Microsoft's FPGA Translates Wikipedia in less than a Tenth of a Second

> What other content would you like to see?  

They claim something impressive ("Translate Wikipedia in less than a Tenth of a Second") but give no details about the task nor the system.

If the claim is not total marketing nonsense I would imply that they mean translating from one language to another (e.g. English to German).

From the article link (and the picture) you could also imply that one FPGA (or the card in the hand of the guy) does this. But this is simply unbelievable. So the question is: How many FPGAs are involved? With out this, the claimed time is simply not meaningful, as double the number of FPGA will mean half the time (every Wikipedia article can be translated individually, so it is easy to execute the task in parallel...).

But I guess this is all not Microsoft's fault, but the problem of that specific link. I found following which gives much more insight at the end of the page:
https://www.top500.org/news/microsoft-goes-all-in-for-fpgas-to-build-out-cloud-based-ai/

There it says that 4 FPGAs (Stratix V D5, ca. 500k LE) would require 4 hours to translate Wikipedia. The 0.1 seconds are achieved with a huge cloud of such FPGA equipped systems...

Of course still impressive, but not the same as most people might think after reading the headline. (And it also makes me wonder about the future of the Altera/Intel low cost FPGAs, when to want to sell a Stratix into every server...)

Regards,

Thomas

www.entner-electronics.com - Home of EEBlaster and JPEG Codec

On 2016-10-18, quiasmox@yahoo.com <quiasmox@yahoo.com> wrote:
>
> I did read it. It was no more specific than the headline.
>

It could be that it was merely translating from wiki markup to HTML


-- 
This email has not been checked by half-arsed antivirus software 

On 10/19/2016 7:47 PM, thomas.entner99@gmail.com wrote:
>> What other content would you like to see?
>
> They claim something impressive ("Translate Wikipedia in less than a Tenth of a Second") but give no details about the task nor the system.
>
> If the claim is not total marketing nonsense I would imply that they mean translating from one language to another (e.g. English to German).
>
> From the article link (and the picture) you could also imply that one FPGA (or the card in the hand of the guy) does this. But this is simply unbelievable. So the question is: How many FPGAs are involved? With out this, the claimed time is simply not meaningful, as double the number of FPGA will mean half the time (every Wikipedia article can be translated individually, so it is easy to execute the task in parallel...).
>
> But I guess this is all not Microsoft's fault, but the problem of that specific link. I found following which gives much more insight at the end of the page:
> https://www.top500.org/news/microsoft-goes-all-in-for-fpgas-to-build-out-cloud-based-ai/
>
> There it says that 4 FPGAs (Stratix V D5, ca. 500k LE) would require 4 hours to translate Wikipedia. The 0.1 seconds are achieved with a huge cloud of such FPGA equipped systems...
>
> Of course still impressive, but not the same as most people might think after reading the headline. (And it also makes me wonder about the future of the Altera/Intel low cost FPGAs, when to want to sell a Stratix into every server...)

For sure the release is short of engineering data... it *is* a marketing 
pitch.  The point is they plan to be providing a combination of FPGA and 
CPU which will run much faster and use less power than the CPU alone. 
No, they aren't offering hard numbers and the task of translating 
wikipedia is not really the best benchmark for serving up or searching 
web pages.  It is meant to offer a metric that even laymen can relate to.

In other words, it's meant to sound good to those who would not 
understand more engineering information.

Microsoft has no incentive to sell FPGAs.  Their incentive is to provide 
the software on faster hardware.  If the hardware doesn't pan out, 
Microsoft gets nothing but expenses.

-- 

Rick C