23
In e na ional Jou nal o Ad ance and Applied Resea ch
www.ijaa .co.in
ISSN – 2347-7075
Impac Fac o – 8.141
Pee Re iewed
Bi-Mon hly
Vol. 6 No. 38
Sep embe - Oc obe - 2025
Ene gy-e icien AI Ha dwa e using Silicon Oxyni ide In eg a ed Pho onics
A chana Chaudha i
Assis an P o esso ,
Depa men o Compu e Science, D . D. Y. Pa il Science & Compu e Science College,
Aku di, Pune-411044
Co esponding Au ho – A chana Chaudha i
DOI - 10.5281/zenodo.17309826
Abs ac :
A i icial In elligence (AI) wo kloads demand unp eceden ed compu a ional powe and
ene gy e iciency. Pho onic ha dwa e has eme ged as a p omising al e na i e o con en ional
elec onics due o i s ul a- as signal p opaga ion, low la ency, and educed ene gy consump ion. In
his wo k, we explo e Silicon Oxyni ide (SiON) as a pho onic pla o m o AI ha dwa e, highligh ing
i s low p opaga ion loss, CMOS compa ibili y, and scalabili y. We discuss SiON wa eguide-based
a chi ec u es o op ical neu al ne wo ks, emphasizing hei po en ial o achie e a o able ene gy–
la ency ade-o s. Ou indings sugges ha SiON in eg a ed pho onics o e s a iable pa h owa ds
nex -gene a ion ene gy-e icien AI accele a o s.
Keywo ds: Silicon Oxyni ide, In eg a ed Pho onics, Op ical Neu al Ne wo ks, A i icial
In elligence Ha dwa e, Ene gy E iciency.
In oduc ion:
The exponen ial g ow h o AI
applica ions has c ea ed an u gen need o
ene gy-e icien ha dwa e accele a o s.
Con en ional CMOS-based sys ems ace
bo lenecks in e ms o powe consump ion
and la ency. In eg a ed pho onics o e s a
dis up i e al e na i e by le e aging ligh o
compu a ion and communica ion. Among
pho onic pla o ms, Silicon Oxyni ide (SiON)
has gained a en ion due o i s low-loss
p opaga ion, b oad anspa ency window, and
CMOS-compa ible ab ica ion.
E olu ion o AI Ha dwa e (CMOS → GPUs
→ TPUs → Pho onics):
1. CMOS-based CPUs: Ea ly AI algo i hms
(be o e ~2010) an p ima ily on gene al-
pu pose CPUs. CMOS scaling (Moo e’s
Law) enabled inc eases in clock speed and
ansis o densi y. Bu CPUs a e se ial
a chi ec u es, which is ine icien o AI
wo kloads ha equi e massi e pa allelism.
This CMOS based CPU has limi a ion o
ene gy bo leneck and slow aining o la ge-
scale neu al ne wo ks.
2. GPUs (G aphics P ocessing Uni s):
A ound 2010, esea che s began using GPUs
o deep lea ning [5]. GPUs a e massi ely
pa allel, handling housands o h eads
simul aneously. They a e capable o aining
o deep con olu ional neu al ne wo ks
(CNNs) in days ins ead o mon hs. Bu he
main limi a ions a e high powe consump ion
(hund eds o wa s pe chip), memo y
bo lenecks, and scaling ine iciency o e e -
la ge AI models.
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
A chana Chaudha i
24
3. TPUs (Tenso P ocessing Uni s): Google
in oduced TPUs in 2016 as ASICs op imized
o AI wo kloads. Specialized o ma ix
mul iplica ions (co e ope a ion in neu al
ne wo ks). In his, we achie ed highe
pe o mance(wa ) as compa ed o GPUs o
in e ence and aining. Bu s ill we e lagged in
bounding by esis i e-capaci i e (RC) delay,
hea dissipa ion, and ene gy cos s o mo ing
da a (memo y wall).
4. In eg a ed Pho onics o AI: To o e come
CMOS scaling limi s, esea che s s a ed using
non- adi ional ha dwa e like
Neu omo phic chips (IBM T ueNo h, In el
Loihi) which a e b ain-inspi ed spiking
ne wo ks. Also used he analog in-memo y
compu ing ha compu e inside memo y
a ays o educe da a ans e .
Apa om his i we use Pho onic
ha dwa e which uses ligh o compu a ion,
p omising ul a-low la ency and ene gy
e iciency. We can use he concep o ONN
(Op ical Neu al Ne wo ks) o AI ha dwa e,
ha pe o m ma ix mul iplica ion ia
in e e ence in wa eguide meshes (e.g.,
MZIs).[7] This causes speed o ligh
p opaga ion wi h ul a as in e ence, low
ene gy pe ope a ion ((< 1 J/MAC in heo y)
and na u al pa allelism due o wa eleng h-
di ision mul iplexing. Bu he ad an ages o
Pho onics o e elec onics a e Ene gy
E iciency: Op ical in e connec s educe he
ene gy cos o da a mo emen , a majo
bo leneck in elec onic accele a o s. Low
La ency: Pho ons a el a he speed o ligh
wi h minimal delay, enabling ul a- as
in e ence. Scalabili y: Mul iple wa eleng hs
can ca y pa allel in o ma ion s eams on he
same wa eguide.
