optimizing existing pdn to recover area of soc in …atul bhargava1, abhishek nigam1, fabrizio...
TRANSCRIPT
RESEARCH POSTER PRESENTATION DESIGN © 2012
www.PosterPresentations.com
Set-Top-Box(STB) SoC designs are extremely
complex with multi-million gates, higher core
utilization of around 75-80 %, and multiple clock
domains. Die-Area optimization of such SoCs becomes
a key challenge.
Our experience on these SoCs consistently
demonstrates that:
• The PDN (Power Delivery Network) lie well within
the IVD budget of 15% of supply voltage.
• We continue to see routing congestions which is
shared by both signal/clock due to PG grid
overdesign
• Using an early Power Grid Prototyping based
approach, we got results that predicted strong
potential for optimizing area of the die without
compromising on cost and performance.
In this work, we first leveraged an existing version of
the STB SoC and optimized the PDN to (a) quantify the
increased signal and routing resources created and (b)
its affect on IR drop. In the second phase, we
implemented these PDN optimization changes on a
shrinked version of the SoC and took it through the
PnR and timing closure flow to validate the area
recovery of these SoCs in production.
a) Phase 1- Optimizations using an existing version of
SoC
• Optimize M3 PDN for routing and tracks
availability for 7ML and 8ML stack configurations
of existing SoC. (M2-M7 stack via for PDN OR
M3-M7 stack via)
• Quantify the effect on IR when M7 pitch is made
2x to reduce PG stack-vias.
b) Phase 2- Reduced version of SoC (Die area reduced by
12%)
• Took the reduced version of the SoC through PnR,
Timing and IR for standard grid (M7 pitch 7.2u)
and optimized grid (M7 pitch 14.4u) to validate the
Area Recovery.
Motivation
The IVD analysis with M7 pitch 7.4u and 14.4u were done
using RedHawk-Grid Prototyping System tool, and the
worst IR drops were within budgets.
Atul Bhargava1, Abhishek Nigam1, Fabrizio Viglione2 ,Shiv Om Sharma1, Anant Narain3
Optimizing Existing PDN to Recover Area of SoC in Production
1STMicroelectronics, Noida, India 2STMicroelectronics, Grenoble, France3Ansys, Inc., Noida , India
Results and Findings
Methodology
Phase 1a: M3 PDN Routability Analysis:
Power Grid-A allows denser routing for all configurations
Phase 1b: Optimizing M7 PDN:
Grid AAvailable tracks per layer
M2(H): 10/16 (standard cell rails)
M3(V): 152/170 (M3 stripes, pitch 17u)
M4(H): 44/48 (stacked via)
M5(V): 152/170 (stacked via)
M6(H): 38/48 (stacked via)
Grid BAvailable tracks per layer
M2(H): 10/16 (standard cell rails)
M3(V): 82/90 (stacked via, pitch 9u)
M4(H): 12/16 (stacked via every 2 rows)
M5(V): 82/90 (stacked via)M6(H): 6/16 (stacked via every 2 rows)
Grid A
Total Available tracks:
112/144 = 77.8% (H)
304/340 = 89.4% (V)
Grid BTotal Available tracks:
28/48 = 58.3% (H)
164/180= 91.1% (V))
Routability Analysis for 8ML and 8T Libraries.
Phase 2: Reduced SoC – Routing Analysis
The SoC die-area was reduced while the area of the IP
remained the same. This increased routing demands near
the Hard IP regions and caused congestions at high-cell
density areas
Routing resources
demand at SoC level
Congestion at High
Cell Density area
Phase 2: Reduced SoC – routing analysis/congestions/timing and IR drop comparisons between standard M7 PG pitch (7.2u) SoC and 2x M7 PG pitch (14.4u) SoC.
Results and Findings
M7 pitch
7.2u
M7 pitch
14.4u
Congestion reductions in channels and notches with
2x M7 pitch due to reductions in M3-M7 PG stack
vias by 50%
M7 pitch
7.2u
M7 pitch
14.4u
Congestion reductions with 2x M7 pitch for same region
in two cases with the signal nets remaining the same in
two cases.
M7 pitch 7.2u
M7 pitch 14.4u
No timing impact with
2x M7 pitch !
M7 pitch 7.2u M7 pitch 14.4u
The IR drop with 2x M7 pitch
only increased by 5mV
(within budget)
• Die-Area optimization of STB SoCs has become a key challenge.
• A two-phase methodology and results for recovering die area of such SoCs in production has been presented.
• In first phase, we leveraged an existing version of the STB SoC and optimized the PDN for 8ML and 8T libraries.
• Our results demonstrate 77% available horizontal tracks with M3 PDN routing versus just 58% without it.
• Also the increase in IR drop is well within the 15% voltage drop budget when the M7 pitch is doubled to reduce the M3-M7 stack via density by 50%.
• In second phase, both the M3 PDN routing and 2x M7 PDN pitch are implemented on a reduced-area version of the same chip and taken through PnR, timing and IR flow to demonstrate the feasibility of recovering the area of the SoC without compromising on performance.
Summary