Upg ading SPHERE wi h he second s age AO sys em
SAXO+: Ha dwa e in e ace in eg a ion
P. Fon anillasa, F. Fe ei aa, A. Se ina, D. G a adou a, G. Chau ing, E. Diolai ib, R. G a on ,
M. Loupiasc, J. Millid, F. Wildie, A. Boccale ia, and P. Ve mo a
aLIRA, Obse a oi e de Pa is, Uni e si ´e PSL, CNRS, So bonne Uni e si ´e, Uni e si ´e Pa is
Ci ´e, 5 place Jules Janssen, 92195 Meudon, F ance
bINAF - Osse a o io di As o isica e Scienza dello Spazio (OAS), 40129 Bologna, I aly
cUni . Lyon, Uni . Lyon1, ENS de Lyon, CNRS, Cen e de eche che As ophysique de Lyon
UMR5574, F-69230, Sain -Genis-La al, F ance
dUni e si ´e Joseph Fou ie -G enoble 1 /CNRS-INSU, Ins i u de Plan´e ologie e
d’As ophysique de G enoble (IPAG), Sain Ma in d’He es, F ance
eObse a oi e de Gen`e e, CH-1290 Sau e ny, Swi ze land
INAF, Osse a o io As onomico di Pado a (OAPd), 35122 Pado a, I aly
gLabo a oi e Lag ange, Uni e si ´e Co e d’Azu , CNRS, Obse a oi e de la Co e d’Azu , 06304
Nice, F ance
ABSTRACT
SAXO+, he upg ade o he SPHERE adap i e op ics (AO) sys em, is abou o en e i s Assembly, In eg a ion,
and Tes ing (AIT) phase ollowing he Final Design Re iew (FDR) a ESO in May 2025, wi h i s ligh scheduled
o 2028 on he VLT. The SAXO+ p ojec aims o enhance he pe o mance o SPHERE, pa icula ly in e ms
o con as , by implemen ing a second-s age adap i e op ics sys em ha will ope a e in conjunc ion wi h he
exis ing i s -s age sys em (SAXO).
The Real-Time Con olle (RTC) is a key componen o he SAXO+ sys em. I is based on he COSMIC
pla o m o he ha d eal- ime laye and on he ESO RTC Toolki o he so eal- ime laye . The SAXO+
RTC has been designed o be lexible and modula , suppo ing he implemen a ion o a ious con ol s a egies,
including echniques based on a i icial in elligence (AI). This second s age will ea u e a py amid wa e on
senso (PWFS) ope a ing a high ame a es—up o 3 kHz—, a new de o mable mi o (DM) wi h 952 ac ua o s
and a sFPDP link o ensu e coo dina ion wi h he i s s age. The in eg a ion o hese in e aces emains
a signi ican challenge, equi ing ca e ul coo dina ion be ween ha dwa e iming, da a o ma s and eal- ime
p ocessing cons ain s o gua an ee o e all sys em s abili y and pe o mance.
In his pape , we p esen he SAXO+ RTC so wa e de elopmen cu en s a us and he ha dwa e in e ace
in eg a ion. We also p esen p elimina y pe o mance esul s o he RTC, including la ency and ji e measu e-
men s o sFPDP and came a acquisi ion. Finally, we discuss he nex s eps in he de elopmen o he SAXO+
RTC, ocusing on sys em in eg a ion and es ing du ing he p ojec ’s AIT phase.
Keywo ds: Adap i e op ics, SPHERE+, SAXO+, VLT, RTC, COSMIC, sFPDP, Came a Link
Fu he au ho s in o ma ion: (Send co espondence o P. Fon anillas o F. Fe ei a)
P. Fon anillas: paul. on anillas a obspm.
F. Fe ei a: lo ian. e ei a a obspm. , +33 (0)1 45 07 71 78
1. INTRODUCTION
Adap i e op ics (AO) sys ems a e a co ne s one echnology o high-con as imaging on g ound-based ele-
scopes. SPHERE1has deli e ed ema kable esul s in he di ec imaging and cha ac e isa ion o exoplane s and
ci cums ella disks. SAXO+2is concei ed as a second-s age AO upg ade o SPHERE ha will ex end and
ampli y hese capabili ies: by adding a high-o de , as - esponse co ec i e s age ope a ing downs eam o he
exis ing SAXO sys em, SAXO+ aims o imp o e he ins umen sensi i i y a smalle angula sepa a ions.
