Hacker News Reader

Comment by akiselev | Hacker News Reader

Relative to custom silicon, yes, FPGAs are way more power hungry. However, there are many things you can do much more power efficiently in an FPGA than on CPU or GPGPU like video decoding/encoding, crypto, digital signal processing, and many other real time critical systems. FPGA sits in a price/performance niche that has high unit costs and power requirements but little upfront investment compared to custom silicon and a little more than general computing but with the benefit of throughput/latency

View on HN

Where do FPGAs sit for running neural networks? Or, does it make sense to manufacture Field Programmable Node Arrays rather than Field Porgrammable Gate Arrays?

I think that GPUs are already basically field programmable node arrays. Maybe for certain instances FPGAs would be better but the neural network would have to be designed at the logic gate level so it'd be significantly more costly in engineering time than a algorithm running on the GPU or CPU. FPGAs are for relatively simple logic running at incredibly high throughput with very low latencies. It's extremely common to move more complicated logic to a "soft" CPU on the FPGA that runs C code compiled to ARM/SPARC/POWER etc. and controls the low level hardware.

GPUs are getting closer to being math- and draw-optimized FPGAs. I don't think they'll be separate products too far in the future.

GPUs and FPGAs are very far from each other architecturally and are good at very different things. They will certainly be separate products for the next 3 decades as they were for the last 3 decades.

At the desktop/server level, I agree, which is why I don't get the whole Intel/Altera synergy thing.

But in more specific embedded cases -- meaning the vast majority of FPGA applications -- they are nothing alike. If I need to run a multichannel digital downconverter on a power budget of 1 watt, no GPU available in the foreseeable future is going to handle that role as well as an FPGA will.

Perhaps someday a hypothetical I/O-optimized GPU will be able to do the jobs that are currently done on FPGAs and ASICs, but definitely not now.

I don't get the whole Intel/Altera synergy thing

Let's recall a previous acquisition, one that I totally didn't understand. An analyst had a really good quip about it:

JPMorgan analyst Christopher Danely upgraded Intel to overweight following the company's earnings, although he still struggles to reconcile Intel's recent acquisition spree.

"Intel might as well have bought Whole Foods," he said of the McAfee deal.[1]

Compared to McAfee, acquiring Altera was brilliant.

As a public company Intel is always facing scrutiny from Wall Street. They need to constantly increase both revenue and earnings. What better way than via a strategic acquisition? Instant growth.

And Altera was a logical target. Altera sells high-margin silicon. Intel sells high-margin silicon. Intel was even fabbing some of Altera's parts, so they were quite familiar with the company. And maybe some smart people could eventually think of some synergy. E.g. once upon a time Intel made the IXP1200, a "network processor". What if, instead of designing custom silicon for something like that, perhaps a CPU + FPGA on a single die would be sufficient?

Makes much more sense than Intel acquiring McAfee.

[1] http://www.reuters.com/article/us-dealtalk-intel-idUSTRE73K7...

I think they'll continue to be seperate products far into the future because they're fundamentally different approaches. Aside from anything else, FPGAs are optimized for latency whereas GPUs are optimized for overall throughput.

Can someone explain the "intersection" of neural networks and FPGAs/GPUs. I feel like I'm seeing this referenced a fair amount but haven't heard a good explanation on why this hardware is attractive for these workload.

Because neural networks use "simple" math that is massively parallelizable. Your computer's CPU may have eight cores, but my older, mid range graphics card has over a thousand cores.

If your work can be broken down into a thousand peices that can be computed independently, than using a GPU can get you ludicrous speedups.

This is a bit of an over simplification, because the 'cores' of a GPU aren't really comparable to a CPU core. It's more like scaling up a factory - at a single point on the production like you go from 1 person doing a job to 100 people doing the same job with the same tools. It means you can do more of the same thing, but can't do 100 completely different things.

Here is a good overview:

http://gamedev.stackexchange.com/a/17255

Thanks

I think it is something people just like to say because both sound a big magical.

Neural nets are matrix multiplications and addition. They are generally memory constrained on GPUs and Main <> GPU memory is also a bottleneck.

I don't see how an FPGA will really help there.

I can't find the link right now, but I remember reading a more recent article on NextPlatform.com about how FPGAs are much more memory constrained than GPUs are, which essentially negates any slight power efficiency advantage they might have over GPUs.

Also, don't FPGAs bring along additional operational constraints ? Modern large/vast FPGAs will take anywhere from 2 to 4 hours to be Pd in the F. This makes managing new releases even more challenging.

"Pd in the F"? What does that mean?

FPGA = Field Programmable Gate Arrays, Pd in the F => Programmed in the Field

what about FPGAs allows these operations to be more power efficient on than A GPU or even CPU? Could you elaborate. I found your statement interesting. Thanks.

Essentially, FPGAs are as "close to the metal" as you can get - they are the metal - which means you can configure circuits that utilize the minimal number of 'gates' needed to go from a specific set of allowable input to a specific set of allowable output.

To make a programmable gate, however, requires more power-inefficient transistor designs, usually using more transistors than an equivalent gate on a general purpose CPU (e.g. a programmable gate set to be an OR gate will be less efficient than an OR-only gate).

FPGAs are therefore more power-hungry per gate, but since you don't need the overhead of transistors that can process arbitrary input to arbitrary output, you can make them very efficient for particular tasks.

Reply on news.ycombinator.com