Upgrading the CDF Track Trigger to 3-D for High Instantaneous Luminosity Ben Kilminster Ohio State University CDF Collaboration L = 1x1032 cm-2s-1 Last week ! Project includes physicists and engineers from: Ohio State, Baylor, Fermilab, Illinois, Purdue, UC Davis Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.1 CDF Central Outer Tracker (COT) 8 “superlayers” of cells 4 with axial wires: r - f measurement 4 with stereo wires: z measurement Each cell Ben Kilminster 0.88 cm drift (avg.) Max drift time ~220 ns 12 sense wires/cell: 96 measurements 2540 cells, 30240 channels Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.2 eXtremely Fast Tracker = Level 1 Track Trigger CDF Trigger System Role of tracking Top, W/Z, Exotic Physics triggers require High momentum electron and muon Level 1 trigger candidates Bottom Physics require low momentum tracking at the Level 1 trigger electrons muons hadronic tracks Detector Elements CAL COT MUON XFT SVX Muon Prim CES XCES XTRP CAL Track Muon L1 Trigger Primitives Electrons: XFT track + EM cluster Muons: XFT track + muon stub L2 Trigger Tracks XFT Track + Silicon Hits Global Level 1 SVT CAL Global Level 2 Ben Kilminster Cornell JC : CDF XFT Upgrade TSI/Clock 11 Feb 2005; p.3 Overview of Existing XFT Hit Finding: Mezzanine Card Hits are classified as prompt or delayed (i.e. “2-bin”) Good hit patterns are identified as segment, then segments are linked as tracks Segment Finding In the axial layers patterns of wires within one superlayer Track Finding Looking across 3 or 4 axial layers, search for patterns of segments consistent with Pt>1.5 GeV/c Tracks only found in 2-Dimensions (r,f) Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.4 The Hit and Segment Finders Track segments are found by comparing hit patterns in a given layer to a list of valid patterns or “masks”. Hits : 2 bins - Prompt or delayed Mask : A specific pattern of prompt and delayed hits on the 12 wires of an axial COT layer generated by simulating tracks Pixel: represents the phi position of the track at the midpoint of the cell. “Prompt” hit Layer Cells Masks Pixels 1 192 2304 166 2 288 3456 227 3 384 2304 292 4 480 2880 345 “Delayed” hit Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.5 Segment Finder Output In the inner two layers, each mask corresponds to 1 of 12 pixel positions in the middle of the layer. The pixel represents the phi position of the track. In the outer 2 layers, each mask corresponds to 1 of 6 pixel positions and 1 of 3 slopes: (low pt +, low pt -, high pt). When a mask is located, the corresponding pixel is turned on. Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.6 The Segment Linker Tracks are found by comparing fired pixels in all 4 layers to a list of valid pixel patterns or “roads”. • Chamber is divided into 288 1.25–degree “identical” Linkers •Highest track reported for each linker •-> Max of 288 tracks per event • Each linker uses a look-up table of ~1200 roads Pixels must match Slopes must match Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.7 XFT System Electronics Mezzanine Cards 168 cards Classifies hits as prompt/delayed Final Finder system 24 SL1-3 boards 24 SL2-4 boards Heavy reliance on PLDs Allows for some redesign: new patterns for number of misses, wire sag, faster gas, etc Final Linker System 24 Linker boards Heavy reliance on PLDs Allows for new road set based on new beam positions Have already developed 2 new roads sets due to accelerator changes. Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.8 XFT Performance in CDF RunII Performance of the XFT in CDF’s RunII has been excellent 1. Momentum resolution 1.74%/GeV/c 2. Phi Resolution < 6mRad 3. Efficiency ~ 95% (almost 100% for high Pt tracks) Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.9 Why an Upgrade? The XFT was designed for a luminosity of: L=1x1032cm-2s-1 with 396 nsec bunch crossings <int/crossing> ~ 3 Then it was supposed to switch to : L=2x1032cm-2s-1 132nsec bunch <int/crossing> ~ 2 But 132ns was not feasible Accelerator Performance Max luminosity attained: 1x1032cm-2s-1 Expect to reach L=3x1032cm-2s-1 at 396nsec bunch crossing <int/crossing> ~ 9 Factor of 3-4 above design Ben Kilminster Design @ 396 ns Design @ 132 ns Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.10 Z ee at low lum. 0 add. Int./crossing Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.11 Z ee at low lum. 2 add. Int./crossing Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.12 Z ee at low lum. 5 add. Int./crossing Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.13 Z ee at low lum. 10 add. Int./crossing Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.14 Z ee at low lum. 10 add. Int./crossing Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.15 Two-Track Triggers Two Track Trigger for B Physics 2 tracks pT>2.5 GeV Opposite charge pT(1)+ pT(2)>6.5 GeV df < 135 Run II Data 160mb Quadratic growth (overlaps + fakes) s(L=5E31)/s(L =0)=1.5 Extrapolate: 100mb linear s(L=1.5E32) = 225mb 34kHz Real (from overlapped MB) s(L=1.5E32) = 500mb 75kHz We are stuck with FIXED Bandwidth (Accept Rate): L1: 30 kHz Ben Kilminster Cornell JC : CDF XFT Upgrade L2: 1 kHz L3: 100 Hz 11 Feb 2005; p.16 What’s needed: Primary problem is the growth of fake tracks due to pattern recognition problems: fake Fake tracks can be made from pieces of different real physical tracks. Need a way to model the highest luminosity running Simulations of the XFT System Study the effect of high luminosity running on a variety of triggers Two-Track Trigger (low PT) Single Track Trigger (high PT: Track Only) Electron Trigger (Track + Other object) Trigger Rate is the Consider modifications/additions to the current system “Figure of Merit” Reduce fake rate Maintain resolution Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.17 XFT Simulation of High Lum. All events are passed through a hit-level simulation Start with COT hits Gives exactly the same answer as hardware when run with same masks, roads and XFT hits Outputs XFT hits, pixels, and tracks for axial and XFT pixels for stereo Simulate High luminosity by Merging events “main” event with zero bias Merge COT hits (combine overlapping hits) simulate XFT with various options Add tracks from individual events together Offline tracks serve as “truth” for the event This method allows us to probe up to 4E32 Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.18 Upgrade Strategy Studied various upgrade paths Replacing Axial System (Use additional 264ns between crossings) Add information from Stereo Layers of COT Replace linker with finer segment linking Doing some of the above We picked this. Advantages of Stereo Upgrade Gives us the fake track rejection we need. Allows for more additional handles in trigger Track pointing to muon stubs & calor (L2) Possibility of Two-Track Invariant Mass Cuts (Understudy) Parallel Path to Axial System Commission parasitically. Only Minor changes to Axial System Stereo layers Current XFT uses 4 axial layers only Slight Changes to existing firmware. KEEP A WORKING SYSTEM No extended downtime Ben Kilminster Cornell JC : CDF XFT Upgrade Upgrade adds 3 stereo layers. Use full 396ns between crossings to send more precise hit information. 11 Feb 2005; p.19 Impact of Stereo The stereo can have an impact in two ways: Confirmation Segment: Since often fake XFT tracks are the result of linking two unrelated low Pt segments, requiring another high Pt stereo segment in the allowed window around an axial track can be very powerful. one stereo layer rejection We will use this at Level 1 Provide Z-pointing to tracks: Since EM and muon calorimeters are segmented in Z, coarse pointing can be very helpful in eliminating fakes We can use this effectively at Level 2 Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.20 Stereo association studies Pixel displacement is a measure of eta of track Expected pixel position (z = 0) Measured pixel position (z 0) SL7 pixel (SL7) Displacement from stereo angle SL6 Reals pixel (SL5) SL5 SL5 has opposite displacement from SL7 SL4 Measured pixel position (z 0) SL3 Fakes pixel (SL3) Displacement from stereo angle Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.21 Overview of Upgrade Hit Stage Provide 6 times bins per wire per bunch crossing instead of the present 2 for stereo layers maintain existing 2-bin system for axial layers Pass stereo hit information to Finders via optical cable maintain