Figu e 1: Schema ic ep esen a ion o he SAXO+ con ol sys em
As showns ig. 1 he SAXO+ second s age complemen s he i s -s age co ec ion pe o med by SAXO. I
in eg a es a 34 by 34 ac ua o s Bos on Mic omachine de o mable mi o (DM) and a C-RED One came a unning
a 3kHz. The second-s age a chi ec u e a ge s esidual abe a ions le by SAXO, enabling deepe con as in
he inal science image.
The Real-Time Con olle (RTC) is cen al o SAXO+’s ope a ion. The SAXO+ RTC will con ol only he
second s age, while he S anda d Pla o m o Adap i e Op ics Real- ime Applica ion (SPARTA3) will con inue
o be in cha ge o he i s s age. SPARTA and he SAXO+ RTC will be synch onized ia a sFPDP link,
unning a 1kHz. The RTC is implemen ed as a wo-laye sys em: a ha d eal- ime compu e (H-RTC) based
on he COSMIC4pla o m o gua an ee de e minis ic, low-la ency p ocessing, and a so eal- ime compu e
(S-RTC) buil on he ESO RTC Toolki o p o ide moni o ing, supe iso y unc ions and lexible expe imen
o ches a ion. This design aims o gi e SAXO+ he low-ji e pe o mance equi ed o kHz-class ope a ions
while emaining modula enough o hos mul iple con ol s a egies.
To ully exploi he capabili ies o he ha dwa e, SAXO+ will suppo a sui e o six con ol laws ha span
classical and mode n app oaches, wi h he abili y o compa e hem in ope a ion. The six laws a e:
In eg a o wi h Op imized Modal Gain In eg a o (OMGI)
In e se p oblem app oaches5
LQG con olle 6
Da a-d i en con olle 7
PO4AO con olle 8
Dual-S age Supe ised and Rein o cemen Lea ning con olle 9
This pape p o ides an upda e on he RTC so wa e de elopmen s a us and ocuses on he ha dwa e in e -
ace in eg a ion o he SAXO+ RTC. Fi s ly, in Sec.2we desc ibe he so wa e de elopmen cu en si ua ion.
Secondly, we p esen he sFPDP and came a in e aces in eg a ion and hei pe o mance esul s in Sec.3and
Sec.4 espec i ely. Finally, in Sec.5we will in oduce a GPU pipeline es alida ing he en i e RTC loop.
2. SOFTWARE DEVELOPMENT CURRENT STATUS
This sec ion aims o p o ide an o e wiew o he cu en RTC a chi ec u e.
2.1 O e iew
On one side, he SAXO+ Ha d Real-Time Con olle (H-RTC) handles c i ical eal- ime asks wi h s ingen
iming equi emen s. Since some o he SAXO+ con ol me hods ely on a i icial in elligence, which is highly
demanding in e ms o bandwid h, he H-RTC has been designed o use GPU as i s compu ing uni . The H-RTC
consis s o wo NVIDIA H100 GPUs, each p o iding 3 TB/s memo y bandwid h while a ailable CPUs only o e
400 GB/s. SAXO+ H-RTC is buil upon he COSMIC pla e o m, ha gua an ees de e minis ic pe o mance
and low-la ency p ocessing.
On he o he side, he SAXO+ So Real-Time Con olle (S-RTC) manages non-c i ical eal- ime asks,
including sys em moni o ing, supe iso y con ol, and non-c i ical compu a ion. The S-RTC is spli in o 4
componen s:
he H-RTC Ga eway: he main in e ace be ween he H-RTC and he S-RTC, esponsible o da a exchange
and command elay
he Compu e Node: esponsible o execu ing non-c i ical Da a asks and o loading wo k om he H-RTC
he S-RTC Ga eway: he main in e ace be ween he S-RTC and ex e nal sys ems, esponsible o da a ex-
change and command elay
he S o age Node: esponsible o da a s o age and e ie al, ensu ing ha all ele an da a is accessible o
he S-RTC
In o de o educe he o e all cos and inc ease main ainabili y, he S o age Node and he S-RTC Ga eway will
be hos ed on he same physical machine.
