High-Speed Se ial In e aces o E en -D i en
Neu omo phic Sys ems
M. Jabłoński1, T. Se ano-Go a edona2, and B. Lina es-Ba anco2
1AGH Uni e si y o Science and Technology, Depa men o Au oma ics and Bioenginee ing, K akow, Poland
2Ins i u o de Mic oelec ónica de Se illa, IMSE-CNM (CSIC and Uni . Se illa), SPAIN. be [email protected]
Abs ac —Neu omo phic Enginee ing is he discipline o building
senso y p ocessing a i icial sys ems inspi ed in he neu al
p ocessing ound in li ing beings. Biological neu al b ains show
massi e connec i i y among neu ons, which is no ealis ic o
mimic using wi es wi hin silicon chips o be ween chips. Add ess-
E en -Rep esen a ion is a echnology widely used among
neu omo phic enginee s o emula e such massi e in e connec i i y
by ime-mul iplexing as digi al channels by ansmi ing
“Add ess E en s” be ween neu ons ha mimic he neu al spikes
ansmi ed in biology. He e we show on-going p og ess on bi -
se ial SATA AER in e -FPGA communica ions o mul i- ile
scalable neu omo phic sys ems.
I. INTRODUCTION
Add ess E en Rep esen a ion (AER) is a communica ion
echnique p oposed in he ea ly 90s [1], widely adop ed by
neu omo phic enginee s. O iginally, he AER echnique would
mul iplex in ime he spike ac i i y o silicon neu ons (in he
ange o a ew Hz) in o a high-speed asynch onous hand-shaked
mul i-bi pa allel digi al bus, ansmi ing each spike in a ew
nano seconds. Since hen, AER based sys ems ha e g own in
complexi y and esou ces, equi ing highe and highe
communica ion bandwid hs and logic esou ces.
Wi h he g ow h o neu omo phic sys ems, p ocessing and
communica ion in as uc u es pe o mance equi emen s ha e
inc eased. In AER-based sys ems he connec i i y is one o he
main pe o mance bo lenecks. In e aces wi h pa allel digi al
buses ha e limi ed bandwid h due o bus equency limi a ions,
in e -bi ji e and skew, as well as bus leng h design es ic ions.
On he o he hand, bi -se ial in e aces allow o se e al Giga
bi s pe second speeds as well as adjus able e en bi leng hs.
The wo-wi e 4-phase handshake p o ocol, highly ypical in
pa allel AER links, can be subs i u ed by a low con ol
mechanism when he e a e LVDS links in bo h di ec ions [2].
In his pape we show on-going p og ess on AER bi -se ial
links be ween FPGAs, es ed a 1.5Gbps, capable o ansmi ing
32-bi Add ess E en s a a a e o up o 37.43Meps (mega e en s
pe second). These links exploi pa allel- o-se ial anscei e s
a ailable in Spa an-6 FPGAs. E en s a e gene a ed by a 32-bi
bi -pa allel Tes Pa e n Gene a o (TPG) unning a a highe
unco ela ed clock, which in e aces o he anscei e h ough a
clock-domain-c ossing in e ace and high-speed synch onize s.
The link has been es ed on a cus om PCB ha includes one
Spa an-6 150 (wi h se ial anscei e s) FPGA and 4
bidi ec ional SATA connec o s, which we call he “AER-Node”
PCB, and is in ended o g id assembly o many o hem.
II. AER-NODE PCB
The AER-Nodeboa d was designed wi h he aim o
demons a e ha spike-p ocessing unde AER is easible and
con enien o high speed ame- ee ision, il e ing, p ocessing
and ac ua ion, using spikes om ision senso s o DC mo o s.
Scalabili y is p o ided by ou SATA connec o s o
bidi ec ional LVDS high-speed communica ions o enable a
mesh o AER-Nodeboa ds [9]. Func ionali y can be inc eased
wi h p ope daugh e boa ds, h ough wo pa allel 28-bi
connec o s and wo 8-bi da a connec o s.
(a)
(b)
Fig.1:(a) AER-Nodeboa d wi h 4 bidi ec ional SATA and 2 pa allel AER
connec o s. (b) Tes assembly whe e all 4 bidi ec ional SATA links a e
in e changing unco ela ed e en s in bo h di ec ions.