existing Ansley cables for axial layers Segment Finding Stage Using 6 times bins, measure phi (pixel) position and slope of segments in 3 outer stereo layers maintain existing axial segment finding system Axial Segment Linking stage maintain existing axial track finding system Add new Stereo Linker Association Stage Associate existing axial tracks from the axial system with stereo segments to reduce number of fake track Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.22 Current XFT Configuration Ansley trigger cable (220 ft) Data @45MHz LVDS 168 TDC from COT axial layers ~2 m copper Cable Data @33MHz (channel link) 24+24 Axial Finders 24 Linkers 3 crates 3 crates XTC 24 crates Ben Kilminster Cornell JC : CDF XFT Upgrade Neighboring cards connected over backplane 24 Linker Output Modules ~10 m of cable to XTRP 11 Feb 2005; p.23 XFT Upgrade Configuration Ansley trigger cable (220 ft) Data @45MHz LVDS ~2 m copper Cable Data @33MHz (channel link) 168 TDC from COT axial layers 24+24 Axial Finders 24 Linkers 3 crates 3 crates Neighboring cards connected over backplane 2 bin XTC 24 crates 126 TDC from COT stereo layers 6 bin XTC 2 Ben Kilminster New cable (~150ft) Optical Data ~45MHz 12+12+12 Stereo Finders ~3m optical Cable @60.6MHz 2 crates Cornell JC : CDF XFT Upgrade 24 SLAMs ~10 m of cable to XTRP SLAMs replace Linker Output Modules Data to L2 11 Feb 2005; p.24 Timing is Tight ! Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.25 Getting TDC data to XFT TDC Rx Tx XTC2 XTC XTC XTC ~150 ft Optical fiber COT data split into two paths: trigger and data Trigger info is put into 6 time bins by the new XTC2 mezzanine card Data driven to Stereo Finders by optical fiber link from optical transmitter card to optical receiver card XTC 2 : Measured start/stop time for each window Consistent with 3ns timing resolution Ben Kilminster XFT Stereo Finder Rx Rx card on Finder: Receives optical data Tests show good performance Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.26 Improving the Hit Binning Algorithm Change binning of hits 6 Bins Maximized time window intervals to avoid some bins being under-utilized Better fake rejection than trivial binning: 20% “Not sure” window Most hits fire two time bins <pulse width>~40ns Bin width: 24ns If hit exists in previous bin, ignore hits at beginning of bin (“not sure window”) Both improvements give an additional ASDQ ~40% fake rejection compared to no Threshold “not-sure” window 24ns 18ns Time Not sure Window Bin N Ben Kilminster Cornell JC : CDF XFT Upgrade Bin N+1 11 Feb 2005; p.27 Improving Segment Finding Patterns Segment Finder Chips 2 Time Bins, masks 6 Time Bins, Masks Finder Axial SL1 166 1344 Finder Axial SL2 227 1844 Finder Axial SL3 292 2056 Finder Axial SL4 345 2207 New Finder Chips Driving factor in chip choice 7 times more “masks” == unique segment patterns from track simulation Segment Finder Chips Implementation Logic Elements Used Speed Stereo Layer 4 6 bins Flex 10K50 2,500 / 2,880 87% used 16.5 ns Axial Layer 4 2 bins Stratix 2S60 10,602 / 48,352 22% used 33 ns Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.28 A look at the SLAM Board (OSU responsibility) Design Storage VME Interface (Control Code, and State machine interface) Optical IO Clock Gen/Dist Output Stereoconfirmed Tracks Linker Input SLAM Chip Tracks Stereo Finder Segments Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.29 Impact on A Specific Trigger Scenario C Two-Track Trigger Lumi [1E32 cm-2s-1] 0.5 1.0 1.5 2.0 3.0 2-Bin s [mb] 0.12 0.28 0.50 0.78 1.5 Using 6-Bin Stereo s [mb] 0.08 0.13 0.21 0.33 0.65 Ratio 0.37 0.38 0.39 0.40 0.42 Uses only 2 of 3 Stereo Layers 3x1032 cm-2s-1 Luminosity Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.30 Conclusions Accelerator performance has been excellent But…high luminosity at 396nsec bunch spacing leads to many interactions/crossing We need to upgrade the XFT to take advantage of the great opportunity The XFT Upgrade will meet the needs of high L running This upgrade gives us the required factor of 3 rejection of fakes System can be installed and commissioned with little impact on the current XFT Not all capabilities have been explored Current rejection only making use of 2 of 3 stereo layers. Expect another factor of ~2 by using stereo extrapolation in Level 2 Mass triggers are also possible at Level 1 and/or Level 2 Schedule: Prototypes under design. Make extensive use of experience from existing system. Expect completed system in ~15 months System Operational by Jan 2006. Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.31 BACKUPS Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.32 XFT Data Path TDC Transition module Timing and multiplexing Fiber optic transmit board Use on: TDC TM, SLAM interface Fiber optic receiver board Use on: Finder, L2 Pulsar Optical data Prototypes built and tested Look good. Investigating higher data rates Production runs will await vertical slice test. Ben Kilminster Electrical data Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.33 Fibers Used in the XFT Upgrade Fibers from XTC to Finder 200ft + 2ft breakout each end Total installation: 38 bundles of 4 fibers + 36 bundles of 6 fibers Pulling fibers expected to take ~7 days Will need fibers pulled for commissioning transition boards Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.34 XFT Stereo Finder Board Fiber 1 Fiber 2 RX Mezzanine Fiber 3 1 18 2 3 18 18 4 Fiber 4 Finder A FPGA Download VMEbus Slave & Control (8 cell) FPGA Download 18 Finder B 5 Fiber 5 18 FPGA Download Fiber 6 RX Mezzanine Fiber 7 Fiber 8 6 7 8 9 Fiber 9 Finder C 18 18 18 FPGA Download Finder D 18 5 x 32 FPGA Download Fiber 10 RX Mezzanine Fiber 11 10 11 Fiber 12 Pulsar Driver Finder E 18 18 FPGA Download 1 x 18 FPGA Download 10 x 15 TX Mezzanine 4 x 18 Ben Kilminster Pixel Driver Pixel and Slope Output to the L2 Pixel and Slope Output to the SLAM Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.35 Firmware Progress The Stereo Finder and SLAM designs make extensive use of reprogrammable devices. Major progress in firmware over the last 6 months Functional designs have been implemented Timing has been studied Quartus II Version 4.1 Build 181 06/29/2004 SJ Full Version Ben Kilminster Stereo Chip Compilation ; Family ; Stratix II ; Device ; EP2S60F484C3 ; Timing Models ; Preliminary ; Total ALUTs ; 10,602 / 48,352 ( 21 % ) ; Total pins ; 150 / 335 ( 44 % ) ; Total memory bits ; 33,088 / 2,544,192 ( 1 % ) ; DSP block ; 0 / 288 ( 0 % ) ; Total PLLs ;0/6(0%) ; Total DLLs ;0/2(0%) SLAM Chip Compilation ; Family ; Stratix ; Device ; EP1S40F1020C5 ; Timing Models ; Production ; Total logic elements ; 25,081 / 41,250 ( 60 % ) ; Total pins ; 341 / 782 ( 43 % ) ; Total memory bits ; 6,912 / 3,423,744 ( < 1 % ) ; DSP block ; 0 / 112 ( 0 % ) ; Total PLLs ; 1 / 12 ( 8 % ) ; Total DLLs ;0/2(0%) Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.36 The Axial Finder Chip Pixel Output Mask finding 140 inputs Dead COT wire list L1 and L2 storage Axial Finder: implemented using Altera FLEX 10K70 chip. Stereo Finder: Altera Stratix EP2S60 chip Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.37 Stereo Finder Chip Finder CHIP BLOCK 1 Clock Ticks= 16.5ns T=5 Clock Ticks T=24 Clock Ticks Cell 3 Wire Data 6 time Bins A Cell 2 Cell 1 Cell 0 A 16 bits B C 16 bits 16 bits Input Alingment 3-FIFOS 16 bits B C 16 bits 16 bits Wire Data Storage Registers Cell Cell Cell Cell 8 to 1 Multiplexer 428 427 426 425 12 Pixels to Pixel Chip Pixel MASK 133 bits Set 72 bits Dead Wire Reg. L2 Buffers(4) Pulsar MASK 133 bits Set Aligns data from 3 TDC modules wrt 16.5 ns clock T=3 Clock Ticks VME 96 Bit FIFO 96 bits 8 slices deep 96 bits T=2 Clock Ticks 32 Pixels to Pulsar Chip 3 to 1 mux T=1 Clock Ticks + 23 Clock Ticks to process all 8 cells To register 6 time bins for 12 cells, requires 24 time slices Ben Kilminster + 7 Clock Ticks to process all 8 cells 72 bits Cell 431 Cell 430 Cell 429 Cell 424 VME concentration on path to SLAM T=3 Clock Ticks T=1 Clock Ticks Data multiplexed to provide 133 bits of information for segment finding in one core cell Cornell JC : CDF XFT Upgrade For each of 8 core cells, finder algorithm is run producing 12 pixels of phi and/or slope output to SLAM module 11 Feb 2005; p.38 Pixel Driver Chip Pixel Driver BLOCK T=2 Clock Ticks Data Vaild from Finder Chips Pixels from 5 Finder Chips sent through two paths to SLAM 5 + T=1 Clock Ticks Channel A & B Control