Figu e 2: RTC communica ion
Figu e 2highligh s in e ac ions be ween H-RTC and S-RTC componen s. S-RTC se e s communica e
h ough Da a Dis ibu ion Se ice (DDS) opics, using DDS publishe s and subsc ibe s. SPHERE INS can
in e ac wi h he RTC sys em h ough he S-RTC Ga eway.
The S-RTC is buil upon he ESO RTC Toolki , a obus amewo k designed o he de elopmen o eal- ime
con olle s. I in e aces wi h he H-RTC o ensu e seamless ope a ions, o e ie e and s o e eleme y. I is
composed o se e al ypes o componen s:
RTC Supe iso : p o ides an en y poin o high-le el con ol, moni o ing o he RTC sys em, e o eco e y
and is used o access he Run ime Con igu a ion Reposi o y (RTR). The e is only one RTC supe iso in
SAXO+ S-RTC.
Teleme y Republishe : eads eleme y da a in MUDPI o ma om he HRTC, and epublishes i o one
o mo e S-RTC nodes, using DDS eliable Mul icas . The e a e h ee Teleme y Republishe componen s: one
handling aw da a, ano he one ea ing calib a ed da a and a las one managing loop da a, which will be
speci ic o each con ol law.
Teleme y Subsc ibe : ecei es da a om one o mo e Teleme y Republishe s and w i e hem in o a sha ed
memo y, o hen be consummed by da a asks. As o epublishe s, he e a e h ee subsc ibe s: one o aw
da a, one o calib a ed da a and one o loop da a.
Teleme y Reco de : eads eleme y da a om speci ic sha ed memo y queues and s o e i in o FITS
iles. Reco de s can also ead da a om o he sou ces and s o e hem in o FITS o CSV iles. The e a e wo
eco de s in SAXO+ S-RTC: one o s o e pixels eleme y da a and ano he one o s o e loop eleme y da a.
Da a asks: esponsible o a ious non-c i ical compu a ions and da a p ocessing asks ha suppo he
o e all ope a ion o he RTC sys em. They ecei e eleme y and RTR da a. Da a asks esul s a e s o ed in
he RTR a e p ocessing. The e a e en Da a asks o he in eg a o wi h OMGI con ol law and mo e o
come o he o he con ol laws.
H-RTC and S-RTC a chi ec u es and beha iou ha e al eady been desc ibed in de ail in Re .10 We will ocus
he e on he cu en de elopmen s a us.
2.2 Cu en s a us
The de elopmen o SAXO+ RTC is well unde way. The h ee S-RTC se e s a e ope a ional and in use. Each
o hem is composed o wo AMD EPYC 9354 p ocesso s each con aining 32 physical hea s. Fo H-RTC, a call
o ende s is scheduled o ea ly 2026. A H-RTC p o o ype composed o wo NVIDIA A100 GPUs is cu en ly in
use. The ope a ing mode in ol ing he in eg a o wi h OMGI con ol law and associa ed asks has been de ined
as he baseline; i is he implemen a ion p io i y. The o he con olle s and hei associa ed Da a asks will be
de eloped a a la e s age.
Recen de elopmen s ha e been made on he S-RTC side. Mos o he baseline Da a asks ha e been imple-
men ed, including:
In e ac ion ma ix Calib a ion: calib a es he in e ac ion ma ix, linking each adap i e mi o ac ua o
o he Py -WFS esponse.
Command ma ix compu a ion: gene a es he command ma ix om he in e ac ion ma ix.
Re e ence slopes compu a ion: calib a es he Py -WFS e e ence slopes using illumina ed and da k images.
Modal gain compu a ion: compu es modal gains based on measu ed u bulence pa ame e s.
De ec o Calib a ion: measu es de ec o backg ound and la , co ec s dead pixels, and calib a es he alid
pixel map.
ROI Calib a ion: de ines and calib a es he Region O In e es (ROI) o he came a senso (128x128 pixels
squa e).
Pupils Posi ion: moni o s he s abili y o he 4 pupils in he came a image.
Tip-Til O loade : ex ac s om he de o mable mi o commands he a e aged ip, il and de ocus.
These da a asks, as he o he S-RTC componen s a e designed independen ly, o ensu e modula i y and ease o
main enance.