Among pho onic ma e ials, Silicon
Oxyni ide (SiON) is p omising because o
Low p opaga ion loss, Wide anspa ency
window and CMOS-compa ible ab ica ion.
Du ing 2017–2019, in eg a ed op ical
ma ix mul iplica ion we e done using silicon
and silicon ni ide pla o ms. La e a e 2020
Hyb id op ical–elec ical a chi ec u es o
deep lea ning in e ence had been in use.
Eme gence o p og ammable pho onic chips
wi h hund eds o unable elemen s a e poin ing
owa ds scalable ONNs. Bu ole o Silicon
Oxyni ide (SiON) in ONN a e p omisingly
inc easing. In ONN (Op ical-Nano-Ni ide)
esea ch, Silicon Oxyni ide (SiON) (SiNxOy)
is a c ucial ma e ial due o i s unable op ical
and elec ical p ope ies, b idging he gap
be ween silicon oxide and silicon ni ide. As
SiOn has a b oad e ac i e index ange o
wa eguides and B agg g a ings in in eg a ed
op ics, a high-quali y dielec ic insula o o
high-powe elec onics and MEMS by
con olling i s oxygen- o-ni ogen a io, used
o op oelec onic de ices and su ace
passi a ion in elec onics.
Me hodology:
Why SiON ? : SiON is a balanced
ma e ial be ween SiO2, Si3N4. I has
unable e ac i e index (1.45–2.0) be e o
con ollable con inemen . Low p opaga ion
loss (<0.1 dB/cm achie able). CMOS-
compa ible ab ica ion using s anda d
Model o SiON
Pho onic wa eguide
MZI mesh wi h
pa allel channels
Ene gy & la ency
ela ion
Compa a i e s udy
o pho onic
wa eguide
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
A chana Chaudha i
25
LPCVD/PECVD p ocesses. [3] Wide
anspa ency window ( isible (blue) o nea -IR
2 µm). The mally s able and less p one o
nonlinea abso p ion han silicon. Mos
op ical-AI chips use Si o SiN. SiON ma e ial
mos ly es ablish a low-loss, CMOS- iendly
pla o m and cha ac e ize i s nonlinea i ies
which makes a solid base o SiON- o -AI
con ibu ion.
Why SiON pho onic Wa eguide? :
Mos ly Wa eguide A chi ec u es should be
Single-mode ib/s ip ype o good
con inemen , low loss. SiON pho onic
wa eguide p o ides he same. Along wi h his,
i used as Mul imode wa eguides o mode-
di ision mul iplexing. Fo A ayed
Wa eguide G a ings (AWGs), SiON is
popula in elecom WDM componen s. MZI
meshes & econ igu able ci cui s which a e
sui able o op ical neu al ne wo k laye s. Fig-
1 shows he basic s uc u e o SiON based
pho onic wa eguide.
Why SiON Pho onic wa eguide o
AI? : Ad an ages o AI Pho onics includes
low cumula i e loss in deep meshes which is
he key o la ge ma ix mul iplica ion ci cui s
in ONNs. Tunable index con as which
balances oo p in (no as bulky as SiN, no as
nonlinea /uns able as Si). Hyb id in eg a ion
po en ial like de ec o s, modula o s, and
lase s can be bonded. Nonlinea i ies
(B illouin/Ke ) ha p omises o op ical
nonlinea ac i a ion unc ions.
SiON-based A chi ec u es o AI Ha dwa e:
We econ igu e Mach–Zehnde
in e e ome e (MZI) meshes o implemen
a bi a y uni a y (o nea -uni a y) linea
ans o ms, enabling ma ix– ec o
mul iplica ion (MVM) in a single op ical pass.
On SiON, MZIs a e ealized wi h low-loss
wa eguides and he mo/elec o-op ic phase
shi e s. To s uc u e his wa eguide we ha e
o sol e he decomposi ion like any uni a y U
∈ CN×N can be ac o ized in o a sequence o
2×2 beam-spli e (MZI) elemen s and phase
shi s (e.g., Reck o Clemen s decomposi ion),
enabling linea laye s o a neu al ne wo k.
SiON educed cumula i e inse ion loss in
deep meshes, enabling la ge N be o e signal
sinks below de ec o noise.