Fo each bidi ec ional SATA link, Xilinx p o ides a
pa allel/se ial anscei e w appe (called he e “GTP ile”),
syn hesized by he Co e Gene a o u ili y.
Fig. 2: “GTP ile” anscei e w appe p o ided by Xilinx, showing
inpu and ou pu signals and ela ed clocks.
Fig. 2 shows one such “GTP ile” oge he wi h i s inpu and
ou pu signals and ela ed clocks. All GTP iles wi hin he FPGA
use an ex e nal e y low ji e e e ence clock wi h di e en ial
clock lines. These di e en ial clock lines eed in o special ab ic
wi hin he FPGA, and once inside, a e ans o med om
di e en ial o single ended be o e eeding he GTP ile. In ou
AER-Node boa ds we ha e used ex e nal clocks o ei he
100MHz o 150MHz. The 100MHz e e ence clock can be used
o LVDS da a a es o line = {1.0, 1.25, 2.0, 2.5}Gbps, while he
150Mhz one would p o ide line = {0.75, 1.5, 3.0}Gbps. O he
e e ence equencies a e also possible. In he p esen wo k we
used an AER-Node boa d wi h an ex e nal clock o 150MHz.
The GTP ile includes a ansmi e side (TX) which eads 8n
bi s o pa allel da a (DATA_ou ) and se ializes i ou o he
LVDS_ou di e en ial line. When using 8b/10b encoding [11]
an addi ional n-bi lag ‘cha isk’ signals whe he he
co esponding 8-bi wo d should be ea ed as a egula da a o a
comma command. The GTP ile also includes a ecei e side
(RX) which pe o ms he opposi e. The GTP ile can be
syn hesized o allow o di e en sizes o he pa allel
DATA_ou /in wo d. The pa allel da a can ha e mul iple o 8
bi s, whe e hese mul iples can be ei he n = 1, 2, 4, o 8, hus
allowing ‘DATA_ou /in’ o sizes 8, 16, 32, o 64 bi s. In his
wo k we used always n = 4, which co esponds o 32-bi pa allel
da a wo ds.
The GTP ile in e nal PLLs ans o m he ex e nal e e ence
clock in o he high speed line equency o LVDS bi -se ial
ansmission. I also p o ides wo e e ence clocks o use
ci cui y o equency and n . When using 8b/10b encoding line
= 10n , and he 8n bi s o ‘DATA_ou /in’ and n bi s o ‘cha isk’
a e ead a he ising edges o clock . Fo example, i using a
150MHz ex e nal e e ence clock and se ing line = 1.5Gbps wi h
n = 4, hen = 37.5MHz and 4 = 150MHz. Consequen ly, his
would allow o a heo e ical maximum e en h oughpu o
37.5Meps (mega e en s pe second), being each e en o 32-bi s.
As we will see in Sec ion IV, he maximum e en h oughpu is
sligh ly less, which is due o he need o pe iodically ansmi
commas o e en alignmen , clock co ec ion and low con ol.
III. BIT-SERIAL BI-DIRECTIONAL AER LINK
AER links be ween chips ha e adi ionally been pu ely
asynch onous [1], using bi -pa allel po s oge he wi h wo
lines o 4-phase handshaking. A se ial-wo d app oach was also
p oposed [3], which would educe he numbe o pa allel bi s by
hal . In he p esen app oach we a e using a clocked “GTP ile”
ha in e aces on one side h ough LVDS o ano he FPGA
“GTP ile”, and on he o he side o in-FPGA clocked use
ci cui y. This clocked use ci cui y may un a he same
equency he GTP ile is eading (TX) DATA_ou o
p o iding (RX) DATA_in, o may un a a di e en equency
(ei he co ela ed o and 4 , o o ally unco ela ed).
Ul ima ely, he FPGA use ci cui y would in e ace o an
ex e nal pa allel AER de ice (such like an AER e ina senso
[4], [5]) which is ully asynch onous.
