gfex - cern...a single board system. both efex and jfex are multi-board systems. 6 see reiner...
TRANSCRIPT
gFEXthe ATLAS Calorimeter Trigger Global Feature
Extractor
Helio Takai, Michael Begel and Hucheng Chen Brookhaven National Laboratory
21st International Conference on Computing in High Energy Physics Okinawa, Japan - April 13-17,2015
1
The ATLAS Experiment
2
Phase I UpgradeLuminosity ~ 2 x1034 cm-2s-1
25 ns bunch crossing Pile up of <𝜇>~60
Level 1 rate =100 kHz HLT rate = 1 kHz
Motivation
3
• High pT bosons and fermions are a key component of ATLAS physics. • W,Z and H bosons, top quarks and exotic particles. • Many analyses with boosted objects.
• Analyses that addresses this kind of physics use large R jets with R > 1.
• The ATLAS Level 1 trigger is designed for narrow jets, with limited acceptance for large objects.
y-1 -0.5 0 0.5 1 1.5 2 2.5
φ
0
0.5
1
1.5
2
2.5 Preliminary Simulation ATLAS = 1.75 TeVZ’m event, tt → Z’
= 186.5 GeVWbm = 77.3 GeV, Wm
= 1.0R tkAnti-Calorimeter clusters
= 0.2RSubjets, C/A bosonW
jetbTop radiationISR
L1 Jet
Mass [GeV]0 50 100 150 200 250 300 350 400
Entri
es /
10 G
eV
0
50
100
150
200
250
300
1 subjet2 subjets3 subjets4+ subjets
Simulation PreliminaryATLAS,-1 L dt = 20.3 fb∫ = 8 TeVs
|<1.2η>200 GeV, |T
Trimmed jet p
narr
ow je
ts
W/Z
Top
Increase trigger efficiency for Fat-jets in ATLAS
Mass [GeV]0 50 100 150 200 250 300 350 400
Entri
es /
10 G
eV
0
50
100
150
200
250
300
1 subjet2 subjets3 subjets4+ subjets
Simulation PreliminaryATLAS,-1 L dt = 20.3 fb∫ = 8 TeVs
|<1.2η>200 GeV, |T
Trimmed jet p
Mass [GeV]0 50 100 150 200 250 300 350 400
Entri
es /
10 G
eV
0
50
100
150
200
250
300
1 subjet2 subjets3 subjets4+ subjets
Simulation PreliminaryATLAS,-1 L dt = 20.3 fb∫ = 8 TeVs
|<1.2η>200 GeV, |T
Trimmed jet p
No
trev
iew
ed
,fo
rin
tern
al
circu
la
tio
no
nly
[GeV]jet
Tpuncalibrated
100 200 300 400 500E
vent E
ffic
iency
0.0
0.2
0.4
0.6
0.8
1.0
=80⟩µ⟨=14 TeV s tt R=1.0 (5% trimmed)Tkanti-
>20 GeVT
p D=0.3 subjet with Tk1 ≥
|<2.5jet
η| [GeV]<220
jet100<m
L1_J100 (Run 1 L1Calo sim.)L1_HT200 (Run 1 L1Calo sim.)L1_G140 (seed>15 GeV)
ATLAS PreliminarySimulation
[GeV]jet
Tpuncalibrated
100 200 300 400 500
Eve
nt E
ffic
iency
0.0
0.2
0.4
0.6
0.8
1.0
=80⟩µ⟨=14 TeV sbb ν l→WH R=1.0 (5% trimmed)Tkanti-
>20 GeVT
p D=0.3 subjet with Tk1 ≥
|<2.5jet
η| [GeV]<150
jet100<m
L1_J100 (Run 1 L1Calo sim.)L1_HT200 (Run 1 L1Calo sim.)L1_G140 (seed>15 GeV)
ATLAS PreliminarySimulation
Figure 2: Event-level Level-1 trigger e�ciency turn-on curves as a function of the o✏ine trimmed jetpT for (left) tt̄ events and (right) WH ! `⌫bb̄ events. Both processes are simulated with a pile-up levelequivalent to an average number of interactions per bunch-crossing, or hµi, of 80. The e�ciency curvesare shown as a function of the o✏ine trimmed anti-kt R = 1.0 jet pT , where the o✏ine trimmed jet isrequired to have a mass between 100 < mjet < 220 GeV (left, top) or 100 < mjet < 150 GeV (right,Higgs). The trimming parameters specify that any subjets with a pT fraction of the original jet lessthan 5% are to be discarded. The subjets are defined using the kt clustering algorithm with a nominalradius parameter of D = 0.3. In each case, three trigger selections are shown: (blue open circles) fullsimulation of the existing Run I Level-1 calorimeter jet trigger with a 100 GeV threshold, (black opensquares) full simulation of a sum jet transverse energy (ET ) trigger, or HT trigger, with a 200 GeVthreshold, formed using Run 1 Level-1 calorimeter jets, (red closed circles) and a Phase I gFEX-basedreconstruction algorithm with a 140 GeV threshold [1]. The gFEX reconstruction implements a simpleseeded cone algorithm with a nominal radius of R = 1.0 and with a seed selection of 15 GeV applied tocalorimeter towers with area 0.2 ⇥ 0.2 in ⌘ ⇥ �. The 140 GeV gFEX trigger threshold is chosen to matchthe L1 J100 single subjet turn-on curve.