As shown in ig. 3, componen s communica e h ough sha ed memo y opics and a un ime eposi o y. In he
S-RTC a chi ec u e, he da a asks li ecycle and da a- low o ches a ion a e implemen ed in C++. The algo i hm
ke nels a e de eloped in Py hon and in oked om he C++ side. This sepa a ion combines he pe o mance and
p edic abili y o C++ o ho pa hs and sys em in eg a ion wi h he lexibili y and apid p o o yping a o ded by
Py hon o algo i hm de elopmen . The e o e, each da a ask includes wo kinds o uni es s:
One Py hon es based on Py es (py es ): checks he Py hon algo i hm co ec ness, he nume ical accu acy
o he ou pu , he handling o edge cases and excep ions.
One C++ es based on Google Tes and Google Mock (g es ): checks he da a ask compu a ion beha iou , i
e i ies ha he Py hon algo i hm is co ec ly called ia a Py hon w appe and hen i checks i he esul s
a e well sa ed. The g es also es i he wa ning and e o messages a e well igge ed.
Figu e 3: Da a ask es ing a chi ec u e
This dual es ing s a egy ensu es ha bo h he algo i hmic in eg i y and he componen compu a ion be-
ha iou a e igo ously alida ed. A unc ional es pipeline using COMPASS simula o has been implemen ed
on he H-RTC. This es s a egy will be p esen ed in Sec. 6.
In addi ion o he RTC baseline de elopmen , he o he p io i y is o in eg a e he ha dwa e in e aces
enabling communica ion wi h he o he componen s o he AO sys em.
3. sFPDP INTERFACE INTEGRATION
In his sec ion we will desc ibe he sFPDP in e ace in eg a ion in o he H-RTC and i s pe o mance esul s.
3.1 Se up
As he SAXO+ RTC needs o communica e wi h he i s s age, an in e ace needs o be implemen ed. Fo
non-dis up i e in eg a ion, he link be ween he wo s ages will be made using he exis ing Se ial F on Panel
Da a Po (sFPDP) swi ch embedded in SPARTA. The sFPDP p o ocol is a high-speed communica ion s anda d
op imized o eal- ime da a ans e be ween sys ems. I enables low-la ency, de e minis ic ansmission o la ge
da a s eams wi hou he o e head o complex ne wo k p o ocols.
Thus, he RAVEN-I boa d has been selec ed o be his sFPDP in e ace. I is a high-pe o mance da a
acquisi ion and ansmission solu ion de eloped by TECHWAY. Designed o lexible and simple in eg a ion in o
complex a chi ec u es, he RAVEN-I s ands ou o i s abili y o p ocess high-speed analogue signals wi h high
eliabili y. I is designed o in e ace wi h digi al equipmen using he sFPDP p o ocol in demanding eal- ime
applica ions, which i s pe ec ly wi h SAXO+ equi emen s.
The RAVEN-I ca d has been in eg a ed in o he SAXO+ H-RTC p o o ype. The boa d includes 4 sFPDP
channels, each con igu able o ecep ion o ansmission. The maximum da a a e pe channel is 2.5 Gb/s, in
con o mi y wi h VITA 17.1 and 17.3 s anda ds. The ca d uses op ical links ia SFP connec o s. I s hos in e ace
is based on PCIe Gen2 x4 connec o , o e ing a maximum link speed o 6.2 Gb/s.
Se e al es s ha e been conduc ed o ensu e he boa d mee s SAXO+ equi emen s. The i s goal o hese
es s was o alida e ha he boa d was able o g ab he da a coming om SPARTA CODE Tes Tool. The
second goal was o check i he he boa d was capable o acqui e da a a 1KHz, acco ding o SAXO equency.
SPARTA CODE Tes Tool is a so wa e applica ion de eloped by ESO o acili a e he es ing and alida ion
o sys ems ha use SPARTA. Fo his es , he Tes Tool is used as a eal- ime ame gene a o . As shown in
Figu e 4, i sends sFPDP packe s ia a SFP op ical link o he RAVEN-I boa d. The sFPDP packe s con ains
a s uc u ed payload o 2768 by es including:
Figu e 4: sFPDP in eg a ion Se up
1377 16-bi o HODM
2 16-bi alues o ITTM
2 16-bi alues o DTTM
2 16-bi alues o PTTM
The signal sen is a sinusoid. As i s s oke is 1 and he bi ange is 16, he signal is cen e ed symme ically
a ound 32767 wi h alues anging om 16384 o 65534.