He e we wan o explo e he case whe e he “GTP ile” has
o connec , inside he FPGA, o a synch onous pa allel AER
ci cui clocked a he maximum possible equency. Fo his, we
syn hesized a “Tes Pa e n Gene a o ” (TPG) which we we e
able o clock a 375MHz. This clock was gene a ed om a
second X al on he PCB, which is unco ela ed o he low-ji e
di e en ial one shown in Fig. 2. Fig. 3(a) illus a es he es
app oach by showing a simpli ied diag am o he ci cui y
connec ing o he TX pa o he “GTP ile” in Fig. 2. The e is a
high speed Tes Pa e n Gene a o clocked a TPG = 375MHz,
which p o ides a bi -pa allel DATA o 32-bi s h ough
handshaking (Req and Ack). These handshaking signals go
h ough high-speed synch onize s, each clocked by ei he TPG
o 4 , as shown in he igu e. The TX FSM clocked a 4
in e aces wi h he “GTP ile” on one side, and on he o he side
in e aces h ough a clock domain c ossing ci cui y wi h he
high speed TPG. Since “GTP ile” eads and w i es da a
acco ding o he ising edges o , he TX FSM will p o ide a
signal “phase” o he ci cui s unning a equency 4 o signal
when he ising edge o is app oaching. The TX FSM and
clock domain c ossing ci cui y a e bo h clocked by he same
clock 4 , and in e change in o ma ion abou hei in e nal s a us
h ough signals “ alid” and “busy”. The “clock domain
c ossing” ci cui in his case is a he simple and is gi en Fig.
3(b).
Fig. 3: (a) In e acing ci cui y be ween high speed clocked Tes Pa e n
Gene a o and GTP ile. (b) De ails o clock domain c ossing ci cui .
IV. EXPERIMENTAL RESULTS
In his Sec ion we p o ide ini ial expe imen al
cha ac e iza ion esul s o he case o line = 1.5Gbps, 4 =
150MHz, and = 37.5MHz.
A. In eg i y o LVDS Signals in FPGA BUILT-IN
GTPs
The quali y o bi -se ial ansmission using GTP in e aces
embedded in FPGAs depends on se e al ac o s. The p ima y
sou ce o e o s is he equency misma ch o he clocks
in ol ed and hei ji e . F equency misma ch (usually wi hin a
ew ens ppms) can be success ully esol ed by clock co ec ion
echniques based on elas ic bu e s and a ailable wi hin he
GTP iles [7], [8]. Physical p ope ies o he link, such as
quali y o PCB aces, wi es and connec o s also impac signal
in eg i y. A basic me ic ypically used o he e alua ion o bi -
se ial in e aces is BER (Bi E o Ra e), which indica es he
numbe o e oneous bi s ha ha e been ansmi ed du ing a
es pe iod - he lowe he alue he be e he link. Fo his
pu pose Xilinx p o ides dedica ed Tes Pa e n Gene a o s based
on 7-bi PRBS (Pseudo Random Bi S eam), which can be
gene a ed and e i ied inside a GTP anscei e i connec ed in
loop-back mode.
A me hod equen ly used o di ec ly measu e signal
in eg i y is he Eye-Diag am analysis, which gi es mo e
de ailed analog in o ma ion on ime and ol age ma gins o a
pa icula link and i s se up (p e-emphasis, equaliza ion). I
allows also de e mining he op imum sampling poin o a oid
iola ion o ime- ol age ma gins embedded in o an Eye-
Diag am. This me hod can howe e ha e he d awback ha he
shape o he diag am can be a ec ed by a aching he sampling
p obe, as his is an “in asi e” me hod. Fo example, Fig. 4
shows measu ed eye-diag ams a wo di e en sampling poin s.
(a)
(b)
Fig. 4: Eye-Diag ams o LVDS link measu ed a di e en posi ions: (a) a
FPGA RX inpu pins, (b) a in e -PCB SATA connec o a TX side.
An al e na i e non-in asi e echnique o s udying physical
channel p ope ies is he Ba h-Tube analysis. A Ba h-Tube
cu e can be acqui ed by measu ing he BER alue o a ixed
pe iod o ime and sweeping a ange o Sampling Poin alues.
The esul is a ba h-shaped cu e ha usually yields e y low
alues o BER (i.e. o 0.0) in he middle o he diag am, i.e.
a ound 0.5UI (Uni Window). This co esponds o hal o he
pe iod o he clock cycle, eco e ed om he ecei ed da a bi -
s eam. T h e cen e o he egion whe e he BER cu e eaches
0.0 is he op imum Sampling Poin o he se up o a pa icula
loop-back link (see Fig. 5). Plo s we e collec ed wi h he
ChipScope ha dwa e debug in e ace and so wa e plug-in
iBERT [8] o Spa an6 de ices.