2
No
trev
iew
ed
,fo
rin
tern
al
circu
la
tio
no
nly
1 Level 1 jet trigger e�ciency turn-on curves
[GeV]jet
Tpuncalibrated
100 200 300 400 500
per-
Jet E
ffic
iency
0.0
0.2
0.4
0.6
0.8
1.0
=80⟩µ⟨=14 TeV s tt R=1.0 (5% trimmed)Tkanti-
>20 GeVT
p D=0.3 subjet with Tk1 ≥
R>2∆|<2.5; isolated by jet
η|
L1_J100 (Run 1 L1Calo sim.)
1 subjet
2 subjets
3 subjets≥
L1_G140
1 subjet
2 subjets
3 subjets≥
ATLAS PreliminarySimulation
[GeV]jet
Tpuncalibrated
100 200 300 400 500
per-
Jet E
ffic
iency
0.0
0.2
0.4
0.6
0.8
1.0
=80⟩µ⟨=14 TeV s tt R=1.0 (5% trimmed)Tkanti-
>20 GeVT
p D=0.3 subjet with Tk1 ≥
R>2∆|<2.5; isolated by jet
η|
L1_J100 (Run 1 L1Calo sim.) 1 subjet 2 subjets
3 subjets≥
ATLAS PreliminarySimulation
Figure 1: Per-jet e�ciency turn-on curves in Monte Carlo (MC) simulation for multiple Phase I upgradeLevel-1 jet trigger options. A global feature extraction (gFEX) reconstruction algorithm (closed redmarkers, left) from the TDAQ Phase I Upgrade Technical Design Report (TDR) [1] with a 140 GeVthreshold is compared to full simulation of the Run I Level-1 calorimeter jet trigger (open blue markers,left and right) with a 100 GeV threshold. The gFEX reconstruction implements a simple seeded conealgorithm with a nominal radius of R = 1.0 and with a seed selection of 15 GeV applied to calorimetertowers with area 0.2⇥ 0.2 in ⌘⇥�. The 140 GeV gFEX trigger threshold is chosen to match the L1 J100single subjet turn-on curve. Pair-produced top quark MC simulation samples are simulated with a pile-up level equivalent to an average number of interactions per bunch-crossing, or hµi, of 80. For eachalgorithm, the e�ciency curves are shown as a function of the o✏ine trimmed anti-kt R = 1.0 jet pT
with di↵erent o✏ine subjet multiplicities. The trimming parameters specify that any subjets with a pT
fraction of the original jet less than 5% are to be discarded. The subjets are defined using the kt clusteringalgorithm with a nominal radius parameter of D = 0.3. For subjet counting, the subjets are required tohave a subjet pT > 20 GeV. The o✏ine trimmed jets are required to be isolated from any other o✏inejet by at least a radial distance of �R > 2.0 radians and to be within the pseudorapidity range |⌘| < 2.5.The turn-on curves measure per-jet e�ciencies after requiring a that the the Level-1 gFEX jet be within�R < 1.0 of the o✏ine trimmed jet.
1
boosted top
Simulation Studies
H bb -
Simulations performed for 14 TeV, and for a <𝜇> = 80.
Larger trigger acceptance
4
Simulation Studies
Correlation between the event energy density (x) and estimated by gFEX (y). The correlation is better than 90%.
Jet energy pile up subtraction
offline calculation
gFEX
cal
cula
tion
offline calculationgF
EX c
alcu
latio
n
5
MuCTPi
Barrel
Endcap
Muon
LiquidArgon
TileCal.
