support higher frequencies.

The result were costly functional monsters with buggy drivers and VMware eventually gave up and put all their overlay network and software switch logic in software, because it only cost 2-3% of CPU overhead, but allowed rapid code changes, which ASICs didn't (today at 100/400Gbit they are reversing that again).

Many of the fabric vendors went bust, Intel survived relatively well with much less offload functionality, as it had never put their faith into really fat ASICs that much. Instead they prefer CPU side offload-engines on edge SoCs, which ASICs vendors obviously could not duplicate.

But with the hyperscalers 10GBit quickly became far too small and merchant silicon took over the switches at x0 and x00 Gbit/s, there was little money to be made at 10Gbit/s. Servers sell with DPUs these days that are 400Gbit/s going on 800.

Aquantia with NBase-T tried to change the game consumerizing >1Gbit Ethernet by reducing the power cost for modulation to around 3 Watts at 10Gbit/s, lower at 5Gbit and almost to 1Gbit/s levels for 2.5Gbit/s. They also offered fully integrated chips for NICs and switches (cascadable 4 port designs), which changed the price game significantly and essentially made 10Gbit "consumer ready"... except that consumers evidently weren't buying, perhaps because the only network they cared about was the broadband connection and East-West traffic never an issue.

Nobody at Microsoft or Intel was really interested in better office LANs, so somehow better than Gbit LAN never made it to the workplace, either.

Intel and and the big ASIC manufacturers refused for ages to do anything NBase-T, because they wanted to sell high revenue. When Intel finally did 2.5Gbit/s, they must have put interns to the job because they did it badly and buggy, while RealTek was, well, RealTek 2.5Gbit no better or worse than 1Gbit.

What Intel is still selling at 10Gbit is mostly chips they designed 15 years ago and might still make at 90nm somewhere. They just can't compete in power efficiency to Aquantia and BroadCom is even worse: they needed either active cooling or server chassis air-flow to survive when i tried them last more than a decade ago and I doubt they redesigned or shrunk their 10Gbit ASICs since then.

They are also mostly dual port chips by design requiring twice the single port PCIe lane allocations, because that's what everybody wants in servers, but doesn't care for in consumer hardware. Not that I've ever had a single Aquantia card fail on me: much lower power consumption evidently helps.

No, they are not drop-in replacements. But at least these days (since they went into the upstream kernel some years ago) they work much better out of the box with just about every Linux I've tried, while 2.5-10Gbit NIC drivers from Intel have become a true nightmare with hundreds of device IDs for dozens of chipsets and revisions and at least a handful of distinct drivers.

Aquantias support nothing fancy in terms of offloading, iSCSI or FCoE, they are basically 1Gbit technology scaled to NBase-T: exactly what I want for my home lab and now only on a 1x PCIe v4 slot (or a 20Gbit USB port).

abufrejoval said:

What Intel is still selling at 10Gbit is mostly chips they designed 15 years ago and might still make at 90nm somewhere. They just can't compete in power efficiency to Aquantia

No, I don't think so. The motherboard I wanted, and one of their more popular NICs, is based on the X550-AT2, which is 28 nm and launched in Q4 2015 (7.5 years ago). Its TDP is 11 W for 2 ports and it supports multi-gigabit.
https://www.intel.com/content/www/us/en/products/sku/84329/intel-ethernet-controller-x550at2/specifications.html
I'm not trying to have a **** match between Intel and Aquantia, as much as you seem to be trying to bait me into one. I'm just pointing out what seemed to be in short supply for much of 2021 and 2022. Take it as a datapoint, if you will.