Controller - State Machine A B C 12 bits 12 bits 3-FIFOS A B C 12 bits From Finder Chips C D E 3 to 1 12 bits Multiplexer 12 bits 12 bits 12 bits 12 bits Channel A to SLAM 12 bits 3-FIFOS 12 bits A B C Channel B to SLAM 12 bits 12 bits 3 to 1 Multiplexer 12 bits 12 bits Total time = 2 clock ticks for FIFO(in and out) and 18 clock ticks to send the 18 Slices of Pixel Data Pixel Data from 3 Finder Chips (18 core cells) are accumulated in each FIFO Ben Kilminster Cornell JC : CDF XFT Upgrade Pixel Data sent to SLAM in two 15° slices containing 18 core cells of pixels for association to axial tracks 11 Feb 2005; p.39 L2 Output Chip demonstration of something works - can be optimized Pulsar CHIPwhich BLOCK T=2 Clock Ticks L1 Accept From FINDER CHIPs A B C D E + T=1 Clock Ticks + T=1 Clock Ticks Control Controller - State Machine 32 bits 32 bits 32 bits 32 bits 32 bits 5-FIFOS A B C D E To Backplane & Auxillary Mezzzanine Board 32 bits 32 bits 5 to 1 32 bits Multiplexer 16 bits 2 to 1 Multiplexer 32 bits 32 bits 16 bits 16 bits 32 bits Total time to send data on a L1 Accept = (96 bits/cell * 8 cells/Finder * 4.5 Finders / 16 bits) * 16.5ns = 3.564us • 32 bits of Pixel data is stored as a slice in the FIFOs, 3 slices per Cell, 8 Cells from each Finder. • On L1 Accept, FIFO outputs to multiplexers, then sent to PULSAR board. Otherwise, FIFO slice overwritten. Ben Kilminster 96 bits per cell * 8 cells per Finder Chip * 4.5 Finder Chips = 3,456 bits per Finder SL7 board 16 bits every 16.5ns, so it will take --> 3,456/16 * 16.5ns = 3.564us Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.40 SLAM Chip 3layers by 3 cables Stereo pixels @16nsec Stereo pixels @16.5nsec Single Linker L3 pixels All L3 pixels Cable Reader t 3 Latch Pixels t 1 All L5 pixels All L7 pixels Remap Pixels t 1 Single Linker L5 pixels Single Linker L7 pixels 4 Phi bits by 48 Pt Bits Array Process PtByPhi Array t 0.5 SubRoad Finder t 0.5 Loop over 12 linkers tBen 12 Kilminster Stereo Confirmation Bit @8.25nsec Linkers 0-5 to XTRP @33nsec Reformat Linker t 3 Cornell JC : CDF XFT Upgrade Linkers 6-11 to XTRP @33nsec 11 Feb 2005; p.41 Firmware Progress The Stereo Finder and SLAM designs make extensive use of reprogrammable devices. Major progress in firmware over the last 6 months Functional designs have been implemented Timing has been studied Quartus II Version 4.1 Build 181 06/29/2004 SJ Full Version Ben Kilminster Stereo Chip Compilation ; Family ; Stratix II ; Device ; EP2S60F484C3 ; Timing Models ; Preliminary ; Total ALUTs ; 10,602 / 48,352 ( 21 % ) ; Total pins ; 150 / 335 ( 44 % ) ; Total memory bits ; 33,088 / 2,544,192 ( 1 % ) ; DSP block ; 0 / 288 ( 0 % ) ; Total PLLs ;0/6(0%) ; Total DLLs ;0/2(0%) SLAM Chip Compilation ; Family ; Stratix ; Device ; EP1S40F1020C5 ; Timing Models ; Production ; Total logic elements ; 25,081 / 41,250 ( 60 % ) ; Total pins ; 341 / 782 ( 43 % ) ; Total memory bits ; 6,912 / 3,423,744 ( < 1 % ) ; DSP block ; 0 / 112 ( 0 % ) ; Total PLLs ; 1 / 12 ( 8 % ) ; Total DLLs ;0/2(0%) Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.42 Finder Timing 61 (16.5 ns) clock ticks = 1007 ns from when first TDC data arrives to when output of receiver on SLAM has 18 cells Stereo Finder Algorithm Finder Algorithm Track Segment Similar to axial XFT 6 time bins input (72 bits per 12-wire cell) vs 2 time bins(24 bits) 8 12-wire cells vs 4 cells per FPGA 16.5ns clock vs 33ns clock Much larger Mask set L2 Pulsar Data(96 bits output) Implement in newest ALTERA Stratix 2 FPGA +Neighbor Ansley Cable Core Ansley Cable - Neighbor Ansley Cable Gray Wires indicate 12 wire mask for the depicted track segment Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.44 Segment Finding Algorithm Check 4-wire patterns (generated by simulating tracks) in top, middle, and bottom 4 wires SL Check all allowed combinations of the three 4-wirepatterns Output 1,2, or 3 misses Found pixel phi & slope ! • For each SL (3,5,7), there are 3 designs for each allowed number of misses (1,2,or 3) • 2 designs can be loaded: one of the above plus a testing algorithm which allows us to input test vectors and verify output through rest of the system Ben Kilminster Cornell JC : CDF XFT Upgrade 11 Feb 2005; p.45