LTDBDPS
TBB
JetFeatureExtractor
ElectronFeatureExtractor
GlobalFeatureExtractor
OpticalPlant
nMCM
pre-Processor
Hub
ROD
Hub
ROD
Hub
ROD
CMX
CMX
Jet EnergyProcessor
ClusterProcessor
L1Topo
L1CTP
L1A
toDAQ
ROI
Hit C
ount
s
TOBs
supercells
η x φ0.2x0.2
η x φ0.1x0.1
e/γ, τ
Jets, τ, ΣE ,E T Tmiss
Large Jets,E Tmiss
L1Calo
gFEX in ATLAS L1Calo
gFEX is a component of the ATLAS L1 Calo system in Phase I upgrade. It complements the electron and jet feature extractors. It is a single board system. Both eFEX and jFEX are multi-board systems.
6
See Reiner Hauser’s talk on ATLAS data flow in Track 1, tomorrow.
See Eduard Simioni’s presentation in Track 1
tomorrow.
See Julian Glatzer’s presentation in this
session
October 16, 2014 – 13 : 53 DRAFT 5
#3 #1 #2 #4 #3 #1 #2 #4
Figure 4 Left: Seeding step for identifying large-R jets by selecting towers over a threshold ET value.Right: Summing the energy around the seeds within ∆R ! 1.0.
#3 #1 #2 #4
Figure 5 The final large-R jets. Each jet is stored on the Processor FPGA that produced the seed.
techniques. For this study, the seeding step required ET > 15 GeV and a noise threshold of ET > 3 GeV
was applied to all gTowers entering the sum. The results are shown in Fig. 6. The open symbols represent
the efficiencies as would be measured by the Run 1 L1Calo for a threshold of 100 GeV (L1 J100). The
red closed symbols correspond to a 140 GeV gFEX-jet threshold (L1 G140), which was matched to the130
L1 J100 plateau for QCD-like jets with only one subjet. The turn-on curves in Fig. 6 (left) measure
per-jet efficiencies as a function of the offline trimmed R = 1.0 jets, isolated (∆R > 2.0) and matched to
the online objects within ∆R < 1.0. There are three curves for each of the J100 and the G140 selections,
corresponding to the number of subjets within the offline trimmed jets (1 through 3). These subjets are
[GeV]jetTpuncalibrated
100 200 300 400 500
per-J
et E
fficie
ncy
0.0
0.2
0.4
0.6
0.8
1.0
=80⟩µ⟨=14 TeV s tt R=1.0 (5% trimmed)Tkanti-
>20 GeVTp D=0.3 subjet with Tk1 ≥
R>2∆|<2.5; isolated by jetη|
L1_J100 (Run 1 L1Calo sim.) 1 subjet 2 subjets
3 subjets≥
L1_G140 1 subjet 2 subjets
3 subjets≥
ATLAS PreliminarySimulation
[GeV]jetTpuncalibrated
100 200 300 400 500
Even
t Effi
cienc
y
0.0
0.2
0.4
0.6
0.8
1.0
=80⟩µ⟨=14 TeV s tt R=1.0 (5% trimmed)Tkanti-
>20 GeVTp D=0.3 subjet with Tk1 ≥
|<2.5jetη|
[GeV]<220jet100<m L1_J100 (Run 1 L1Calo sim.)L1_HT200 (Run 1 L1Calo sim.)L1_G140 (seed>15 GeV)
ATLAS PreliminarySimulation
[GeV]jetTpuncalibrated
100 200 300 400 500
Even
t Effi
cienc
y
0.0
0.2
0.4
0.6
0.8
1.0
=80⟩µ⟨=14 TeV sbb ν l→WH R=1.0 (5% trimmed)Tkanti-
>20 GeVTp D=0.3 subjet with Tk1 ≥
|<2.5jetη|
[GeV]<150jet100<m L1_J100 (Run 1 L1Calo sim.)L1_HT200 (Run 1 L1Calo sim.)L1_G140 (seed>15 GeV)
ATLAS PreliminarySimulation
Figure 6 Trigger efficiency turn-on curves comparing the gFEX R = 1.0 jet trigger to the Run 1 styleL1 J100 and L1 HT200 jet triggers (both expected to be unprescaled in Run 2). All samplesuse the ⟨µ⟩ = 80 Upgrade Monte Carlo simulation. The left two plots are for tt̄ while theright plot is WH → ℓνbb̄. The left plot shows the efficiency per “isolated” jet binned in thenumber of subjets identified offline, while the right two plots display the event-level triggerefficiency. The first 12 bunches from each bunch train were removed prior to analysis incorrespondence with the TDAQ TDR [2].
October 16, 2014 – 13 : 53 DRAFT 5
#3 #1 #2 #4 #3 #1 #2 #4
Figure 4 Left: Seeding step for identifying large-R jets by selecting towers over a threshold ET value.Right: Summing the energy around the seeds within ∆R ! 1.0.
