Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Alibaba cloud FPGA: the $200 Kintex UltraScale+

essenceia.github.io

Post reply on HN
Print view View original

31–40 of 102 posts

Re: Alibaba cloud FPGA: the $200 Kintex UltraScale+

#31

Post by le-mark »

HarHarVeryFunny wrote: post #28

Is anyone aware of any examples of FPGAs used to implement neural nets or other connectionist architectures? 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 intel…

Why the downvotes? My understanding ai is not a use for fpgas because models require a lot of ram (preferably fast ram) and fpga themselves don’t have a lot of it. An fpga will have pins you can attach fast ram, but then board layout and layering traces becomes very non trivial. Hence gfx cards (and a lot of them) are more suited.

Re: Alibaba cloud FPGA: the $200 Kintex UltraScale+

#32

Post by throwaway48476 »

le-mark wrote: post #31
HarHarVeryFunny wrote: post #28

Is anyone aware of any examples of FPGAs used to implement neural nets or other connectionist architectures? 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 intel…

Why the downvotes? My understanding ai is not a use for fpgas because models require a lot of ram (preferably fast ram) and fpga themselves don’t have a lot of it. An fpga will have pins you can attach fast ram, but then board layout and layering traces becomes very non trivial. Hence gfx cards (and a lot of them) are more suited.

FPGA will never be as fast as an ASIC. NV data center chips cant do graphics anymore, just compute.

Re: Alibaba cloud FPGA: the $200 Kintex UltraScale+

#33

Post by phendrenad2 »

hmry wrote: post #3
mati365 wrote: post #2

What's special about this?

These chips usually cost $1k-$2k when bought new individually from the manufacturer

So people are going to desolder the chip? This doesn't seem very hobbyist-doable. Are there people out there who build products based on scrapped FPGAs from ebay?

Re: Alibaba cloud FPGA: the $200 Kintex UltraScale+

#34

Post by robinsonb5 »

bschwindHN wrote: post #26
duskwuff wrote: post #4

For what it's worth: if you have a FT2232H adapter hanging around (and if you don't, you should consider getting one), those adapters can easily be flashed to be compatible with Vivado: https://docs.amd.com/r/en-US/ug908-vivado-programming-debugg...

I _do_ happen to have one laying around from another project, weirdly enough. I've never done FPGA work though, when you say it's compatible with Vivado, what does that mean in practice? I can configure an AMD FPGA, or something more?

Vivado offers some embedded debugging and logic analyzer infrastructure, which allows you to tap into signals within the design and monitor them from the PC. (For instance, the Memory Interface Generator will report its initialisation status to the PC and if something failed it will tell you what). This is done over JTAG, and Vivado will only support that when using a select few JTAG dongles.

If you're happy to create your own debugging infrastructure (which isn't that hard - almost all FPGAs have at least two dedicated IR scancodes specifically for user applications) then you can use any JTAG dongle supported by OpenOCD.

Re: Alibaba cloud FPGA: the $200 Kintex UltraScale+

#35

Post by HarHarVeryFunny »

le-mark wrote: post #31
HarHarVeryFunny wrote: post #28

Is anyone aware of any examples of FPGAs used to implement neural nets or other connectionist architectures? 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 intel…

Why the downvotes? My understanding ai is not a use for fpgas because models require a lot of ram (preferably fast ram) and fpga themselves don’t have a lot of it. An fpga will have pins you can attach fast ram, but then board layout and layering traces becomes very non trivial. Hence gfx cards (and a lot of them) are more suited.

FPGA certainly not ideal for an experimental AI architecture, but if data is just flowing thru it (sensory input in, motor control out) then I don't see RAM as an issue. Could be an interesting option for experimenting on a mobile robot.

Re: Alibaba cloud FPGA: the $200 Kintex UltraScale+

#36

Post by HarHarVeryFunny »

throwaway48476 wrote: post #32
le-mark wrote: post #31

Earlier quoted context omitted.

