FPGA Central

Hello,
at present I am working on an Infiniband implementation on a Virtex-II
Pro. As far as I know the build-in RocketIOs are fully compliant to
the Infiniband standard (beacon and variant CRC excluded). There are
two types of CRC, the invariant and the variant type. Although the
RocketIOs are able to calculate at least one type, i.e. the invariant
CRC, you cannot make use of it in this particular case. The reason for
this is that the invariant CRC is needed for the variant CRC type
which you have to handle yourself, anyway. Has anyone ever tried or
finished an Infiniband implementation inside the V2P/RocketIOs and
could share his experience with me? Or is there any example
implementation for the whole CRC calculation inside the FPGA fabric
available? I wonder how one could ever reach the typical 2.5 Gbit/s
Infiniband transmission rate when almost all functions must be handled
inside the FPGA fabric? Still some other questions: Could someone
please help me on the question how often the beacon signal should be
repeated before synchronisation is sufficiently established? And what
about the packet generation? There are several IBA packet structures.
If I assume a plain point-to-point connection it surely would not be
necessary to include all that global routing stuff inside the header.
So what would be the adequate format to use? The "local packet"
structure or maybe just "raw packets"? If someone got the slightest
clue, please post a write an email! Any answer is appreciated. Best
regards, Bruce

Keywords: FPGA, Xilinx, Virtex 2, Virtex2, Virtex-2, Virtex II,
VirtexII, Virtex-II, RocketIO, RocketI/O, RocketIOs, RocketI/Os,
Rocket IO, Rocket I/O, Rocket IOs, Rocket I/Os, Virtex2Pro,
VirtexIIPro, MGT, SerDes, Gigabit Transmission, Infiniband, Beacon
Signal, Synchronisation, Gbit, Gb, Gbit/s, Gbits, Gb/s, Gbs, P2P, IBA

On Fri, 13 Aug 2004 06:09:36 -0700, Bruce wrote:

> Hello,
> at present I am working on an Infiniband implementation on a Virtex-II
> Pro. As far as I know the build-in RocketIOs are fully compliant to
> the Infiniband standard (beacon and variant CRC excluded). There are
> two types of CRC, the invariant and the variant type. Although the
> RocketIOs are able to calculate at least one type, i.e. the invariant
> CRC, you cannot make use of it in this particular case. The reason for
> this is that the invariant CRC is needed for the variant CRC type
> which you have to handle yourself, anyway. Has anyone ever tried or
> finished an Infiniband implementation inside the V2P/RocketIOs and
> could share his experience with me? Or is there any example
> implementation for the whole CRC calculation inside the FPGA fabric
> available? I wonder how one could ever reach the typical 2.5 Gbit/s
> Infiniband transmission rate when almost all functions must be handled
> inside the FPGA fabric? Still some other questions: Could someone
> please help me on the question how often the beacon signal should be
> repeated before synchronisation is sufficiently established? And what
> about the packet generation? There are several IBA packet structures.
> If I assume a plain point-to-point connection it surely would not be
> necessary to include all that global routing stuff inside the header.
> So what would be the adequate format to use? The "local packet"
> structure or maybe just "raw packets"? If someone got the slightest
> clue, please post a write an email! Any answer is appreciated. Best
> regards, Bruce
>
> Keywords: FPGA, Xilinx, Virtex 2, Virtex2, Virtex-2, Virtex II,
> VirtexII, Virtex-II, RocketIO, RocketI/O, RocketIOs, RocketI/Os,
> Rocket IO, Rocket I/O, Rocket IOs, Rocket I/Os, Virtex2Pro,
> VirtexIIPro, MGT, SerDes, Gigabit Transmission, Infiniband, Beacon
> Signal, Synchronisation, Gbit, Gb, Gbit/s, Gbits, Gb/s, Gbs, P2P, IBA

We have a 4X IBA core available that works with both Rocket IO and
external SerDes.

http://www.polybus.com/ib_link_layer_website/

On 13 Aug 2004 06:09:36 -0700, [email protected] (Bruce) wrote:

[snip]
>Or is there any example
>implementation for the whole CRC calculation inside the FPGA fabric
>available? I wonder how one could ever reach the typical 2.5 Gbit/s
>Infiniband transmission rate when almost all functions must be handled
>inside the FPGA fabric?

Hmmm. I worked on designs last century that performed CRCs on
10Gbit/s data streams. FPGAs have improved since then, perhaps
doubling in speed. 2.5Gbit/s sounds rather easy, assuming you know
the tricks.

I tried to summarise the high speed CRC tricks in this post:
http://groups.google.com/groups?thre...sdo4%404ax.com

Regards,
Allan.