<mithro>
ZirconiumX: You do seem better at stats than myself, so would appreciate feedback too
<ZirconiumX>
It's less about that, more about applying the little bit of existing knowledge I have on these things :P
<sf-slack>
<acomodi> mithro: as far as I understand, we would need to add tests in tool-perf to iteratively check an algorithm (in this case the a PnR tool) to verify the assumptions
<sf-slack>
<acomodi> for assumption 1: run a design X times and verify that the X results are the same (or at least are between an interval of acceptance)
<sf-slack>
<acomodi> for assumption 2: run a design X times and for each run change the seed and verify that the output changes as well
<sf-slack>
<acomodi> I assume that the output in this case would be CPD, correct?
<ZirconiumX>
mithro: it's "MARLANN" as in "multiply-accumulate and rectified-linear accelerator for neural network"
borisnotes has quit [Ping timeout: 260 seconds]
gsmecher has joined #symbiflow
<mithro>
acomodi: So for assumption 1 -- we need to run the tool X times with identical input and check we get identical output
<mithro>
acomodi: For assumption 2 -- I think we need to run a whole bunch of runs until we have a high level of confidence
<mithro>
Test 3 is the big one
<ZirconiumX>
mithro: Can I suggest an alternative alternative test 3?
<mithro>
ZirconiumX: Sure
<mithro>
ZirconiumX: the discussion section is trying to explain my high level thinking around why test 3 is a reasonable thing to think about
<ZirconiumX>
Assumption 3: Place-and-route results have approximately normally distributed maximum frequencies.
<ZirconiumX>
(this holds true in my experience)
<ZirconiumX>
Alt. Alt. Test 3: Determine the mean of the normal distribution of maximum frequencies. This will tell you how likely it is that a tool will produce a design that satisfies the timing constraints.
<ZirconiumX>
My assumption here is that the seed is effectively a random input, and related seeds do not produce related results.
<ZirconiumX>
Which seems a much safer assumption than the thinking you outline
<ZirconiumX>
mithro: ^
rw1nkler has quit [Quit: WeeChat 1.9.1]
wavedrom has joined #symbiflow
borisnotes has joined #symbiflow
<mithro>
ZirconiumX: I believe you that there is a distribution of maximum frequencies - I'm not sure that I believe it is normally distributed
<mithro>
I wouldn't be surprised to see binomial type distributions
<ZirconiumX>
I think Fmax should be treated as continuous
<ZirconiumX>
Because there's no inherent yes/no here
<ZirconiumX>
So binomial is definitely the wrong one here
citypw has quit [Ping timeout: 240 seconds]
<mithro>
ZirconiumX: I've definitely seen designs where the optimizer gets stuck with fmax distributed around two seperate fmax values
<ZirconiumX>
That would be bimodal, not binomial :P
<hackerfoo>
Unrelated: I'm seeing about 800GB memory usage for 80 runs of VPR for the Titan tests, so plan on 10G per run for tests, CI, etc.
<hackerfoo>
I had to switch to m1-megamem-96 to make effective use of the cores.
<litghost>
hackerfoo: Most of that memory is just the gaussian blur test
<litghost>
hackerfoo: Are you running with titan nightly or weekly?
<hackerfoo>
It's not e.g. LU230 is at 17GB right now.
<hackerfoo>
I'm running nightly, but with various flag settings.
<hackerfoo>
There isn't just one offender, there are many designs that are well over 10GB.
<hackerfoo>
The ones that use the most memory also tend to run longer.
tcal_ has quit [Ping timeout: 265 seconds]
<hackerfoo>
I think using megamem (there's only one) is a good idea, because ultramem is too much, and with highmem you can only safely use about half the cores.
<hackerfoo>
I'm still only running 80 threads per machine, but I'm going to do 95 (leaving one for other tasks) next time.
<hackerfoo>
For 950 threads total.
syed has joined #symbiflow
syed has quit [Quit: Leaving...]
OmniMancer has quit [Quit: Leaving.]
adjtm_ has joined #symbiflow
borisnotes has quit [Read error: Connection reset by peer]
borisnotes has joined #symbiflow
QDX45 has quit [Remote host closed the connection]
QDX45 has joined #symbiflow
adjtm has quit [Ping timeout: 256 seconds]
LisaMarie has joined #symbiflow
LisaMarie has quit [Client Quit]
adjtm_ has quit [Remote host closed the connection]
adjtm has joined #symbiflow
tcal_ has joined #symbiflow
tcal_ has quit [Remote host closed the connection]