The Raspberry Pi CPUs were designed for smartphones, so they have a limited USB port that was only intended to support that use case.

To the extent these sorts of boards use CPUs that are cheap because they are made in high volume for other applications, we'll continue to suffer from IO limitations for what we want.

For 32 bit ARM, it seems to be shortcuts in the SoC design a.k.a there's a reason some boards cost under $50.

There exist some designs with real SATA (eg. Cubietruck) but those have poor CPUs (A7), and there exist many designs with SATA-via-USB which is never going to be fast. Also micro SD cards have abysmal performance for the sort of random I/O that operating system root filesystems have.

Luckily the situation in 64 bit ARM server land is much better. The APM Mustang and AMD designs have a combination of fast cores and properly engineered I/O subsystems. Real SATA, multiple 10gigE and 1gigE interfaces, PCIe, etc.

I would say it is not so much shortcuts as the fact that the 32-bit SoCs are generally designed for mobile. The only way to get a cheap devboard is to use an SoC which is being produced in high volume, so it doesn't have a prohibitive cost. That means you get the peripherals that a mobile phone or tablet wants, and not the ones that it doesn't, typically. If you insist on using an SoC custom designed for a devboard then the board is probably going to be tens of thousands of dollars.

64-bit is better because there are SoCs directly targeting server usecases which therefore have the kind of peripherals you'd prefer to see in devboards.

SoCs with 8GB onboard is a pretty niche chip, but other than that it's all out there. The Raspberry Pi/2 is actually a pretty poor offering - you can do better, even at the same price, The Pi Foundation just does a better job of marketing.

Specific models are the ODroid XU4, the CubieBoard or Cubietruck, the Banana Pi, etc.

They're not cheap, but the Jetson TK1 is an extremely powerful offering for this category based on the Tegra K1 SoC. You get SATA, USB 3.0, mPCIE, 1GigE, GPU with 192 Kepler cores, etc. If you need some grunt you can even do CUDA compute. Again, I'm still waiting for my Tegra-X1 based Jetson, NVIDIA :<

Demand mostly. Also the way ARM does I/O is a bit different, that said, a memory mapped "smart" disk controller is pretty straight forward, the trick is getting access to the memory bus crossbar switch inside the chip. Bolting one on, all of your peripherals go through one memory bus.

What would be really interesting is if ARM introduces a way to use a high speed serial bus (think PCIe) which would allow for the development of a 'south bridge' type IO chipset for ARM machines.

Because they're usually based on mobile phone or similar SoCs. However, at least 4 companies are promising to deliver Server-class chips this year - Cavium, Avago (ex Broadcom), Qualcomm, and AMD.

(Edited to add) Incidentally, most of the above are designed to have more than 8 cores. Cavium's ThunderX has 48, for example. Both Cavium and Broadcom previously delivered many-core MIPs designes, so this isn't vaporware either.

It really is just a question of SoC choice: the iMX53 (for example) has SATA, but in single quantities just the chip costs the same as a whole Pi.

the iMX6 adds PCIe 2.0.

>For example, the latest Raspberry Pi 2 ... handicaped by the IO speed which is limited to USB

The ODROID XU4 (referenced specifically by this link) has much better I/O throughput than Pi 2. It's still pretty affordable, too (75-80 USD).

> supporting at least 8GB

This is less common, but I figure it should arrive within 12-18 months.

I have a SATA SSD on my Cubietruck and it has gigabit Ethernet.

The Allwinner A80 seems a step backwards to me as they don't seem to have included SATA, though you would be able to address 8GB of DRAM.