Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: FPGAs playing chess--an expert opinion

Author: Magnus Homann

Date: 02:37:30 12/23/99

Go up one level in this thread


On December 22, 1999 at 20:00:05, Greg Lindahl wrote:

>On December 22, 1999 at 11:19:34, Magnus Homann wrote:
>
>>And how long does it take to read the data back? You'll get about 1-10 MB/s on
>>the PCI bus, unless you watch out a bit.
>
>It takes 1 usec to grab the bus, and then you can transfer either way. Since the
>return value is tiny, the answer is: 1 usec.

Where did you get the 1 usec to grab the bus? It can be as low as 100 ns,
if there's no other activity. If there is other activity, it can be higher.

>For a graph of bandwidth vs. size of transfer, see:
>
>http://www.myri.com/myrinet/performance/DMAperf.html
>
>That page doesn't have any numbers for any Alpha 21264 systems, which are
>substantially better.

Depends on what you are talking to. The different numbers should indicate
to you that the transfer rate is implementation dependant. Both on the the
motherboard and the PCI-card.

>>Really? So how much data do you typically move?
>
>Well, Vincent says he needs 160 bytes for one eval. You might move N times this
>at once. So, an average answer from the page above is ~ 10 MB/s for 160 byte
>transfers, and > 50 MB/s for 1600 byte transfers.

...if the PCI-card can support it.

>
>> So far, I have no clue what you expect the FPGA to do.
>
>Neither do I, as I've said repeatedly. However, if you want to know the impact
>of the PCI bus on a design, it's well known. That was my only point in this
>little sub-thread.

Yes, it is well known to me. I just want to mkae sure that you consider all the
aspects of the PCI bus.

Homann



This page took 0 seconds to execute

Last modified: Thu, 15 Apr 21 08:11:13 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.