#3 #1 #2 #4
Figure 5 The final large-R jets. Each jet is stored on the Processor FPGA that produced the seed.
techniques. For this study, the seeding step required ET > 15 GeV and a noise threshold of ET > 3 GeV
was applied to all gTowers entering the sum. The results are shown in Fig. 6. The open symbols represent
the efficiencies as would be measured by the Run 1 L1Calo for a threshold of 100 GeV (L1 J100). The
red closed symbols correspond to a 140 GeV gFEX-jet threshold (L1 G140), which was matched to the130
L1 J100 plateau for QCD-like jets with only one subjet. The turn-on curves in Fig. 6 (left) measure
per-jet efficiencies as a function of the offline trimmed R = 1.0 jets, isolated (∆R > 2.0) and matched to
the online objects within ∆R < 1.0. There are three curves for each of the J100 and the G140 selections,
corresponding to the number of subjets within the offline trimmed jets (1 through 3). These subjets are
[GeV]jetTpuncalibrated
100 200 300 400 500
per-J
et E
fficie
ncy
0.0
0.2
0.4
0.6
0.8
1.0
=80⟩µ⟨=14 TeV s tt R=1.0 (5% trimmed)Tkanti-
>20 GeVTp D=0.3 subjet with Tk1 ≥
R>2∆|<2.5; isolated by jetη|
L1_J100 (Run 1 L1Calo sim.) 1 subjet 2 subjets
3 subjets≥
L1_G140 1 subjet 2 subjets
3 subjets≥
ATLAS PreliminarySimulation
[GeV]jetTpuncalibrated
100 200 300 400 500
Even
t Effi
cienc
y
0.0
0.2
0.4
0.6
0.8
1.0
=80⟩µ⟨=14 TeV s tt R=1.0 (5% trimmed)Tkanti-
>20 GeVTp D=0.3 subjet with Tk1 ≥
|<2.5jetη|
[GeV]<220jet100<m L1_J100 (Run 1 L1Calo sim.)L1_HT200 (Run 1 L1Calo sim.)L1_G140 (seed>15 GeV)
ATLAS PreliminarySimulation
[GeV]jetTpuncalibrated
100 200 300 400 500
Even
t Effi
cienc
y
0.0
0.2
0.4
0.6
0.8
1.0
=80⟩µ⟨=14 TeV sbb ν l→WH R=1.0 (5% trimmed)Tkanti-
>20 GeVTp D=0.3 subjet with Tk1 ≥
|<2.5jetη|
[GeV]<150jet100<m L1_J100 (Run 1 L1Calo sim.)L1_HT200 (Run 1 L1Calo sim.)L1_G140 (seed>15 GeV)
ATLAS PreliminarySimulation
Figure 6 Trigger efficiency turn-on curves comparing the gFEX R = 1.0 jet trigger to the Run 1 styleL1 J100 and L1 HT200 jet triggers (both expected to be unprescaled in Run 2). All samplesuse the ⟨µ⟩ = 80 Upgrade Monte Carlo simulation. The left two plots are for tt̄ while theright plot is WH → ℓνbb̄. The left plot shows the efficiency per “isolated” jet binned in thenumber of subjets identified offline, while the right two plots display the event-level triggerefficiency. The first 12 bunches from each bunch train were removed prior to analysis incorrespondence with the TDAQ TDR [2].
Finding Large Radius Jets
325 ns latency. 7
October 16, 2014 – 13 : 53 DRAFT 5
#3 #1 #2 #4 #3 #1 #2 #4
Figure 4 Left: Seeding step for identifying large-R jets by selecting towers over a threshold ET value.Right: Summing the energy around the seeds within ∆R ! 1.0.
#3 #1 #2 #4
Figure 5 The final large-R jets. Each jet is stored on the Processor FPGA that produced the seed.
