NovaSilicon proposes CDK circuit design kit; two papers show its in-house Agent outperforms GPT-6 Astra on average
NovaSilicon (Shanghai Xinxingyuan Intelligent Technology) proposes the concept of CDK (circuit design kit), and uses two papers to compare its in-house Agent with a general-purpose Agent in two scenarios: analog layout and digital floorplan. On the analog layout benchmark AnalogLayoutWorld, the in-house Novara scores an average of 0.8154 with an acceptance rate of 72.5%, while GPT-6 Astra scores 0.7483 and 70.8%; in the floorplan evaluation, Nova-SI scores an average of 57.97, about 1.9 times GPT-6 Astra's 31.17, while the historical solutions of human engineers score 91.88. The company, founded 5 months ago, has completed 3 consecutive funding rounds totaling hundreds of millions of yuan, with the latest round led by Shunwei Capital.
Why read this
AI score 62
The two papers use the same set of chip design tasks to compare the in-house Agent with a general-purpose Agent, and provide comparable details such as acceptance rate and token consumption.
Topics
Body · AI translation
Add crowdsourced testing promotional copy Add crowdsourced testing promotional copy end Add App-exclusive traffic routing
Published by Machine Heart
Founded in 1987, TSMC only does foundry work. Fabs package their manufacturing processes into PDKs (Process Design Kits) and hand them to customers. The PDK is the fab's core technology, shared with chip companies through a limited interface while the core manufacturing recipes are retained. By using the PDK, chip companies can outsource manufacturing and focus on design; such companies are called Fabless.
Will the division of labor in the design stage also change once?
NovaSilicon's (Shanghai Xinxingyuan Intelligent Technology Co., Ltd., hereinafter referred to as "NovaSilicon") approach is called CDK (Circuit Design Kit). It packages chip design experience and Agents into a single system. From chip specifications to GDS layout, the implementation steps that require repeated iteration in between are handled by NovaSilicon's engineers using the CDK. The concept of CDK was first proposed by NovaSilicon, and the company calls the resulting division of labor Designless — the industry will form a new division of labor, with many design houses shifting from building their own design teams to becoming Designless.
Around the CDK, the company recently released two papers that put the CDK's two Agent systems on the same set of tasks as general-purpose Agents such as GPT-6 Astra. These cover analog layout design and digital chip floorplanning.
The highest average scores in both scenarios came from NovaSilicon's self-developed Agent systems designed based on the CDK: 0.8154 for analog layout (GPT-6 Astra: 0.7483); 57.97 for floorplan, about 1.9 times that of GPT-6 Astra.
In the 5 months since its founding, NovaSilicon has completed 3 consecutive funding rounds totaling hundreds of millions of yuan. The latest round was led by Shunwei Capital, with GGV Capital and Matrix Partners China participating. Previously, NovaSilicon had received investment from 5Y Capital, SenseTime Guoxiang Capital, Shanghai Future Industry Fund, Shanghai Science and Technology Innovation Group, and other institutions.
Since its founding, NovaSilicon has rapidly gathered and cultivated many interdisciplinary engineers and researchers with experience in both AI and chip design.
First to propose CDK: the PDK of the design stage
NovaSilicon's vision is to do once again in the design stage what fabs have done. Agents and simulators are equivalent to equipment on the production line, the CDK is equivalent to the process recipe, the input is chip specifications, and the output is GDS. Table 1 | Comparison of PDK and CDK. Source: NovaSilicon
English text is being prepared.
For the analog direction, the CDK Vega covers topology selection, netlist-to-schematic conversion, device sizing, simulation, and layout. For the digital direction, the CDK Sirius includes back-end design initialization, floorplan, placement and routing, clock tree, physical verification, and timing analysis, as well as front-end specification understanding, modeling, RTL generation, and verification.
The two papers each tested one of these links. The analog layout paper tested Novara within Vega, and the floorplan paper tested Nova-SI within Sirius.
