Tyan FT72-B7015 fail to POST when installing more than 4 GPUs

[Mario]

Hi,

I have a problem with Tyans FT72-B7015 when try to install GTX 970 gpus. I have tried it on two systems (same Tyan b7015).

On one system I managed to install 5 gpus but with 6th put in, the system does not go through post. Fans just spin, then go slower and the system just stays there.

On second system (the same parameters) I have the same problem but with more than 4 gpus installed. When I put 5th the system does not go through post.

GPUs are ok, tested. It does not matter in wchich pci-e slot I put gpus.
Slots are ok. Both systems work very smooth with 4 GPUs.

Any ideas why it happens? I would be greatful for help.


Regards
Mario

[Paul]

A legacy BIOS has limited "low memory" for loading
Option ROMs.

http://forums.2cpu.com/showthread.php?threadid=29994&highlight=loading+scsi+bios

Formatted for clarity:

Q: Why doesn't the BIOS (Option ROM) for my PCI device show
during boot up?

A: The cause of the problem is that in order to be PC Compatible,
the Option Rom space is limited to 128K. This is true for any
motherboard with PC compatible BIOS'.

In the common configurations, a newer AGP card (such as any
GeForce4) will require 64K of Option Rom space, so you have only
64K of Option Rom space left to work with for other devices.
Many SCSI, NIC (w/ PXE), IDE Raid and etc., can easily use
another 40 to 64K of Option Rom space for their needs.

By design the Option Rom should shrink down to a smaller run time
code after the initialization code has run. For example, some Adaptec
cards will require 32K to initialize. Then they shrink down to
12K at run time; whereas some GeForce4 cards require 64K to initialize
and never release to a smaller amount. Please check with the device
manufacturer for the latest firmware upgrade or ask if they have a
smaller Option Rom available. Again this is a limitation of the
PC compatible specification and not a failure of the motherboard
BIOS itself.

Another example here.

http://docs.oracle.com/cd/E20881_01/html/E20896/sbxis.glink.html

*******

It could be a power issue. Make sure the redundant power subsystem
is properly cabled to provide enough power. Perhaps someone has
unplugged some power supply from it. If the thing has BMC, maybe
it can say something about power supply config.

The box in question, has a distribution bus for +12V. And Tyan
would have tested with (8) M2050 Tesla 225W cards installed.
Perhaps there is some difference in the VBIOS on a Tesla card,
versus a gamer 970 GTX ?

If a Tesla has no graphics output, perhaps the Option ROM
is not of conventional design (isn't a VESA BIOS or a GOP BIOS).

I was not able to get *any* tidbits on a Tesla card,
to spot differences like that. I'd have to own one to
figure it out.

Paul

[Mario]

Hi Paul,

thank you for your clues/tips.
We are sure that the power suply is properly cabled.
We will try to check if it is the Option Rom issue also try to check power issues via BMI. I will let you know if the problem is solved or we need more help

regards

[Paul]

If you can, try to contact Tyan tech support, as
they might have your issue on the top of their list.

For high-runner faults, or in cases where the
documents were woefully lacking, a lot of customers
end up phoning with the same problem.

I don't think it's necessarily power, at least at
first. Modern cards with GPUs, have a relatively
large ratio between "compute" mode and idle mode.
A card might draw 10W when down-clocked to 100MHz
idling, and draw 225W when flat out doing
shader computations. I doubt the initialization
sequence would trigger enough of a "power glitch"
to trip the thing right away.

The same goes for CPUs. There's no reason for them
to draw 156W during POST, as only one core might
be executing the BIOS POST sequence.

And while you could attempt to debug the problem
with a Port 80 POST card, I don't think I've ever
run into a case where that led to a resolution.
Typically a hex code comes up on the card
display, that the documents say is "reserved",
and then you've got nothing to work with. A
Port 80 card gives "progress codes" rather than
"error codes", which is why it is worthless.
Some hardware people, use them as "go-nogo" - if
the card has any decent number (not 0x00 or 0xFF)
on it, it means the CPU is executing code. And that's
all it does for you.

Paul

[wolni...@lighthousestudio.pl]

Hi Paul,

thank you for your help. We finally had a chance last week to try again to solve the issue.

Upgrading BIOS to 1.06D Beta helped.
Servers go through POST, see all 8 GPUs and run smoothly

Thanks very much again!
Mariusz