Post
Topic
Board Hardware
Re: BFL announces 28nm 600GH/S blade for $4680
by
BFL-Engineer
on 18/08/2013, 00:13:54 UTC
This is hilarious... I actually have some BFL hardware and even I think this means they are finished..  They are obviously trying to move payments to a medium that cant be refunded (bitcoin or wire xfer).  They are using a design which alot of us *know* cannot dissipate that much heat, and they are moving at a snail's pace with current orders.

Guys... Ive never said this before, but i believe they are on the virge of folding and taking anyone who preorders money with them..

Regarding power consumption, Radeon 5970 and 5870 both consume more power than our card does, the very reason we took this design approach.

TDP of 5970 is 294W and 5870 is 224W.   The card is reported to be 350W which is significantly higher not lower.
Still this is AMD and even with three decades of experience the 7990 (375W TDP) was delayed by six months due to power/thermal issues that they found challenging to resolve.

While 350W is possible in that form factor one would have to be willing to bet that
a) BFL hasn't been overly optimistic in power simulations (unlikely every other product in the past).  350W is cutting it close, 400W would be nearly impossible
b) BFL doesn't run into any cooling or power problems due to the high energy density.  Something that has plagued even veteran companies like AMD and NVidia.

As for 350W being realistic.  Well it is 0.6 w/GH.  BFL current chips are 3.1 w/GH.  A die shrink generally cuts power consumption by 40% (you stated upthread up to 60% but that would be rather optimistic don't you think). 28nm is two die shrinks from current chip.  So 3.1 w/GH * 0.6 * 0.6 = 1.1 w/GH.  Ouch.  1.1 * 600 = 660W.  Now you did indicate you opimtized the chip but that is a rather significant optimization wouldn't you say.  Nearly an 86% performance per watt.  Intel is happy for a 5% to 10% improvement in performance per watt (outside of die shrinks).

Given the aggressive improvement in performance per watt necessary, combined with the lack of any headroom (if it misses by even 20% then it can't be cooled in that form factor at that speed), it would need to be a nearly flawless design and execution from start to finish.   It certainly "can" be done but given BFL past promises on power and cooling well one would be betting that "this one will be different". 

Regarding 5970 and 5870, it was my mistake looking at some charts (I'm not good with GPUs generally), what I meant was 6990. The actual design was modified, stray capacitance and flip-flop instability was resolved (which were causing the majority of consumption). The numbers we have are lower, and were reported here with margin. The migration between Stratix 3 and Arria 2 GX (Original single vs. MiniRig FPGA cards), proved a 50% reduction in power. As both were doing SHA256, taking 50% for this case would be a reasonable number. Including the corrections made to stray capacitance and flip flop instability, the final figures arrive at an even lower number. The number announced by us is the worst case scenario based on what we have in our hands; however, as stated, these are estimations.


Regards,
Nasser