techniques. For this study, the seeding step required ET > 15 GeV and a noise threshold of ET > 3 GeV
was applied to all gTowers entering the sum. The results are shown in Fig. 6. The open symbols represent
the efficiencies as would be measured by the Run 1 L1Calo for a threshold of 100 GeV (L1 J100). The
red closed symbols correspond to a 140 GeV gFEX-jet threshold (L1 G140), which was matched to the130
L1 J100 plateau for QCD-like jets with only one subjet. The turn-on curves in Fig. 6 (left) measure
per-jet efficiencies as a function of the offline trimmed R = 1.0 jets, isolated (∆R > 2.0) and matched to
the online objects within ∆R < 1.0. There are three curves for each of the J100 and the G140 selections,
corresponding to the number of subjets within the offline trimmed jets (1 through 3). These subjets are
[GeV]jetTpuncalibrated
100 200 300 400 500
per-J
et E
fficie
ncy
0.0
0.2
0.4
0.6
0.8
1.0
=80⟩µ⟨=14 TeV s tt R=1.0 (5% trimmed)Tkanti-
>20 GeVTp D=0.3 subjet with Tk1 ≥
R>2∆|<2.5; isolated by jetη|
L1_J100 (Run 1 L1Calo sim.) 1 subjet 2 subjets
3 subjets≥
L1_G140 1 subjet 2 subjets
3 subjets≥
ATLAS PreliminarySimulation
[GeV]jetTpuncalibrated
100 200 300 400 500
Even
t Effi
cienc
y
0.0
0.2
0.4
0.6
0.8
1.0
=80⟩µ⟨=14 TeV s tt R=1.0 (5% trimmed)Tkanti-
>20 GeVTp D=0.3 subjet with Tk1 ≥
|<2.5jetη|
[GeV]<220jet100<m L1_J100 (Run 1 L1Calo sim.)L1_HT200 (Run 1 L1Calo sim.)L1_G140 (seed>15 GeV)
ATLAS PreliminarySimulation
[GeV]jetTpuncalibrated
100 200 300 400 500
Even
t Effi
cienc
y
0.0
0.2
0.4
0.6
0.8
1.0
=80⟩µ⟨=14 TeV sbb ν l→WH R=1.0 (5% trimmed)Tkanti-
>20 GeVTp D=0.3 subjet with Tk1 ≥
|<2.5jetη|
[GeV]<150jet100<m L1_J100 (Run 1 L1Calo sim.)L1_HT200 (Run 1 L1Calo sim.)L1_G140 (seed>15 GeV)
ATLAS PreliminarySimulation
Figure 6 Trigger efficiency turn-on curves comparing the gFEX R = 1.0 jet trigger to the Run 1 styleL1 J100 and L1 HT200 jet triggers (both expected to be unprescaled in Run 2). All samplesuse the ⟨µ⟩ = 80 Upgrade Monte Carlo simulation. The left two plots are for tt̄ while theright plot is WH → ℓνbb̄. The left plot shows the efficiency per “isolated” jet binned in thenumber of subjets identified offline, while the right two plots display the event-level triggerefficiency. The first 12 bunches from each bunch train were removed prior to analysis incorrespondence with the TDAQ TDR [2].
STEP 1 - Find seeds. Seeds are towers with energy above a set threshold.
STEP 2 - Sum energy from neighboring towers. Concurrently, estimate pileup energy.
STEP 3 - Subtract pileup energy and “join” results. The final result is stored on processing FPGAs.
October 16, 2014 – 13 : 53 DRAFT 22
Hybrid FPGA
FPGA–4 FPGA–3
FPGA–2FPGA–1
/2
\2
/2
\2
/112/112 /112/112
/112
/112
/112
/112
/
20
\
20
\
20
/
20
/
20
\
20
\
20
/
20
Monitoring & TTC communications (MGT)Algorithm connections (GPIO)(50 DDR pairs data + 6 pair clock)
TTC transmission (GPIO)(8 DDR pairs data + 2 pair clock)
IPbus communications (GPIO)(8 Tx & 8 Rx + 4 handshake)
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
o
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
o
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
o
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
o
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
o
L1Cal
oL1C
alo
L1Cal
o
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
o
L1Cal
o
Figure 13 Interconnections used for information transfers between the Processor FPGAs and betweenthe Processor FPGAs and the Hybrid FPGA. The numbers indicate the sum of pins associ-ated with each GPIO connection or the number of MGT.
configuration of gTowers that could be sent to the Hybrid FPGA is detailed in Table 9 (only half the bits
are needed as EM and HAD layers will be pre-summed). Sending these gTowers to the Hybrid FPGA
at the 1 MHz L0A rate in Run 4 requires that each Processor FPGA transmit up to 5.13 Gb/s (including
8b/10b encoding). Each Processor FPGA must also send the TOBs requiring an additional bandwidth of510
0.615 Gb/s. It is also useful, for monitoring purposes, to histogram TOBs in random events at a prescaled
rate independent of the L0A.
Output from the Hybrid FPGA is via an Ethernet connection to an online monitoring system including
a web server running on the Hybrid FPGA. Ethernet connectivity will be implemented via an external
PHY chip (such as the Marvell R⃝ 88E1116R [31]) with an interface to IPcores “Tri mode eth MAC”515
and an “AXI based FIFO” between the IPbus core and the Ethernet UDP layer in the Hybrid FPGA.