Why the downvotes? My understanding ai is not a use for fpgas because models require a lot of ram (preferably fast ram) and fpga themselves don’t have a lot of it. An fpga will have pins you can attach fast ram, but then board layout and layering traces becomes very non trivial. Hence gfx cards (and a lot of them) are more suited.

FPGA will never be as fast as an ASIC. NV data center chips cant do graphics anymore, just compute.

Sure - not as fast or flexible, but relatively cheap and reprogrammable - good for experimentation, maybe for building a real-world proof-of-concept of a learning architecture ?

Re: Alibaba cloud FPGA: the $200 Kintex UltraScale+

#37

Post by taneq »

matt-p wrote: post #30
taneq wrote: post #21

Earlier quoted context omitted.

Also beware that FTDI has bricked swathes of gear that (often inadvertently) used counterfeit chips, via driver updates. So if you buy gear that uses an FTDI chip then if you’re unlucky want it turns out to be counterfeit, your thing might get bricked. :/

Good, FR. Then you can send it back to who you bought it from. All vendors should do this.

The issue is (as I understand it) that some early supply chain distributors either unwittingly bought shipments of counterfeit parts, or otherwise ended up stocking them as genuine. Then legitimate hardware manufacturers used them. And let's be upfront here, any part using FTDI USB interfaces is low volume hardware that's either indie or, like, niche scientific gear. These are not chips you use when you have a 20-person hardware procurement department with the time to do a deep dive into every shipment of every component. These are chips you use when you're used to serial ports and have had USB foisted upon you and just want to sell the thing you built.

Sending it back only punishes the OEM you bought the weird niche thing off, and they likely had no indication that the chips they paid genuine prices for were counterfeit. If they try to get a refund from Alibaba or DigiKey or whoever, they'll politely be told to fuck off, if they get a response at all.

I totally understand why FTDI chose to take this path, but I also think it was a poor choice overall and I now avoid FTDI products (which back in the day I happily used, before all this) just in case I get caught up in the BS.

Re: Alibaba cloud FPGA: the $200 Kintex UltraScale+

#39

Post by hmry »

stephen_g wrote: post #29
hmry wrote: post #3

Earlier quoted context omitted.

These chips usually cost $1k-$2k when bought new individually from the manufacturer

Yeah Xilinx’s pricing curve is incredibly steep. Probably makes sense since they’re used in things like defence (radar, electronic warfare) so they can sell them in low quantities for huge prices and people will still buy them. But with decent volume they get pretty cheap - for example there are video converters they have FPGAs in them that you can buy for less than you can buy the FPGA part in single qty - which mea…

Wanted to buy a GOWIN FPGA recently, like on Sipeed's dev boards. Mouser wanted $37 for it, more than twice what the whole devkit costs.

Meanwhile the list price from the manufacturer is $3, but it's completely impossible to order anything as an individual. Not sure why they designed a hobbyist beginner-friendly dev board around a part from a company that makes all their products impossible to buy...

Re: Alibaba cloud FPGA: the $200 Kintex UltraScale+

#40

Post by Aurornis »

HarHarVeryFunny wrote: post #28

Is anyone aware of any examples of FPGAs used to implement neural nets or other connectionist architectures? 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 intel…

> Is anyone aware of any examples of FPGAs used to implement neural nets or other connectionist architectures?

Do you mean as a specialized processor for high speed inference, or as if the neural net including weights was directly synthesized into the FPGA architecture?

Neither use case would be as powerful as other options. FPGAs don’t provide infinite leeway for implementing completely custom logic. They have a finite number of blocks that can be reconfigured in certain ways as synthesized by the software. This architecture isn’t a good fit for the large LLMs we think about which need a lot of high bandwidth memory access. You can connect an FPGA to high speed memory, but the current crop of GPUs are going to be much better at the job.

Post reply on HN