The sFPDP acquisi ion applica ion de eloped o his es was w i en in C++ using he Ra en API p o ided
by TECHWAY. The applica ion o e s a comp ehensi e se o ea u es o p ocess da a in eal ime. The sys em
au oma ically eads and analyses sFPDP packe heade s, ex ac ing he sequence coun e and e i ying he
in eg i y o 2768-by e payload. The applica ion e ie es and s uc u es he da a. Each packe ecei ed is
ime-s amped wi h nanosecond p ecision, enabling ine empo al aceabili y o measu emen s. The sys em
au oma ically gene a es JSON iles con aining imes amps co ela ed o sequence coun e s, acili a ing pos -
acquisi ion analysis. The applica ion con igu es Techway boa d channel 0 in ecei e mode, disables he Cyclic
Redundancy Check (CRC) o op imize pe o mance, and hen s a s a eal- ime acquisi ion loop. Then, he
sys em pe o ms a sequence check o de ec los packe s, imes amps each ecep ion wi h nanosecond accu acy,
and can p ocess da a ei he con inuously o o a de ined numbe o packe s.
As SPARTA CODE Tes Tool is no able o p oduce da a a 1kHz, we ha e implemen a ed a new se up o
check whe he he boa d is capable o acqui ing da a a 1kHz.
Figu e 5: sFPDP eal ime acquisi ion se up
As shows Fig. 5, o his new se up he RAVEN-I boa d ope a es in loopback mode, using a ib e op ic cable
physically connec ing channel 0 ( ecei e ) o channel 1 ( ansmi e ). In his con igu a ion, he da a ansmi ed
on channel 1 is immedia ely ecei ed on channel 0, enabling an isola ed e alua ion o he boa d’s eal- ime
acquisi ion pe o mance. Fo his pu pose, we de eloped a cus om da a emula o speci ically o his es , which
espec s he SPARTA sFPDP packe s s uc u e desc ibed abo e. This in e nal emula o gene a ed a con inuous
s eam o amp signals, managing he iming au onomously and p ecisely.
3.2 Resul s
As illus a es Fig. 6, he signal ecei ed by he RAVEN-I boa d wi h he ini al con igu a ion, co esponds o he
inpu . The sinusoid is pe ec ly econs uc ed and cen e ed a ound 32767, wi h alues be ween 16384 o 65534.
We ecei e all he signals we send wi h he es ool, wi hou any loss and in he igh o de . This demons a es
he abili y o he TECHWAY Ra en-I boa d o acqui e da a om SPARTA
Figu e 7shows he imes amp measu ed a each i e a ion, wi h he boa d in loopback mode. We can see
ha he ecep ion equency is 1000Hz, as expec ed, and i is also pe ec ly s able.
Figu e 6: sFPDP ecei ed signal
Figu e 7: sFPDP eal- ime acquisi ion esul s
The a e age ji e is 0.11
µ
s and he maximum measu ed is 0.17
µ
s, which illus a es he iming s abili y o
he acqui ed da a. These es s con i m he boa d’s capabili y o handle eal- ime da a acquisi ion in a s able
manne .
4. CAMERA INTERFACE INTEGRATION
In his sec ion we desc ibe he came a in e ace in eg a ion in o he H-RTC and i s pe o mance esul s.
4.1 Se up
C-RED One is an in a ed came a de eloped by Fi s Ligh Imaging. We chose i o i s high-speed, high-
sensi i i y and ul a-low noise. I can akes mo e han 3500 images pe second a he ull 320x256 esolu ion.
The came a will be linked o he H-RTC ia came a link and a ame g abbe will acqui e he came a ames.
The EDT VisionLink F4 ame g abbe came a link has been selec ed and in eg a ed on a H-RTC p o o ype
PCIe slo . This ame g abbe has been designed o scien i ic applica ions whe e high esolu ion, high speed
and low la ency a e equi ed, making i sui able o SAXO+ needs. The came a acquisi ion es aims o alida e
ha he ame g abbe is able o collec ames coming om he C-RED One came a a 3 kHz.