The Hybrid FPGA GPIOs in the FPGA fabric can be used for parallel data/control/timing lines with this
design.
gFEX Concept
• Coarse granularity data (𝛥𝜂x𝛥𝜑=0.2x0.2) from calorimeters are received by high speed optical links and processed by four large FPGAs.
• The processing FPGAs are monitored and programmed by a Hybrid FPGA (SoC).
• Results are transferred to the next level in L1 trigger.
8
October 16, 2014 – 13 : 53 DRAFT 3
4×Processor FPGA
rear front
Real-Time Data Path
Readout Data PathAT
CA
Sh
elfM
an
ager
Hu
bT
TC
RO
D
IPMC Card
IPMC
Hybrid FPGA
PHY IPbus Int.
TTC Interface
GBT Interface
Monitoring
opticalfi
ber
plant
Op
tical
Con
n.
/4×72
Op
tical
Rx
/24×12
Deseria
lizer
/264
gT
ow
erB
uild
er
Jet Finder
Pileup Density
Event Moments
TO
BB
uild
er
Seria
lizer
Op
tical
Tx
/4×12
Op
tical
Con
n.
/4×12
L1T
op
o
/1×48
Readout Buffer
Serializer
/4×2
/2×4×1
Op
tical
Con
n.
FE
LIX
TT
C
Op
tical
Rx
Op
tical
Tx/
4×2
2
//3
/7
/ 4 × 1
/1×12
/1×12
/1×24
Ethernet
RJ45PHY
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
o
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
o
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
o
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
o
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
o
L1Cal
oL1C
alo
L1Cal
o
L1Cal
oL1C
alo
L1Cal
oL1C
alo
L1Cal
o
L1Cal
o
Figure 3 A block diagram of the gFEX module. Shown are the real-time (to L1Topo) and readout datapaths (to Hub/ROD and FELIX).
The gFEX is an ATCA module conforming to the PICMG R⃝3.0 Revision 3.0 specification [14]. The
gFEX module will likely be placed in a sparsely populated ATCA shelf so that it can occupy two slots if
needed: one for the board and one for cooling (e.g., large heat sinks), fiber routing, etc.
2.2 Input Data
The gFEX receives data from the electromagnetic and hadronic calorimeters on optical fibers. For most75
of the detector, the towers — called gTowers — correspond to an area of ∆η×∆φ = 0.2×0.2. Calorimeter
locations with towers different from this standard size are detailed in Sec. 3.3.
Calorimeter data are transmitted as continuous serial streams. The gFEX logic must be aligned with
the word boundaries in this serial data for conversion into parallel data. The scheme for achieving this is
not yet specified, however, possible mechanisms are being explored.80
The calorimeter data are accompanied by a CRC code. This is checked in the Processor FPGA
immediately following conversion from serial streams into parallel data. All data with a detected error
are sent to the readout path without modification but are zeroed on the real-time trigger path. In addition,
information about errors is collected as follows:
Input Error Count Incremented for any clock cycle with at least one error in any input channel.85
Input Error Latch A bit is set for any channel with at least one error. These bits remain set until
cleared via an external signal.
Error Check Result Formed from an OR of all error checks for the current BC.
Global Input Error Formed from the OR of all bits in the Input Error Latch.
The Input Error Count, Input Error Latch, Error Check Result, and Input Error bits can be read via an90
IPbus command [15] and will be reported via the Hybrid FPGA. An IPbus command will be able to clear
any or all of these registers. The Error Check Result and Input Error Count are included in the readout
data for the current BC. Additionally, the quality of the input data is assessed:
gFEX: Processing Chain and Interfaces
9
See Jorn Schumacher’s presentation in Track 1
tomorrow.
Floor Plan of Prototype gFEX � Input/Output
o 4 72-fold MPO connectors (back) o 2 48-fold MPO connectors (front) o 24 miniPOD Rx ←− gFOX o 4 miniPOD Tx −→ L1Topo o 1 miniPOD Rx ←− FELIX o 1 miniPOD Tx −→ FELIX o JTAG, UART, RJ45 (front) o ATCA Zone 1 & J23 ADF Plus (back)
� FPGA o 4 Virtex-7 XC7VX690T-FFG1927
(speed grade -3) o ZYNQ XC7Z045-FFG900
� Notes: o neighboring slot empty ⇒ more space
for fiber routing and better cooling o switched to eFEX ATCA form factor
proposal o identifying vendors for 72-fold
MPO;1.1 dB light loss expected (purchase for gFEX, jFEX, FOX?)
