Title: L2eta: Current Focus
1L2ßeta Current Focus
- Last Fridays meeting at UMD
- Bob Hirosky, Drew Baden, Rob Bard, John Giganti
- Discussed details of various SBC and I/O card
options - Re-iterated desire for simple, off-the-shelf
re-implementation - More detailed discussions of hardware
requirements - Discussion of ß-2-ß PIO reduction for
multiprocessor models - AGP-style solutions set aside - leverage
existing PCI experience for minimum development
time - Prime contenders include
- SBCs single or dual PIIIs dual or
quad G4s - I/O (modified) s-link PMC(s) Catalina
64bit PMC - or similar (Transtech)
(vs redesign)
2Summary of Hardware findings
CPU boards SINGLE-gtQUAD Power PC
Variety SINGLE-gtDUAL PIII Variety I/O
CARDS PMC CARDS w/ Internal FIFO/MEMORY Catalin
a (64 bits wide at 33MHz) S-link (32 bits wids
at 33 MHz) Transtech (64 bits wide at 66 MHz)
NEW FIND!
3Single/Dual/Quad PPC 750 450 MHZ PPC 7400
to come
DUAL/QUAD Power PC Option
30 faster than alpha in INTs about the same in fp
4- Power PC option comments cost
8K(dual)/14K(quad) - Pro
- Quad CPU board can replace a whole crate, no need
to alpha-alpha PIO - use shared memory instead - 2 PMC slots are on completely separate busses -
limits bus latency - 64bits/66MHz - higher bandwidth for DMA
- 32bits/33MHz - lower bandwidth for other things
- Con
- Longer startup curve due to switch to PowerPC
linux? - Compiler options under Linux? Linux compatible
images? - Single vendor for this QUAD board, several for
single/duals
5A Dual Pentium III Option
Dual 1GHz PIII
specint95 46.6
vs alpha 15.4
6- PIII option comments
- Pro
- Huge amounts of CPU now on the shelf
- More familiar Linux
- KAI compiler under Linux
- Con
- PCI bus is only 32bits/33MHz - would have to live
w/ estimated 65MB/s Max bandwidth (vs 80 ish now) - No QUAD option yet - cant replace a whole crate
- Maybe single vendor for the dual PIII board, many
for singles
7Catalina QC64-PMC
8- Catalina PMC card comments cost 4.5K
- Pro
- 64 bits wide w/ user programmability in onboard
FPGA - FPGA programmable from PCI bus!
- Lots of FIFO space on board, we could do w/o
FIFOs on our adapter card - Should be able to get LINUX drivers
- Con
- Runs only at PCI 33MHz, same as we have now
- Its a one way card, need another PMC card for
output
9CERN S-link to PMC, PMC to S-link
Input Card
Output Card
10- S-link PMC card comments
- Pro
- Lots of support from CERN/LHC
- We could use this as a base design if we later
build PMC cards ourselves - Con
- Runs only at PCI 33MHz / 32 bits
- way cards, use two of them or modify an existing
design
11Transtech-DSP PMC-FPGA I/O card
12- Transtech PMC card comments cost 2.5-5K,
depending on FPGA - Pro
- 64 bits wide w/ user programmability in onboard
FPGA - 66 MHZ PCI interface
- Its a two-way card
- Choice of onboard FPGA
- Lots of fast on board SRAM (can it replace
fifos) - Promise of LINUX drivers
- This company is working w/ DNA computers
- Con
- FPGA programming might be a bit more difficult if
we use the onboard memory to implement or FIFOs - Not delivered yet (but promise several week time
scale - this board is a modification of an
existing DSP board, not made from scratch)
Newest find - need to do more research on this
one, but it looks promising
13SYSTEM OPTIONS
QUAD PowerPC Catalina card output PMC QUAD
PowerPC Transtech card ?Fastest I/O ?option
of removing alpha-alpha PIO ?separate PCI
buses, may make this a non-issue ?
Linux/Compiler issues ? Less certain about
future enhancements to SBC PIII S-link
Card ? Compilers / Familiar Linux ? Maximum
symmetry w/ existing system ? 18-25 PCI BW hit,
65MB/s max (MBdma needs 20MB/s)
Continuing work on draft TDR for mid Feb...