Authors: Nova-Silicon Team
General-purpose Agent draws analog layouts: 91.1% submitted, 21.3% passed acceptance
The first paper's benchmark is called AnalogLayoutWorld.
Layout is the final step in analog chip design. Circuit engineers first finalize the transistor-level netlist, which is essentially a wiring table specifying which devices are used, their sizes, and how they are connected to one another. Layout engineers then draw the construction blueprints based on it, drawing each device as geometric shapes on the silicon wafer, placing them in position, connecting the metal wires, and outputting a GDS file that can be handed to the fab.
AnalogLayoutWorld has 40 tasks, covering amplifiers, reference sources, power management circuits, sampling circuits, and oscillators. The Agent receives a fixed netlist and electrical specifications to perform layout design, and scoring looks at six items: whether a GDS is submitted, LVS (whether the layout and netlist match), DRC (whether process rules such as line width and spacing are met), area compactness, engineering quality, and whether post-simulation meets the target. The engineering quality item was defined jointly by the team and analog layout engineers, and examines device matching, critical signal routing, and power supply, among other things.
Beyond the score, there is also an acceptance line. Only when DRC, LVS, and all post-simulation operating conditions pass, and the layout's aspect ratio and morphology are also qualified, is it considered accepted.
The team tested 13 combinations of general-purpose models and Agent frameworks, and CDK: Novara from Vega was also run on the same set of problems. Table 2 | Main results on AnalogLayoutWorld (excerpt). Data source: Table 1 of the paper
Novara had an average score of 0.8154, ranking first. Its acceptance rate was 72.5%, also the highest. GPT-6 Astra had an average score of 0.7483 and an acceptance rate of 70.8%. GPT-6.1 Sol was 0.7023 and 62.5%. The two configurations of GLM-5.3 had average scores between 0.46 and 0.48, and all other configurations were below 0.30.
Novara's LVS pass rate was 100%. On the engineering quality item, Novara scored 0.816, while the second-place GPT-6 Astra scored 0.566.
Figure 1 | Item-by-item comparison between Novara and five general-purpose Agent configurations. DRC, LVS, and Post-sim are pass rates, and Score is the average score over 40 problems. Image source: the paper
GPT-6 Astra's post-simulation pass rate was 82.5%, higher than Novara's 75.0%, but its acceptance rate was only 70.8%. The paper explains that the part the GPT series loses between simulation and acceptance is mainly layouts with unqualified aspect ratios and morphology. For Novara, from 75.0% to 72.5%, layouts that passed simulation could basically all be accepted.
Simulation all passed, but GPT submitted a 209-micrometer thin strip
Among the 40 problems was a current-starved ring VCO. All three attempts by GPT-6 Astra passed DRC, LVS, and both post-simulations, but the submitted layout was about 209 µm wide and less than 10 µm high, with an aspect ratio between 22.63 and 26.31, while the upper limit specified by the evaluation was 4. Such a long, thin strip is very difficult to fit together with surrounding circuits, and all three attempts received only 0.3 points.
From the trajectory records, Astra first tried folding a long resistor in the circuit, but the folding scheme failed LVS length parameter matching, so it fell back to a straight resistor 203.19 µm long and arranged the remaining devices in a line along the resistor. This is a typical behavior of pursuing hard specifications while ignoring the requirements of real analog circuit tape-out.
On the same problem, the layout submitted by Novara on its first attempt was 34.14 × 42.33 µm, with an aspect ratio of 1.24, and scored 0.9015. Novara pays more attention to modeling that conforms to the layout design ideas engineers actually use, and also obtained layout results closer to engineering use.
Figure 2 | Four layouts of the ring VCO. On the left is Novara's first attempt, and on the right are the three attempts by GPT-6 Astra; all four passed DRC, LVS, and post-simulation. Image source: the paper
An industrial design with 99 hard macros: average score about 1.9 times that of GPT-6 Astra
The second paper switches to digital chip back-end and tests floorplan.