Besides he Sampling Poin , he ollowing pa ame e s also
signi ican ly impac ansmission quali y: Vol age Swing, TX-
P eemphasis, and RX-Equaliza ion. They should be adjus ed o
gi e minimum BER a a pa icula speed o he link. Fig. 5
shows some sample Ba h-Tube diag ams o a ious RX-
Equaliza ion pa ame e s. Fig. 6 shows he o e lap o he ba h
g aph wi h he eye diag am.
Fig. 5: Ba h-Tube diag am o loop-back connec ion a 1.5Gbps, TX-
P eemphasis 2.5dB, Vol age Swing 929mV
Fig. 6: Ba h-Tube cha embedded in o Uni -Window o he Eye-Diag am,
o 1.5Gbps, TX-P eemphasis 2.5dB, ol age swing 929mV.
B. E en E o Ra e (EER) Cha ac e iza ion
Fo ou bi -se ial in e -FPGA link i is in e es ing o
cha ac e ize he “E en E o Ra e” (EER), as his would
include all possible e o sou ces, no only physical, bu also
logical, as he p ope ope a ion o he low con ol p o ocol, he
clock co ec ion scheme, o he mul i-by e e en alignmen [2].
In o de o es EER we used he TPG discussed in Sec ion II
clocked a a e y high clock equency o TGP = 375MHz. The
es se up in ol ing wo AER-Node PCBs (NB1 and NB2) is
shown in Fig. 7. The pa allel-se ial anscei e link in each
FPGA includes he “GTP ile”, he “TX block” (clocked a 4 ,
he as e use clock p o ided by “GTP ile”) and an “RX
block” (clocked a , he slowe use clock p o ided by “GTP
ile”). Besides he TPG, he e a e also wo “Synch onous Tes
Pa e n Checke s” (STPC) clocked a equency . The STPC a
he end o he chain in each FPGA es s EER o he whole
ansmission chain, while he STPC be ween “GTP ile” and
“TX block” es s clock domain c ossing issues.
Fig. 7: Tes se up consis ing o wo AER-Node PCBs (NB1 and NB2), each
wi h one bi -se ial link (“GTP ile”, “TX block” and “RX block”) oge he
wi h a “Tes Pa e n Gene a o ” (TPG) clocked a e y high equency TPG
= 375MHz and wo “Synch onous Tes Pa e n Checke s” (STPC) clocked
a he slowe clock p o ided by “GTP ile” = 37.5MHz.
Table I: Measu ed E o s, 32-bi E en Th oughpu and skew be ween
pai s o AER-Node Boa ds on all 4 links
In e nal
STPC
e o s
EER RX Measu ed
Th oughpu
(eps)
Clk
skew
(ppm)
NB2 → NB1 0 0.0 37426439 -16.134
NB1 → NB2 0 0.0 37427647 16.133
NB4 → NB3 0 0.0 37428143 -29.409
NB3 → NB4 0 0.0 37425942 29.408
Long e m es s o se ial connec ions a 1.5Gbps we e
pe o med concu en ly be ween wo AER-Node PCBs o each
all 4 se ial links, as connec ed in Fig. 2(b), du ing 20 hou s o
o e . No e o s we e de ec ed a none o he STPCs, as shown
in Table I: columns “In e nal STPC e o s” and “EER” (which
co esponds o he e o s coun ed a he STPC a he end o he
pa hs). The h oughpu measu ed a he in e nal STPCs was
always 37,427,043 eps, which co esponds exac ly o he
expec ed TX side gene a ed e ec i e da a a e (excluding he
COMMA densi y we we e using). Howe e , he measu ed
h oughpu a ies sligh ly on he pa h ends STPCs (RX side),
which a e he ones shown in Table I unde “RX Measu ed
Th oughpu ”, because hese a e numbe s a e measu ed using
he clocks a he des ina ion AER-Node boa d whose clock
di e s by a ew ens o ppms. The measu ed di e ence be ween
he TX and RX h oughpu s is gi en unde column “Clk skew”
in ppms.
