Alibaba Cloud Fpga: the $200 Kintex Ultrascale+

https://docs.amd.com/r/en-US/ug908-vivado-programming-debugg...

Basically they achieved massive improvements in throughput and latencies vs RocksDB, for some operations even an order of magnitude.

They even offered this as a service, but eventually pulled out. Not sure if they are still using it, but I was surprised that pretty much nobody else, as far as I am aware, is using any hardware acceleration for repetitive database workloads.

A Kintel UltraScale+ is quite a prize though, really nice write up.

But why? Are these boards frequently removed and physically moved around in data centers?

Or was that an eBay seller thing to protect it?

I guess there aren't any GPIO's wired up to headers or such? Any suggestions on a project for it that's a good fit?

dmesg log :

[ 0.388790] pci 0000:00:00.0: [14e4:2712] type 01 class 0x060400

[ 0.388817] pci 0000:00:00.0: PME# supported from D0 D3hot

[ 0.389752] pci 0000:00:00.0: bridge configuration invalid ([bus 00-00]), reconfiguring

[ 0.495733] brcm-pcie 1000110000.pcie: link up, 5.0 GT/s PCIe x1 (!SSC)

[ 0.495759] pci 0000:01:00.0: [dabc:1017] type 00 class 0x020000

...

Since most people might not be intimately as familiar with PCIe terminology, allow me to quickly document what is going on here.

0000:00:00.0: is the identifier ...

[14e4:2712]: is the device’s [vendor id:device id], these vendor id identifiers are assigned by the PCI standard body to hardware vendors. Vendors are then free to define there own vendor id’s.

The full list of official vendor id’s and released device id can be found : https://admin.pci-ids.ucw.cz/read/PC/14e4 or in the linux kernel code...

"""

Wouldn't it be really nice if dmesg had more info about devices?

What's exciting about FPGAs is that they make "custom" chips affordable for colleges and even individuals... At this point in time, with the AI world fixated on LLMs, it seems that there's an opportunity for small groups/individuals to use FPGAs to experiment with more animal-like ground-up approaches to intelligence.

What's cool is that the Raspberry-PI V is just fast enough to run modern desktop software, including (ack!) Microsoft Teams. I was able to give a live demo of my FPGA design by sharing the desktop during a conference call.

My only complaint is, of course, lack of SoC documentation from Broadcom.

Say what you will about Intel, but they do provide full chip documentation if you go through their NDA process. With it, I made a very nice FPGA PCIe environment on Xeon Ivy Bridge: it let me save all of the PCI config registers, then reconfigure the FPGA, then reprogram them and have the chip show up on the bus without having to reboot the server or re-enumerate. The trick is to temporarily set the root complex bit to disable PCIe error detection during the reconfiguration process.

(Was using Altera Stratix-IIgx at the time).

There is possibly another way to this, but haven't tried it:

https://stackoverflow.com/questions/32334870/how-to-do-a-tru...

But even aside from that, with the documentation you have access to a lot of error reporting which is extremely useful during PCIe bring-up.

SFP transceivers are just that- transceivers, they will spit out what you put in, so you could even build some interesting bespoke fiber optic signaling gear with these, and a PCIe backplane, or just solo.

This ai craze makes me think that:

- there's not much of a moat long term in inference hardware (can't comment about training)

- the models themselves are a commodity that can be swapped any day

- the products are coming but not at the pace one thought. It's been 3 years since Chatgpt launched and we still have no decent service where I can inject documents and ask questions.

Reminds me of recycling used Mining- or LED-Controllers as they can be bought on Ali and have quite capable FPGAs.

The ebay listing mentions 128 MB, but I wonder if this can be trusted since some other information in the "item specifics" overview is clearly bogus ("7-Function Digital Multimeter Shield").