Overview of the ATLAS Fast Tracker (FTK)hep.uchicago.edu/cdf/shochet/ftk_talks/FTK_overview_07...Overview of the ATLAS Fast Tracker (FTK) (daughter of the very successful CDF SVT)

July 24, 2008 M. Shochet 1

Overview of the ATLAS Fast Tracker (FTK)(daughter of the very successful CDF SVT)


What is it for?• At the LHC design accelerator intensity:

– New phenomena: ∼ 0.05 Hz– Total interaction rate: ∼ 1 GHz (40 MHz beam crossings)

• Many possible new phenomena produce heavy b quarkswhich can only be distinguished from the bulk of the background by reconstructing the individual tracks.

• We are proposing to significantly enhance the ability of ATLAS to rapidly identify b quarks in the trigger.– Currently done in commodity PC’s. This is slow and becomes

slower as the accelerator intensity and thus track density increase.

The problem!

beam pipe few mm


ATLAS


Inner Tracking Detectors

Pixels barrel SCT barrel Pixels disks

3 pixel layers (space point)

8 strip layers (1 coordinate)

⇒ 11 layers, 14 coordinates


Getting data into FTKRODs

SCT →

Pixels →

ROBs

ROBs

FTKon L1 accept

silicon hits silicon tracks

Level 2

ask for ROI’s

dual-output HOLA designed by Tang


Number of input fibers• Pixels: 120• Strips: 92

Number of crates• ∼ 12


How does it work?

• First do pattern recognition, then fit possible candidates.

•Prestore patterns (roads) in large content-addressable memory.

The Pattern Bank (4-layer)

coarse resolution hits full resolution hits(superbins)


Massively parallel pattern recognition

AM = BINGO PLAYERS

HIT # 1447

PATTERN NPATTERN 1PATTERN 2

PATTERN 3

PATTERN 5

PATTERN 4

track fitter

1 superbin per silicon layer

Majority logic allows up to 1 missed layer.


Functional layout

Track dataROBIN

Track dataROBIN

Raw dataROBINs

~Offline quality Track parameters

SUPER BINSDATA

ORGANIZER ROADS

ROADS + HITS

EVENT # N

PIPELINED AM

HITS(LVDS links)

DO-board

EVENT # 1

AM

-boardPixels & SCT

DataFormatter

(DF)

50~100 KHzevent rate

RODsRODs

cluster findingsplit by layer

overlap regions

S-links

Track Fitter


Possible layout for a core crate (after DFs)

AM

-B7

AM

-B8

AM

-B1

AM

-B0

DO

5D

O4

DO

3D

O2

DO

1D

O0

Trac

k Fi

tter

AM

-B2

AM

-B3

AM

-B4

AM

-B5

AM

-B6

AM

-B9

AM

-B10

AM

-B11

AM

-B12


Data Formatter

• Receives raw hits from the detector (RODs)• Finds hit clusters

pixels silicon strips

• Store cluster centroids• Separates clusters by silicon layer & sends to Data

Organizers on 6 LVDS data busses (22 bits each)


Data Organizer• Receives hits from Data Formatters.• Stores hits at full resolution in a way that is rapidly

accessible by pattern number.• Sends hits at coarser resolution (superbin) to pattern

recognition unit (Associative Memory).• Receives patterns from AM, retrieves full resolution hits,

and sends road number and hits to the Track Fitter.


Track Fitter• Receives road # and associated hits from Data Organizers.• Computes all hit combinations

• Calculate the track parameters curvature, azimuthal angle, polar angle, z0, impact parameter

and the goodness of fit (χ2) using a linear correction to the mean for that sector (excellent precision over a sector).

sector: a physical silicon module in each layerpixels: 1" x 2.5 " strips: 2.5" x 5 "


14 measurements, 5 parameters ⇒ 9 constraints (→χ2)

Pi: 5 track parameters & 9 constraints (χ2 is sum of squares)xj: the hit coordinate in layer jaij, bi: the stored constants for each sector, calculated in

advance from a large sample of training tracks (simulation or data)

• Cut on goodness of fit; among the combinations in a road, select the track with the best χ2.

• Output good tracks to ROBIN.

14

1i ij j i

jp a x b

=

= +∑


Readout Buffer (ROBIN)• Stores tracks for access by the level-2 trigger PCs.