CONCLUSIONS
We ha e shown co ec ope a ion o bi -se ial AER links
be ween cus om made PCBs o mul i- ile PCB assembly
in e connec ed h ough 4 SATA links. Ini ial expe imen s
ope a ing a 1.5Gbps a e epo ed. The links a e s imula ed
h ough asynch onous handshaking echniques inside each
FPGA by Tes Pa e n Gene a o s clocked by unco ela ed
clocks. These unco ela ed clocks we e se a he maximum
possible equency we we e able o se , 375MHz, in o de o es
he channel a i s maximum possible h oughpu . Exhaus i e
es s we e pe o med o e i y he co ec ope a ion, including
eye diag am measu emen s, ba h- ube measu emen s, bi e o
a es, and e en e o a es.
ACKNOWLEDGMENTS
This wo k has been suppo ed by Spanish g an s (wi h suppo om he
Eu opean Regional De elopmen Fund) BIOSENSE (TEC2012-37868-C04-
02/01), Andalusian g an NANO-NEURO (TIC-6091), Polish g an om he
Minis y o Science and Highe Educa ion o he Republic o Poland AGH UST
no. 11.11.120.612, EU CHIST-ERA g an PNEUMA (PRI-PIMCHI-2011-0768),
EU FP7 g an HBP (The Human B ain P ojec ) 604102, and H2020 EU g an
ECOMODE 644096.
REFERENCES
[1] M. Si ilo i, “Wi ing conside a ions in analog VLSI sys ems wi h
applica ion o ield-p og ammable ne wo ks,” Ph.D. Disse a ion,
Compu a ion and Neu al Sys ems, Cali o nia Ins . Technol., Pasadena,
CA, 1991.
[2] T. Iakymchuk, A. Rosado, T. Se ano-Go a edona, B. Lina es-Ba anco,
A. Jimenez-Fe nandez, A. Lina es-Ba anco, and G. Jimenez-Mo eno,
“An AER handshake-less modula in as uc u e PCB wi h 8 2.5Gbps
LVDS se ial links.” P oc. o he IEEE In . Symp. Ci c. and Sys . (ISCAS),
pp. 1556–1559, June 2014.
[3] K. Boahen, “Poin - o-Poin Connec i i y Be ween Neu omo phic Chips
Using Add ess E en s,” IEEE T ans. on Ci c. and Sys . Pa -II, ol. 47,
no. 5, pp. 416-434, May 2000.
[4] P. Lich s eine , C. Posch, and T. Delb uck, “A 128x128 120dB 15us
la ency asynch onous empo al con as ision senso ,” IEEE J. Solid S a e
Ci cui s, 43(2) 566-576, 2007.
[5] T. Se ano-Go a edona, B. Lina es-Ba anco, “A 128x128 1.5% Con as
Sensi i i y 0.9% FPN 3µs La ency 4mW Asynch onous F ame-F ee
Dynamic Vision Senso Using T ansimpedance P eampli ie s,” IEEE
Jou nal o Solid-S a e Ci cui s, ol. 48, No. 3, pp. 827 – 838, 2013.
[6] R. Se ano-Go a edona, e al. “CAVIAR: A 45k-Neu on, 5M-Synapse,
12G-connec s/sec AER Ha dwa e Senso y-P ocessing-Lea ning-Ac ua ing
Sys em o High Speed Visual Objec Recogni ion and T acking”, IEEE
T ans. on Neu al Ne wo ks, ol. 20, No. 9, pp. 1417-1438, Sep. 2009.
[7] DS160. Spa an-6 Family O e iew, .2.0. Xilinx, Oc obe 2011.
[8] UG386 Spa an-6 FPGA GTP T anscie e s. Ad ance P oduc
Speci ica ion, .2.2. Xilinx, Ap il 2010.
[9] C. Zama eño-Ramos, A. Lina es-Ba anco, T. Se ano-Go a edona, B.
Lina es-Ba anco, “Mul icas ing Mesh AER: A Scalable Assembly
App oach o Recon igu able Neu omo phicS uc u ed AER Sys ems.
Applica ion o Con Ne s”, IEEE T ans. on Biomedical Ci c. and Sys .,
ol. 7, No. 1, pp 82-102. Feb. 2013.
[10] Daniel B. Fasnach , Ad ian M. Wha ley, Giacomo Indi e i. “A Se ial
Communica ion In as uc u e o Mul i-Chip Add ess E en Sys ems”.
Ci cui s and Sys ems (ISCAS), 2008 IEEE In e na ional Symposium on.
[11] P. A. F anaszek, e al., “By e o ien ed DC balanced (0,4) 8b/10b
pa inioned block ansmission code,” US Pa en 4,486,739, Dec. 4, 1984.