M. van Veelen

1
Failing scaling and Roadmapping to new
Architectures

• Can we use the existing architectures with new
  and future technology for the next generations of
  processing machines in Radio Astronomy?

Martijn van Veelen Jaap Bregman
2
(No Transcript)
3
Spatial multiplexing (partition antennas)
FAILING SCALING
4
Spectral multiplexing
Transpose
FAILING SCALING
5
Scaling chip design principles

• It is hardly possible to synchronize the
  processing. Different areas on chip operate on
  their own decoupled clock.
• In order to achieve low power/cm2 we no longer
  push to run a maximum chip clock

6
Coherent synchronous processing?
FAILING SCALING
7
The emerging router
FAILING SCALING
8
Beamforming and two-step DFT
FAILING SCALING
9
X
F
FAILING SCALING
log(computations)
break-even point OFX OXF
log(antennas)
WSRT
ALMA
LOFAR
SKA
10
Required FX operations
FAILING SCALING
32bit
3
2
70-700MW
?
?
11
Trends in the correlator architecture

• Spatial correlator function no longer determines
  the geometrical layout
• The technology (chip-speed) no longer determined
  the geometrical layout
• Coherent correlation is no longer implemented by
  synchronouns processing
• The connectivity between correlator and
  calibration is a priority driver rather than the
  input of the correlator.
• Function and technology are no longer determining
  the architecture
• The power consumption / cooling and the
  correlator output connectivity drive the
  architectural optimizations.

12
Function ? Form
technology
?
?
?
function
technology
architecture
function
13
Function ? Form
technology
Spatial nulling
Spectral blanking
?
?
?
function
FX
XF
Round robin
XF
FPGA
CPU
14
(No Transcript)
15
Time-To-Market
16
(No Transcript)
17
(No Transcript)
18
NRE
technology
HW platform
SW platform
Processing software
function
Application Specific
lifecycle flexibility
reliability
COTS HW
maintenance
system costs (manufacturing)
Power consumption
TCO
19
What is effective RD?

• The development paths for existing correlators
  based on Hero-approach vs. the just-in-time
  integration approach.

20
The Hero-approach
21
The just-in-time integration approach
22
RD
Processing Software
Manufacturer Effort (Multi-Customer)
Operating System
COTS Correlator Platform
Optimal RD effort (time x man-power)
HW-centric Control Software
RD effort
Resource Optimized Platform
RD effort
Aging Technology
23
SKA ?
LOFAR
System Costs
WSRT
Development Time
24
Roadmap architecture
Function
Architecture
Technology
time
25
Roadmap
sensitivity
power consumption
availability
surveyspeed
serendipity
Operating Modes
maintenance
F
X
flexibility
Functions
XF
FXF
FX
FFX
?
synchronous
asynchronous
round-robin scheduling
?
Architecture
4D torus
Ring network
fanout
Crossbar
Gb routers
Infinibandswitches
AdvancedTCA
Multi-core microproc
ASIC
gate arrays
FPGAs
Multi-core FPGA
Micro proc
?
Emb.Mult. FPGA
Technology
130nm
90nm
? nm
time
26
Conclusions

• Architectures emerge but not in a predictable
  manner.
• Technology advances stepwise rather than
  continuously, it cant be predicted on a
  timescale of half a decade.
• Given the emerging optimization criteria related
  to TCO, a one-to-one scaling of existing
  architectures will fail.
• The technology advances with Moores Law are not
  sufficient to provide a suitable architecture for
  SKAs central processor.
• We can benefit from platform-specific RD
  invested by industrial machine manufacturers, if
  we pursue shorter TTM.
• Considering the development paths of processing
  machines so far, this indicates the need to shift
  our RD focus
• Roadmapping helps to address the key issues on
  the right level of abstraction, somewhere in
  between the function and technology.

27
Failing scaling and Roadmapping to new
Architectures

• Can we use the existing architectures with new
  and future technology for the next generations of
  processing machines in Radio Astronomy?

Martijn van Veelen Jaap Bregman