2/24/2015 2
gFEX Floor Plan
1. The baseline FPGA is the XC7VX690T-FFG1927 speed grade -3 @ 320 MHz.
2. The SoC is the ZYNQ XC7Z045-FFG900, with dual ARM core and Linux OS (PetaLinux).
3. Input/Output • 4x 72-fold MPO* connectors (IN) • 2x 48 MPO* connectors (OUT) • miniPODs for Rx and Tx • JTAG, UART, RJ45 (Front) • ATCA Zone 1 & J23 ADF+ (Back)
4. The number of input fibers is 264 (at 6.4 Gb/s) or 232 (at 11.2 Gb/s).
10(*)MPO - Multiple-Fiber Push-On/Pull-off
Floor Plan of Pre-prototype gFEX � Input/Output
o 4 72-fold MPO connectors (back) o 1 48-fold MPO connectors (front) o 6 miniPOD Rx ←− gFOX o 1 miniPOD Tx −→ L1Topo o 1 miniPOD Rx ←− FELIX o 1 miniPOD Tx −→ FELIX o JTAG, UART, RJ45 (front) o ATCA Zone 1 & J23 ADF Plus (back)
� FPGA o 1 Virtex-7 XC7VX550T-FFG1927
(speed grade -3) o ZYNQ XC7Z045-FFG900
� Note: o Verify critical technical problems,
such as PCB design, ZYNQ SOC, High speed routing and Vertex 7 FPGA GTH link with high density MiniPods.
o Functional Board. Almost all the important gFEX functions can be tested on this board.
2/24/2015 3
Development Plans
1. gFEX is being developed in stages. The first phase (ongoing) is to prototype a board to assure full integration with L1 Calo. In a second stage a board with four FPGAs will be produced.
2. High speed optical links will be tested. Supported I/O speeds are 6.4, 9.6, 11.2 and 12.8 Gb/s.
3. Integration and link speed tests will take place at the end of 2015.
4. Firmware for data processing is being developed in parallel using commercial boards.
Prototype
11
Pre-prototype Layout
� Schematic Entry : complete � Layout & Routing: 80% � On Schedule for March
2/24/2015 4
Pre-prototype Layout
� Schematic Entry : complete � Layout & Routing: 80% � On Schedule for March
2/24/2015 4
The Prototype Board
Prototype Schematic and Layout are complete! It is implemented in an ATCA board form factor.
12
PCB Stackup � Months of simulation effort to develop
stack up for prototype gFEX module � 26-layer PCB
o high-speed signals in layers 5/7 and 20/22 o back-drilled vias o Megtron 6 PCB material o several vendors make similar boards
� Thickness is ∼ 2.6 mm o mill edges to ATCA standard
Design File: gth_v7_fr4_simulation.ffsHyperLynx LineSim v8.2.1
U1
xilinx_7gth_txxilinx_7gth_tx
1n
1p
U2
xilinx_v7_gth_ami_...rx_p
2n
2pV1TOP
inn...
TL1
61.1 ohms7.435 ps0.050 inCoupled Stackup
TL2
61.1 ohms7.435 ps0.050 inCoupled Stackup
V3TOP
inn...
TL3
65.9 ohms2.108 ns12.000 inCoupled Stackup
TL4
65.9 ohms2.108 ns12.000 inCoupled Stackup
TL5
61.1 ohms1.487 ps0.010 inCoupled Stackup
TL6
61.1 ohms1.487 ps0.010 inCoupled Stackup
V4TOP
inn...
TL7
61.1 ohms1.487 ps0.010 inCoupled Stackup
TL8
61.1 ohms1.487 ps0.010 inCoupled Stackup
V5TOP
inn...
TL9
65.9 ohms351.380 ps2.000 inCoupled Stackup
TL10
65.9 ohms351.380 ps2.000 inCoupled Stackup
TL11
61.1 ohms7.435 ps0.050 inCoupled Stackup
TL12
61.1 ohms7.435 ps0.050 inCoupled Stackup
J1
Xilinx_Virtex-7_...
Port1
Port2
Port3
Port4
J2
Xilinx_Virtex-7_...
Port1
Port2
Port3
Port4
C1
100.0 nF
C2
100.0 nF
GTH-to-GTH Link simulation
14 inch/FR4/with stub/9.6 Gbps 14 inch/FR4/without stub/12.8 Gbps 14 inch/M6/without stub/12.8 Gbps 14 inch/M6/with stub/12.8 Gbps
2/24/2015 5
Challenges
1. gFEX will be a 26 layer board, with a total thickness of ~2.6 mm.