Op ic
al
Inpu
Pho o
De ec o
s
Phase
Shi e
SiON
Wa eguide
SiON based Pho onic Wa eguide o AI Ha dwa e
Si Subs a e
SiO2 Cladding
SiO2 Cladding
SiOXNY Co e
Fig1: S uc u e o SiON
pho onic wa eguide
Fig2: Block diag am o SiON Pho onic wa eguide o AI
ha dwa e
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
A chana Chaudha i
26
Simula ion Model o ene gy-la ency
analysis:
We p oposed a model model o SiON
pho onic AI ha dwa e ha uses MZI mesh
wi h pa allel channels. He e we a e
conside ing he SiON pho onic wa eguide +
MZI mesh as a ma ix mul iplica ion
engine.
Each mul iplica ion in ol es he pa ame e s
like Op ical loss pe MZI (α _MZI),
Phase shi e ene gy E by ela ion EMZI =
CV2
Wa eguide P opaga ion delay by
. [5]
Mul iply-accumula e is gi en by
Fig3: ene gy–la ency cu e
Fig 4: Compa a i e s udy
Discussion:
Fig3 cu e ep esen s he ene gy–
la ency ade-o . Longe wa eguides, mo e
la ency in u n slowe is he in e ence. Bu
sligh ly mo e op ical loss [8]. Phase-shi e
ene gy domina es, gi ing ~5 pJ/MAC
baseline.
Fig4 ep esen s he compa a i e
s udy o ene gy–la ency ade-o wi h
SiON, CMOS CPU, GPU, TPU & Si. CMOS
CPU (~100 pJ, 50 ns), GPU (~20 pJ, 10 ns),
TPU (~10 pJ, 5 ns), Si Pho onics (~1 pJ, 2 ns)
[4][1][6]. I clea ly shows how SiON can b idge
he gap be ween cu en elec onics (high
ene gy, lowe la ency) and u u e pho onics
(ul a-low ene gy, low la ency). Fu he he
SiON cu e ends owa ds be e e iciency
as pa allelism inc eases.
Challenges and Fu u e Di ec ions:
S ill SiON is no as compac as Si, bu
mo e s able han SiN. SiON based ecosys em
is s ill in de eloping mode as compa ed o
Si/SiN wi h limi ed indus ial-scale suppo .
Ac i e de ice in eg a ion like as modula o s,
de ec o s is limi ed in case o SiON as
compa ed o Si/SiN.
Conclusion:
We ha e ou lined he po en ial o
Silicon Oxyni ide in eg a ed pho onics as a
ma e ial pla o m o ene gy-e icien AI
ha dwa e. By enabling low-loss, scalable, and
CMOS-compa ible a chi ec u es, SiON
wa eguides ep esen a p omising s ep
owa ds nex -gene a ion op ical neu al
IJAAR Vol. 6 No. 38 ISSN – 2347-7075
A chana Chaudha i
27
ne wo ks. SiON shows p omising ole because
o i s ailo abili y and mode a e loss, making
i a ac i e o expe imen al AI pho onic
accele a o s.
Re e ences:
1. Adam K zywaniak
[0000−0003−1904−2510], Pawel
Cza nul [0000−0002−4918−9196] and
Je zy P o icz
[0000−0003−2975−9339]
2. Gian-Luca Bona (Swiss Fede al
Labo a o ies o Ma e ials Science and
Technology), Roland Ge mann (IBM),
and B. J. O ein, IBM Jou nal o
Resea ch & de elopmen 47(2.3):239
– 249 DOI: 10.1147/ d.472.0239
3. In e na ional Con e ence on
Mic oelec onics (ICM), 2000 (IEEE
Ca . No.00EX453) "Da a-d i en
dynamic logic e sus NP-CMOS
logic, a compa ison"
4. Jinhwi Kim, Apos olos Galanopoulos,
Jude Vi ek Joseph, Jeongho Kwak
ICTC
10.1109/ICTC49870.2020.9289270
5. K izhe sky, Alex; Su ske e , Ilya;
Hin on, Geo ey E. (2017-05-
24). "ImageNe classi ica ion wi h
deep con olu ional neu al
ne wo ks" ACM. 60 (6): 8490. doi:10.
1145/3065386. ISSN 0001-
0782. S2CID 195908774.
6. Mu ad Qasaimeh, K is o Denol , Jack
Lo, Kees Visse s, Joseph Zamb eno,
and Phillip H. Jones,
IEEE10.1109/ICESS.2019.8782524
7. Rui Tang, Mako o Okano, Chao
Zhang, Kasidi Top ase pong,
Shinichi Takagi, Mi su u Takenaka,
384770862, “Wa eguide mul iplexed
pho onic ma ix ec o mul iplica ion
p ocesso using mul ipo
pho ode ec o s.”
8. Zhiping Zhou, Bing Yin, Qingzhong
Deng, Xinbai Li, and Jishi Cui Vol. 3,
Issue 5, pp. B28-B46 (2015)
10.1364/PRJ.3.000B28