For a digital chip to go from a logical netlist to a physical layout, the first step is to decide where large modules go. These large modules are called hard macros; their shapes are fixed and cannot be scaled. After the hard macros are placed, EDA tools then fill the remaining space with standard cells and wiring. Intuitively, this is like first fixing the positions of dozens of buildings on a plot of land, with the remaining land left for small houses and roads. If the buildings are placed poorly, even if the total area is sufficient, local roads will still be congested.
The paper uses an industrial design: 235,727 instances, with 99 hard macros occupying 71.1% of a 14.123 mm² layout area, and the largest and smallest hard macros differ in area by about 121 times. The layout area is L-shaped, and the missing corner cannot be used.
The evaluation feeds the hard macro placement submitted by the Agent into a prescribed downstream flow: building the power and ground network, performing standard cell placement, estimating global routing, and analyzing setup timing. The evaluation's total score is out of 100, obtained by multiplying four items: congestion, timing, wirelength, and standard cell utilization. The participating systems are divided into three groups. The first group is NovaSilicon's self-developed, locally deployed Nova-SI; the second group is general-purpose engineering Agents, which can query the design database, call EDA tools, and revise after receiving feedback from the downstream flow; the third group has only file read and write permissions. Table 3 | Floorplan evaluation results. Each item is out of 100, and geometrically invalid or failed deliveries are recorded as 0 points. Data source: Table 1 of the paper
Nova-SI's average score was 57.97, GPT-6 Astra's was 31.17, and the other general configurations were between 15 and 23. In the group with only file permissions, 9 of the 12 outputs had hard macros out of bounds or overlapping and were recorded as 0 points. All 8 outputs from Nova-SI passed the geometry check, with a median of 57.52 points. Astra's median was 21.37 points, and 3 of its 4 outputs were lower than Nova-SI's lowest score. When producing solutions, Nova-SI did not use online feedback from the prescribed flow, while the general group containing Astra could call EDA at any time to see results and then revise. In terms of token consumption, Nova-SI averaged 36.13M per output, while the five configurations in the general group ranged from 62.11M to 286.50M.
After removing Nova-SI's best output, the mean was 54.67. The paper multiplied each of the seven scoring scales by 0.5, 1, and 2, tried 2,187 combinations, and found that Nova-SI's mean was 22.30 to 30.38 points higher than Astra's. The highest-scoring of the 8 outputs was 81.09 points. For reference, the historical solution by human engineers on this design was 91.88 points. Figure 3 | Nova-SI's highest-scoring layout (No. 07, 81.09 points), from left to right showing hard macros, standard cells, and a congestion heatmap. Image source: the paper
The future of CDK
CDK is first used in NovaSilicon's own design flow. Based on CDK, NovaSilicon's current service targets are chip design companies and customers with IP needs. There are two initial directions. One is to use the UCIe PHY IP being developed by Vega; UCIe is the industry standard for Chiplet interconnect, and PHY is the physical layer interface within it. The other is design services for 3D compute-in-memory (CIM) chips, and the digital design CDK, Sirius, has already been used in this service.
Large-model vendors are the equipment makers that provide the underlying foundation for CDK, as important as the lithography machines used by TSMC. However, just as PDK and recipes are the core that supports TSMC in changing the semiconductor industry, CDK will also reshape the semiconductor industry from Fabless to Designless.
NovaSilicon brings the Design Kit from process manufacturing into the design flow. This process will, like process manufacturing evolving according to Moore's Law, continue to scale at a rapid pace, realizing a new Scaling route in the field of chip design.article_adlist[
For reprints, please contact this official account for authorization
Submissions or report requests: liyazhou@jiqizhixin.com ]article_adlist Statement: This is an exclusive article for Sina.com; unauthorized reproduction is prohibited. tech_web_article_v2015_block_bottom Non-targeted 300*250 button 17/09 wenjing begin
Non-targeted 300*250 button end