• Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
    Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.

8P AMD possible with H8QGL?

O

overdoze

Weaksauce
Joined
Mar 14, 2012
Messages
83
I have been searchin high and low for this information. Unfortunately, I can not find any leads yet. According to the manual, the motherboard has an HT connector slot next to PCIE slot. I wonder if you can connect 2 motherboard together to make an 8P system.

I'm looking for more information about this HT connector system from supermicro. I just thought it would be helpful to increase production dramatically if you have 2 identical motherboard connected together.
 
We have been toying with the idea around here a bit. I believe pjkenned was playing with some infiniband hardware to explore that possibility. The main issue would be the software to manage it. Most of the software packages such as vSMP cost an upwards of $400+ per socket licencing fees. There is also a custom card that connects to the HT slot and can do hardware level linking. It is called "Numascale", but the cards are $1750+ each, and the fiber cables between the cards are not cheap either.
 
So who still has a spare kidney?

/looks at stiches
 
Untitledone said:
We have been toying with the idea around here a bit. I believe pjkenned was playing with some infiniband hardware to explore that possibility. The main issue would be the software to manage it. Most of the software packages such as vSMP cost an upwards of $400+ per socket licencing fees. There is also a custom card that connects to the HT slot and can do hardware level linking. It is called "Numascale", but the cards are $1750+ each, and the fiber cables between the cards are not cheap either.
Click to expand...

But what about using the slot already on those boards? You'd just need a cable and software...
 
Kendrak said:
So $500 for hardware lvl interface?

Or do we still need the software at 8 x $400 to make it work?
Click to expand...

Research is under way. No ETA on when we will know for sure. However there are cheaper cards that that. We found the larger profile version of the card that I linked ( 7140 is the low profile, and 7040 was the larger one ) for about $19 per card with free shipping. So the hardware is not so expensive for experimentation. More information will be had later on.
 
I'm following this closely, let me know if I can assist.

I noticed this while searching for the 7040 boards that you mentioned. A linux file check-in with this description:

Code: 
IB/ipath: Drop support for the original QHT7040 board  The original QHT7040 had
significant performance issues so there was an additional check in the driver for a newer
serial number.  Support for the small quantities of that board shipped has been dropped,
so this patch removes the special checks to simplify the code.

Signed-off-by: Dave Olson <dave.olson@qlogic.com>
Signed-off-by: Roland Dreier <rolandd@cisco.com>
Code checks for this:
Code: 
[COLOR=#808080][I]/*[/I][/COLOR]
[COLOR=#808080][I]* Later production QHT7040 has same changes as QHT7140, so[/I][/COLOR]
[COLOR=#808080][I]* can use GPIO interrupts.  They have serial #'s starting[/I][/COLOR]
[COLOR=#808080][I]* with 128, rather than 112.[/I][/COLOR]
[COLOR=#808080][I]*/[/I][/COLOR]
Probably a non-issue but you might be careful which used 7040s you purchase if you go that route.

Atlas
 
402blownstroker said:
No, the price for ScaleMP for g34 sockets is $800 per socket. $6400 in total for 8 sockets.
Click to expand...

has anyone asked them if they offer a cheaper "volunteer do gooder" license?

unlikely, I know, but has anyone asked?
 
402blownstroker said:
I asked, but I got zero response.
Click to expand...
Did you call somebody or did you just send a cold e-mail to their public "contact us" e-mails? companies that do most of their sales businness to business tend to be fairly non-responsive unless you talk to a real person. Maybe someone here has the right ties to have the contact info for a sales rep.
 
overdoze said:
I have been searchin high and low for this information. Unfortunately, I can not find any leads yet. According to the manual, the motherboard has an HT connector slot next to PCIE slot. I wonder if you can connect 2 motherboard together to make an 8P system.

I'm looking for more information about this HT connector system from supermicro. I just thought it would be helpful to increase production dramatically if you have 2 identical motherboard connected together.
Click to expand...

Instead of 2 board 4p interconnect through HT and see the system as 8p. Why don't someone write a wrapper to split down a WU in half and send other half through mpi to other 4p machine? I vaguely remember few concepts from my parallel programming class that took more than 10 years ago.
 
vincel said:
Instead of 2 board 4p interconnect through HT and see the system as 8p. Why don't someone write a wrapper to split down a WU in half and send other half through mpi to other 4p machine? I vaguely remember few concepts from my parallel programming class that took more than 10 years ago.
Click to expand...

Easier said than done...
 
extide said:
Easier said than done...
Click to expand...

At one point in time the cores where based on MPI. It has been a while since the cores where rewritten to use threads. I do not see them going back to MPI any time in the future.
 
[QUOTE A member of [H] black ops has ordered 4 of them to test with.[/QUOTE]


i like where this is goin :)
 
@vincel there is intercommunication between thread. Breaking off is not possible since SF is already breaking off to different WU for us to run on.

@all:
Thanks for all the responses. I can see this is going somewhere although pretty complicated. I may have to donate a kidney for the cause :)
 
The licensing costs per core are just ridiculous. And saddening :( I still have all my organs available...
 
I had the chance to try F@H on ScaleMP a while ago. I'll post some data when I find it so you don't have to waste any money trying. I was running multiple nodes and was not impressed. I do remember that the resultant theoretical PPD was lower than the sum of the PPD from the individual systems. As tear pointed out in the other thread, the move to the threaded client works directly against the use of any type of interconnect that's slower than memory.
ScaleMP is a very small company looking for very large customers so I'm not surprised that you didn't get a response.
 
Last edited:
This just in:
&#8194;&#8194;&#8194;&#8194;QHT's proboscis doesn't fit GL's receptacle :D

(need two flex PCI-E extenders: 1x and 16x)
EDIT: scratch that, not that simple, still investigating

(yet another) EDIT:
Best match: RSC-RR1U-HT

Not sure if PCI-E extenders and RSC-RR1U-HT are equivalent but I'll have a chance to determine that.
 
Last edited:
You must log in or register to reply here.
Back
Top