Cluster computer the Turing Pi 2 has just launched on Kickstarter, and with 29 days still to go has breezed past its $64,000 goal with over $1 million pledged at the time of writing. An extra special addition to the Turing Pi 2, as reported by CNX Software, is that it supports a brand-new module, the Turing RK1, which features the Rockchip RK3588 octa-core processor and up to 32GB RAM.
The board is already circulating among tech YouTubers, including friend of the Tom’s Hardware PiCast Jeff Geerling (above). Rapidly becoming something of a heavyweight in the SBC space. The RK3588 sports four Arm Cortex-A72 cores and four A55 cores. It also carries a neural processing unit (NPU) capable of six TOPS, making such a cluster more attractive to the machine learning crowd. The full spec of the RK1 isn’t ready yet, but we expect there to be some flash storage on board, and PCIe 3.0 in some form.
You’re not stuck with the RK1, however. Raspberry Pi Compute Module 4s can be used via an adapter (if you can find them), as the Turing Pi 2 features four SODIMM slots into which you can secure your modules via an adaptor. You can use Nvidia Jetson Nano, TX2 NX, and Xavier NX modules too, or mix and match a combination that suits your needs.
The modules are connected via a Gigabit Ethernet switch that’s built into the mini-ITX board, with enough bandwidth left over for a pair of gigabit RJ-45 ports on the edge of the board. These are complemented by a pair of USB 3 ports (with more available via a header), two mPCIE sockets including one with a SIM card slot for connecting third-party extension boards such as wireless networking or home automation transmitters, two SATA III ports, an HDMI and a MIPI DSI header. There's a 40-pin GPIO header too.
Power comes via a 2-pin ATX header, and consumption is said to be less than 60W when using Nvidia modules. Board management will be via open-source firmware, and the cluster seems perfect for exploring self-hosted home automation applications, photo galleries or media servers, even a Minecraft server.
Closed our funding goal in less than 3 minutes 🚀$200k in 15 minutes$400k in 30 minutesYou are awesome!https://t.co/yIzx4ooY2c#turingpi2launch pic.twitter.com/RkTu3oC78IMay 16, 2022
The Turing Pi 2 is available from $219 for a bare board, with shipping available anywhere in the world. Raspberry Pi CM4 adapters are $12. Remember that crowdfunding a project is not a guarantee of receiving a finished product. Backing a crowdfunded project is akin to an investment, you believe in the project and want it to succeed. You are not purchasing a retail product.
It is really little more than an "uncluttering device" to put 4 PIs (or Jetson Nanos or yet-nots) in a box, while such modules are not included: ~$200 may seem a little much, by the time it's delivered--or RP4's are back in stock.
And for that kind of money you can in fact get a Jasper Lake N6005 based Atlas Canyon NUC that includes a SoC which will easily deliver the same class of compute power in a smaller form factor and finally can be purchased (I really want Gracemonts now!).
It expands to 32GB of RAM and allows you to run that cluster in VMs with very likely similar compute power but much greater flexibility and a much faster interconnect, because RAM is faster than any Ethernet, let alone Gbit. 8 PCIe 3.0 lanes may be less flexible than I'd want, but NVMe storage seems included while 2.5/5/10-NBase-T onboard Ethernet may be a pipe dream yet.
And please remember that clusters are really about fault tolerance first and scale-out computing second: the Turing PI 2 board is fully populated with things that can fail! So if resilience is your goal, you'd be much better off using individual RPs and dual switches.
The VM based cluster will offer just as little physical fault resilience and it won't nearly have as many blinking LEDs, but it might be much more usable for real workloads, once you figured out that training machine learning networks on a cluster of meek CPUs is a) difficult b) slow c) terribly inefficient. Even for inferencing you'd want real NPUs, which none of the available modules provide.
The PI doesn't have any ML acceleration that I know of, Jetson Nanos do kinda, but at levels that are 4-6 generations back and who knows what MediaTek software support will look like, once hardware trickles down to the hobby sector.
Yes, the Jasper Lake won't be great for ML either, but if that's really your thing, at least you can add a Movidius stick or board with some software SDK support. And for ordinary Linux or Windows workloads, Jasper Lake will be both terribly boring (things just work!) and much more flexible: I'm pretty sure you could even run MacOS in a VM, if that's your fancy, even if I wouldn't know why anyone would want to bother.
IMHO the Turing Pi 2 is e-trash about to happen, completely irrational and even borderline as a tinkerer's toy.
I own a Jetson Nano and an RP4 8G and I'll concede their educational value, individually.
But mostly they taught me that almost everything else either much more efficient (mobile SoCs) or much faster ("notebook SoCs"): they are truly underperforming across all use cases, but at least quiet about it.
So going beyond one lesson learned and building clusters from underperformers, only multiplies disappointment and underperformance, it doesn't make them any better or more usable.
You're welcome not to believe me, so please no flames. But please waste a few brain cycles on this, before buying.
Thankyou for putting exactly what I was thinking, but more eloquently!