by Anonymous Observer
While everybody in the EDA business is discussing the Cadence and Mentor merger, in my opinion the glory days of EDA as an independent industry will be over quite soon. The most likely scenario will be that foundries like TSMC will buy EDA companies and offer a perfectly integrated flow -- Verilog to packaged chip -- from one source.
If we look at the technologies of today like 90nm or 65nm, what do we see? The development of a stable technology at these nodes is so expensive that fewer and fewer foundries are still in the business. Only the largest ones like TSMC survive, while smaller foundries like IBM and Chartered are creating alliances to share costs. And, this consolidation process will continue.
As far as I know, only a few companies/alliances remain. The list includes the IMEC alliance (Intel, TSMC, NXP, TI, Panasonic and memory manufacturers), UMC, Selete (an alliance of Fujitsu, NEC, Renesas and Toshiba), and the Open Platform alliance of IBM, AMD, Freescale, Chartered, STMicroelectronics, Infineon and Samsung. This number of alliances exactly matches the number of big EDA companies able to offer a complete RTL-to-GDS flow.
On the other side, it is next to impossible for an EDA company to create a PDK relying only on documentation provided by the foundries. Even simple things (in previous technologies) such as rundeck for metal fill are becoming so complex, that only the foundry itself can implement it. A description of all features, which such a rundeck should have, is too difficult and not interesting for a regular designer. The same is true for more complex files, things like models for lithography simulation, technology description for correct parasitic extraction, and more.
So, developing your own PDK is becoming less and less common practice; usually the PDK is developed by the foundry itself. But the question for the foundry is: which tools from how many different EDA vendors they should support?
PDK development is expensive -- every step must be QA'd. And at the end, the integration of the various parts of the PDK must be QA'd as well. Supporting different tools from different vendors multiplies the effort involved here. So, instead of providing PDKs for tools from different vendors, the foundries either concentrate on a single tool, or create their own standard format for providing technology data. The vendors must then adapt their tools to that format.
Here are examples of both practices:
Even if they could read Calibre rundecks flawlessly, there is still no 100% guarantee that the results are the same as with Calibre, so an extensive QA must be done here as well. Since the market share of the competitor tools is small, TSMC and other do not really care. And the last self-defense maneuver from Mentor is the option to encrypt the rundecks to hide some technology data and prevent the user from changing them. The nice side-effect of this strategy is that no competitor's tool can read the rundecks.
Another problem for the foundries is the availability of the required tools for their technology. Every technology step requires new tools, like model-based lithography simulation, CMP simulation, multi-corner timing analysis, pattern-based DRC, DFM-aware routing, and so on.
So far, the foundry has had to trust that either:
a) A start-up company is founded to create such a tool, the company is then bought by a large EDA vendor and integrated in their flow.
b) Or a large EDA company will, itself, come up with a solution. However, even if the reference flow is available, mergers like the proposed one between Cadence and Mentor, or the Synopsis and Avanti merger, don't guarantee that the tools will be available for the whole lifetime of the process.
It just doesn't sound like an independent EDA industry is a reliable partner for the foundries going forward.
What is the usual scenario for a fabless company, which decides to design a chip? First they go to a fab and ask for conditions, then they go to the EDA vendors and ask for the tools. Of course, every EDA sales and marketing guy will claim that their tools are perfect for the process, the PDKs are available, everything is interoperable and also cheaper than the competition.
Well, 'cheaper' is a very relative term. A per-seat license does not cost $100,000 a year, it only costs $90,000 a year -- which is still lot of money for a startup or a medium sized company. The moment of truth comes at the beginning of the design: the PDK is not as suitable as it seemed to be beforehand, and the tools are not compatible or do not support the required features.
Wouldn't it be much easier for the customer to just make one stop for all of this, namely at the foundry -- a foundry which gives him all the tools and all the PDKs which are perfectly integrated and support the process?
The cost calculation for a total price, including the tools and the manufacturing of the chip, would be completely different as a result, and probably cheaper than buying tools and manufacturing from two different sources. The design entry hurdle is lower, so foundries -- which are interested in full utilization of their fabs -- can hope that more and more customers will make this decision and start designing through them, while the foundry only has to support only one tool chain. It begins to look like a win-win situation for both the customer and the foundry.
The concentration of EDA companies is over. Mentor-Cadence, Synopsis and Magma (I highly doubt that Magma will remain independent in the short term) both offer complete tools chain required for RTL-to-GDS. So, if a foundry buys one of these companies they have a complete solution. Intel was always using own tools, so they are not so dependent on EDA industry. IBM also has a set of own tools, which they could give to the customers, so only Selete, UMC and TSMC remain.
In this scenario, a few questions still remain:
1. What about the automotive and analog markets? There are still foundries left, which need EDA tools.
Well, the EDA industry will not die completely. There will still be small companies left, which can prosper in these areas. Big money is made with the latest technologies. Technologies like 0.25um or 0.18um can be covered by small competitors; there is enough know-how available to manage this.
2. What about front-end design? Cadence and Synopsis have lot of business in these areas as well.
A foundry is not interested in front-end business, so after buying an EDA vendor they probably split the company and the front-end business will remain independent. In fact most likely scenario will be that EDA companies first will be bought by private equity company, taken from public market, reorganized, splitted and different parts either will form an independent company or will be sold. In this case it would be very stupid of big foundry not to buy.
3. Fab business and EDA software development are very different areas. The CEO of a combined company would have to be a genius to manage both areas.
This is a valid question. But IBM and Intel show that in-house development of EDA tools is possible, and their tools are highly competitive with tools offered by Cadence and others.
So from my point of view, it's a highly logical step for foundries to start providing EDA software to their customers. The big days of EDA companies are over, the market is consolidating, and it's time to change the rules of the game. No more shopping between different vendors.
Everything a design company needs, they will get from a single source, and (probably) at a much lower price than before. There will be no more incompatibilities between different tools, no unsupported features of the process, and no more partially-functioning PDKs.
Sounds like a dream, doesn't it?
July 27, 2008
________________________________________________
The author of this article is an independent observer of the EDA Industry who wishes to remain anonymous at this time.
________________________________________________
For EDA Confidential: www.aycinena.com
Copyright (c) 2008, EDA Confidential. All rights reserved.