Track Fitter details• GigaFitter – a simplified version built for the CDF SVT

– 2D reconstruction (transverse to the beamline)– 6 detector layers– 3 track parameters (curvature, azimuthal angle, impact parameter)– 3 constraints → χ2


GigaFitter scheme

• DO roads & hits → input FIFO → RAMs according to detector layer• Combinations (one hit/layer) calculated sequentially & stored in Comb-FIFO• Each combination & the constants are sent to DSP for fitting & selection

Lay0-Ram

Lay1-Ram

Lay2-Ram

Lay3-Ram

...

Com

b -FiFo

INPUT FiFo

Constants RAM

DSP:Fit Tracks

Choose best χ2 track

Lay10-Ram


DSP algorithm for SVT

• Comb-FIFO data serialized: 1 hit and its constants sent to parallel DSP slices each computing a track parameter or constraint.

• An additional DSP computes total χ2 from individual constraints• Total of 7 DSP slices in parallel each working at 200 MHz

C4

C3

18

18

18

18

C2

ACC 39

FIFO156

ACC 39

ACC 39

ACC 39

CTRLEVRST

RSTREADY

C1

DSP48E

39

39

39

39

18

constants

Hit


DSP Slice

• One DSP slice computes a track parameter in 14 clock cycles, plus 4 for the first one.

18

18

39


Xilinx XC5VSX95T• 14720 V5 slices (4 LUT + 2 FF)

– ∼1% used for each fitter

• 1520 kbit of distributed RAM– ∼1% used for each fitter

• 640 DSPs– ∼2.5% used for each fitter in the FTK version

5 parameters, 9 constraints, 1 to calculate overall χ2

– ⇒ ∼ 40 fitters/chip (remember: many combinations per road)

• 1 Mbyte of block RAM– plenty for the SVT prototype– not even close for the FTK!!


Missing hit problem• To obtain high reconstruction efficiency, we must allow one

physical detector layer to miss a hit.• The constants in the parameter and constraint equations are

different when there is a missing hit.• Store 12 sets of constants (all hits, a miss in one of 11 layers).

– A lot of memory that has to be accessed very quickly. Can one look ahead for the constant set that will be needed next?

• Alternative is to estimate the hit location in the missing layer.– How long does it take?


Size of the constant memory• 210 words/constant set (14×14 + 14)• If we need 2-byte precision ⇒ 420 bytes/constant set• One constant set/sector. Currently estimate 100k sectors

in an FTK crate.⇒ 42 Mbytes of fast memory⇒ 0.5 Gbyte if solve missing hit problem with more memory

• We have heard that with the latest FPGAs, there is very fast access to external computer-like memory. Is that true?


How many fitters are needed?• The number of cycles from the time the data from a road is

in the input FIFO until the track parameters are in the output FIFO is approximately:

NfitCycl = Ncomb × Nhits (all 14 calculations done in parallel)

• A road packet takes Nhits + 1 Data Organizer clock cycles to be sent to the Track Fitter. Thus if we are to have 0 deadtime from track fitting, we need the number of parallel fitters (there are 40 in a chip), Nfitters, satisfying:

• This translates into a maximum average Ncomb of Nfitters × ClockRatio × (Nhits + 1)/Nhits

where ClockRatio is the ratio of the fitter to DO clock speeds.• We will have to satisfy this: road width, # of FPGAs.

Nhits 1 NfitCycl 1DO freq fitter freq Nfitters

+= ×


Hit Warrior function• One can easily get many tracks (ghosts) from a single real

particle due to presence of extra random silicon hits.

• Compare a new track with those already found. If new one has 8 or more hits in common with a stored track, keep only the best χ2 track.

• If space permits, add this function to the Track Fitter.


Spy Buffer• Data flows very quickly through this system. By the time

any PC using its output detects a problem, the data would already be long gone from the FTK. That makes diagnosing a problem that is occurring internally in the FTK extremely difficult.

• We found it very useful in the SVT to have deep buffers at the input and output of every board in the system. Then, when a problem is detected, these spy buffers can be frozen in the entire system and read out. The buffer is deep enough so the event with the error is still inside it.

• This allows diagnosing and fixing subtle problems.

Documents

Overview of the ATLAS Fast Tracker (FTK)hep.uchicago.edu/cdf/shochet/ftk_talks/FTK_overview_07...Overview of the ATLAS Fast Tracker (FTK) (daughter of the very successful CDF SVT)