2. The choice of material is critical for the high speed traces within the board. Selected Megtron 6.
13
Summary
The global feature extractor, gFEX, will add to ATLAS the capability to trigger on large radius jets at Level 1.
It is based on FPGA processing. Four large FPGAs will receive coarse data (0.2x0.2) from the EM and HAD calorimeters and select events of interest.
Prototype is now being built. Initial prototype addresses the interfaces with ATLAS L1. After this phase a full processor will be built.
gFEX is processor board that has a large number of high speed I/O lines and could be used in other applications.
14
Extra Slides
October 16, 2014 – 13 : 53 DRAFT 17
Table 6 Count of input fibers by section of the calorimeter for several bandwidth assumptions. Thenumber of bits per gTower is the maximum possible for a given detector configuration andbandwidth.
Link Speed
6.4 Gb/s 9.6 Gb/s 11.2 Gb/s 12.8 Gb/s
Partition Coverage gT
ower
s/F
iber
bit
s/gT
ower
Fib
ers
gT
ower
s/F
iber
bit
s/gT
ower
Fib
ers
gT
ower
s/F
iber
bit
s/gT
ower
Fib
ers
gT
ower
s/F
iber
bit
s/gT
ower
Fib
ers
Barrel EM |η| < 1.6 8 15 64 8 22 64 8 26 64 8 30 64
Tile (Phase I opt.2) |η| < 1.6 16 ? 32 16 ? 32 16 ? 32 16 ? 32
Tile (Phase I opt.3) |η| < 1.6 12 10 48 12 ? 48 12 ? 48 12 ? 48
Tile (Phase II) |η| < 1.6 8 15 64 8 22 64 8 26 64 8 30 64
Standard EMEC 1.6 < |η| < 2.4 7 17 32 7 26 32 7 30 32 7 35 32
Special EMEC 2.4 < |η| < 3.2 10 12 32 14 12 24 14 12 24 20 12 16
HEC 1.5 < |η| < 3.2 12 10 48 18 10 32 18 11 32 18 11 32
FCAL 1 3.1 < |η| < 4.9 12 11 12 16 11 8 16 11 8 22 11 6
FCAL 2&3 3.2 < |η| < 5.0 12 11 12 16 11 8 16 11 8 22 11 6
Total (Phase II) 264 232 232 220
Table 7 Coverage and number of fibers incident on the Processor FPGA.
η φ Fibers at Link Speed (Gb/s)FPGA coverage coverage 6.4 9.6 11.2 12.8
1 −1.6 < η < 0.0 0.0 < φ < 2π 64 64 64 64
2 0.0 < η < 1.6 0.0 < φ < 2π 64 64 64 64
3 1.6 < η < 5.0 0.0 < φ < 2π 68 52 52 46
4 −5.0 < η < −1.6 0.0 < φ < 2π 68 52 52 46
be made following additional performance studies and in consultation with the Calorimeter and L1Calo
communities.
Four configurations of fibers and FPGA as a function of link speed are presented in Table 7. It
is interesting to note that by changing from the 80 MGT Virtex R⃝-7 [27] to the 120 MGT Virtex R⃝UltraScaleTM [28], it is possible to reduce the number of Processor FPGA from four to two at 9.6 Gb/s390
or higher link speed while retaining spare capacity for future calorimeter upgrades.
3.4 Real-Time Output Data
The real-time output from the gFEX are TOBs which contain information from the feature identification
algorithms running on the Processor FPGA (Sec. 2.4). There are two category of TOBs: global and
jet. The global TOBs encode data associated with global event observables such as the pileup density ρ,395
EmissT
, HmissT
, etc. The jet TOBs contain information about the identified jets and any applicable moments.
The TOBs are transmitted to L1Topo on optical fibers assuming a baseline bandwidth of 6.4 Gb/s [29].
It is foreseen that the bandwidth could be increased up to 12.8 Gb/s for the Phase-I upgrade. It is as-
sumed that each Processor FPGA will have one or more independent fibers to each L1Topo FPGA. The
protocol and payload for transmission to L1Topo is the same as the receipt of data from the calorimeters400
as described in Sec. 3.1.
gFEX: number of optical fiber connections
gFEX References
https://twiki.cern.ch/twiki/bin/view/AtlasPublic/JetTriggerPublicResults#Global_Feature_Extraction_gFEX_P
Performance Plots (Simulation)
https://edms.cern.ch/file/1425502/1/gFEX.pdf
gFEX Prototype Technical Specification