Figu e 8: Came a in e ace se up
As shown in Figu e 8, he came a is unning a 3 kHz wi h a c opped ame size o 128x128 pixels, which is
he egion o in e es (ROI) de ined o SAXO+. The C-RED One is ope a ing in global ese Co ela ed Double
Sampling mode (CDS). This mode hal es he maximum acquisi ion a e compa ed wi h simple ead mode, bu
allows o compensa e o he di e en exposu e imes be ween pixels caused by olling shu e mode.
A COSMICS ask has been implemen ed o acqui e ames om he C-RED One came a h ough he EDT
ame g abbe . On s a up, i loads a con igu a ion ile and p og ams came a se ings o e he de ice con ol
channel (image size, ame a e, c opping ange). Then, i s a s a eal- ime loop ha blocks o he nex ame,
alida es i s size, and checks he came a’s ame coun e o de ec d ops. F ames a e published in o sha ed
memo y o downs eam p ocessing.
4.2 Resul s
Figu e 9, shows he came a acqui ion imes measu ed on ecep ion by COSMIC. The mean acquisi ion ime is
333.44
µ
s, co esponding o he 3 kHz ame a e. On he op o ha , he e y low a e age ji e measu ed o
0.58
µ
s suppo s ha he ame g abbe is capable o acqui ing ames in a e y s able way.
Figu e 9: Came a acquisi ion la ency
The esul s demons a es ha he ame g abbe is capable o s ably acqui ing ames a 3000Hz, wi hou
any loss.
5. END-TO-END REAL-TIME GPU PIPELINE
In his sec ion we will desc ibe he RTC pipeline es se up and i s pe o mance esul s.
5.1 RTC pipeline es se up
The inal es is o alida e he en i e eal- ime GPU pipeline, wi h he ame g abbe acqui ing ames a 3 kHz
om he C-RED One came a while he sFPDP ansmission is unning a 1 kHz.
As shows Figu e 10, HRTC ope a es in baseline mode. I includes he ollowing s eps o he came a ames:
he C-RED One came a ames a e acqui ed a 3 kHz
he ames a e copied om CPU o GPU
he ames a e calib a ed and he slopes a e ex ac ed on GPU
a linea econs uc ion is pe o med o econs uc wa e on om slopes
he in eg a o con ol law is applied o compu e DM commands
he DM commands a e copied om GPU o CPU o be sen o he mi o
Figu e 10: RTC pipeline se up
A he same ime, he sFPDP acquisi ion is unning:
he sFPDP acquisi ion is unning a 1 kHz
he i s s age da a is copied om CPU o GPU
a ma ix ec o mul iplica ion is pe o med o synch onize bo h s ages
he da a is copied on CPU o be expo ed o eleme y
5.2 Resul s
The la ency p esen ed in Figu e 11 has been measu ed each ime a ame was a ailable on CPU. The ull chain
achie es a mean la ency o 78.53
µ
s, which is well below he 360
µ
s maximal ime budge . On op o ha , he
a e age ji e measu ed is 2.27
µ
s which demons a es he sys em eliabili y and s abili y.
Figu e 11: RTC la ency esul s
6. CONCLUSION AND PERSPECTIVES
On he so wa e de elopmen side, mos o he S-RTC da a asks and H-RTC supe iso s ha e been implemen ed.
Uni es s ha e also been de eloped, he nex s ep is o p oceed wi h unc ional es s. Fo his pu pose we ha e
in eg a ed he Compass simula o 11 in o he H-RTC in o de o gene a e ealis ic wa e on inpu s. This will
allow us o alida e he RTC’s beha iou on ep esen a i e da a. Fu u e de elopmen s on he H-RTC will aim
o implemen he emaining i e con olle s.
On he ha dwa e in e ace in eg a ion side, he sFPDP and came a in e aces ha e been success ully in eg a ed
in o he H-RTC p o o ype. The es s pe o med demons a e ha bo h in e aces a e capable o acqui ing da a
a he equi ed equencies (1 kHz o sFPDP and 3 kHz o he came a) wi h e y low a e age ji e . The inal
es alida es ha he en i e RTC pipeline can un on GPU wi h a mean la ency well below he 360
µ
s ime
