Graphics chip demands rethinking block design

Graphics chip demands rethinking block design
Graphics processors place unusual demands on designers because of the high data rates required and the complex algorithms that must be routinely executed in graphics applications. To advance state-of-the-art graphics systems, 3Dlabs decided to tackle an 8 million-gate graphics processor, code-named Emma. Emma is the company's greatest silicon challenge to date, being twice as complex as any design done previously.

With Emma's predecessor, we found that meeting the required time scales and completing the routing had been difficult to achieve using traditional EDA software for the logic design and a leading design service provider for the layout. The capacity limitations of the synthesis and place and route tools forced 3Dlabs to employ a complex hierarchical design, requiring that the design be broken into small, multiple blocks and then stitched back together. That resulted in the design service provider's having to decrease the timing goal and try different floor plans, adding months to the design cycle.

Knowing that Emma was going to be more complex than the previous designs and would be manufactured in a smaller-geometry process, we decided to try a new approach: bring the physical design in-house and look for a new design solution. After investigating various offerings in the EDA market, we decided on using the Magma Blast Fusion physical design system. Its ability to deliver optimal timing without requiring iterations back through synthesis, as well as its ability to handle multimillion-gate designs, made the system particularly appealing to the designers working on Emma.

Blast Fusion performs logic optimization, placement and routing. It has a comprehensive suite of built-in design, optimization and analysis tools, all of which operate on a single, unified data model. The unified data model enables the built-in tools to operate concurrently using identical design information. With that integrated architecture and Magma's unique FixedTiming methodology, Blast Fusion determines the optimal timing for the entire design before detailed layout and holds that timing constant during the physical design process, yielding the highest performance.

The FixedTiming methodology relies on Magma's SuperCell models. Both are key to Blast Fusion's ability to predict and deliver optimal performance and to handle very large blocks or entire designs at once.

Instead of determining cell size during synthesis based on estimated wireload models and minimal placement data, Blast Fusion performs initial placement using SuperCells. Each SuperCell represents a specific logic function from the library that has a fixed delay and variable size.

Once initial placement is completed, the optimal timing of the design is determined and reported. Blast Fusion maintains that timing during detailed physical design by varying the size of the SuperCell according to the actual load in the layout. After final placement, when the actual interconnect loads are known, Blast Fusion replaces each SuperCell with a library cell that has the proper size and drive strength.

Blast Fusion can handle large designs because it needs only one SuperCell for each logic function. Traditional approaches evaluate hundreds of cell sizes for each circuit topology.

The design system also produces an electrically balanced circuit and a compact, power-saving layout because each cell is optimally sized for the load it is driving.

Before starting the Emma design, Magma did a benchmark using the most problematic section from the earlier 3Dlabs design: a 3.9-mm x 4.5-mm floor plan block. Unlike the previous tool flow, Blast Fusion had no trouble routing the design. In addition, the timing improved by 20 percent.

The final netlist had 21 percent fewer placeable cells after optimization. The standard-cell area was also reduced. The library input-to-final routed database run-time was just 30 hours, and timing was met without iterating between logical and physical design.

Encouraged by those results, 3Dlabs talked with Magma at length. Both companies agreed that a design the size and complexity of Emma would take a significant commitment.

The partnership plan called for Magma to develop a hierarchical flow. The basis of the work was to split Emma into blocks, check the timing for each block in its own context and then check the timing with respect to the rest of the design. To check the timing of the Emma blocks together, a suitably accurate timing abstraction was needed for each block. The retiming at the top level was the most difficult, since there is no standard procedure and current methods yield varying degrees of accuracy.

To minimize the top-level integration effort and timing difficulties, it was decided early on to make the blocks as large as possible. Emma was split into just four multimillion-gate blocks. Each was more than four times the size of the blocks used in the previous design.

Using the four blocks as the elements for the top-level floor plan, 3Dlabs started the design and synthesis process. Simultaneously, the top level and the abstracts for the top-level blocks were crafted. Though a hierarchical flow is not usually necessary with Blast Fusion, a semiautomatic process was used to create sub-blocks and pin assignments, greatly simplifying the top-level assembly.

Early in the design process, the internal sizes and details of the blocks were unknown; estimates were used based on targeted utilization figures and approximate cell count. As physical details became known, the estimates were refined in the Blast Fusion flow. Consequently, an early abstract floor plan was developed using a practical approach to pin positions and sizing. By design, a floor plan was developed in which each block had a neighbor with a matching pin opposite it to minimize routing traffic between blocks.

With the abstraction from Blast Fusion, I/O cell placement and top-level development began. Decisions about power routing were also made at this time. To allow both flexibility and adequate power distribution, 3Dlabs decided to run power horizontally and vertically with a common grid that would replicate into blocks and would be uniform across the design. The approach offered an additional advantage in that blocks could be moved at the top level without disturbing the routing requirements of lower-level blocks.

From there, logic designers at 3Dlabs began work on the logic, netlist, constraints and back-annotation. Knowing that Blast Fusion would optimize the logic, the designers were able to minimize the logic synthesis portion of the flow. After final netlists for each multimillion-gate block were delivered, four applications engineers — one for each block — began the place and route process using Blast Fusion. The smallest block was completed first and became the test vehicle to prove the flow and to test the timing and physical abstraction capabilities of the Blast Fusion system.

Design rule updates, library updates and a continuous refinement of the flow occurred throughout the process. Blast Fusion eliminated much of the time-consuming and manual trial-and-error processes by delivering useful feedback earlier in the flow. That feedback helped 3Dlabs to optimize the register-transfer-level (RTL) code, libraries and floor plan.

For example, the Emma project used NurLogic libraries. In a three-way partnership, NurLogic, 3Dlabs and Magma specified and tested the library in the flow. Each layout of each standard cell in the NurLogic library was electrically and physically verified. The analysis of each standard-cell layout involved area efficiency, timing performance and suitability to the FixedTiming approach. When necessary, enhancements identified by Blast Fusion were implemented in the NurLogic library.

As the flow and library updates began to stabilize, the other, larger blocks were completed. Of the four Emma blocks, only one had trouble with timing, and there were no routing issues. Overall, Blast Fusion provided a four-to-one improvement in capacity. And after the flow was set up, 3Dlabs was able to get the completed design routed, design-rule-checked, layout-vs.-schematic-verified and timing-clean in four days.

Once the physical design was complete, 3Dlabs verified the results with Avanti's Hercules physical verification tool.