Stacked Die From A Networking Angle

2.5D and 3D are still expensive, but companies have committed to this strategy and are working out the kinks to lower costs and improve customization capabilities.


By Mark LaPedus
The first wave of 2.5D chips using silicon interposers are trickling out in the marketplace.FPGA vendor Xilinx was the first chipmaker to ship a 2.5D device, and Altera, Cisco, Huawei and IBM recently have talked about their respective 2.5D chip developments.

Generally, Altera and Xilinx have taken a somewhat identical and straightforward approach. The two companies are separately developing 2.5D FPGAs, initially based more on homogenous devices. Both are also using Taiwan Semiconductor Manufacturing Co. Ltd.’s turnkey solution to integrate all or part of their 2.5D FPGAs.

Huawei is taking a different avenue, which is arguably more representative of the complex approach that many may take in their 2.5D efforts. The Chinese networking equipment giant is developing a heterogeneous 2.5D device that combines an FPGA from Altera and stacked DRAM from Tezzaron. The interposer comes from Singapore’s Institute of Microelectronics (IME). And fabless ASIC vendor eSilicon is handling the supply chain and integration process.

Putting the pieces together is expected to be a herculean effort. But having explored a multitude of options, Huawei decided to move down an arduous path—and for good reason. “The memory wall is a very serious problem,” said Anwar Mohammed, a senior staff scientist at Huawei. “The gap is becoming wider and wider. And all of the solutions we have for solving the problem are not working anymore.”

For the high-end networking space, Huawei sees a clear but challenging path to solve the problem. “We have to punch a tunnel through the memory wall,” Mohammed said. “For networking applications, 2.5D is the preferred solution.”

The roadblocks
The memory bottleneck and resistivity problems in planar devices have fueled the development of stacked 2.5D and 3D chips using through-silicon vias (TSVs), whether those TSVs run through a die or a separate interposer die in 2.5D designs.

Mike Splinter, chairman and chief executive of Applied Materials, said the 2.5D/3D chip market represents a promising segment for the IC industry, but the business will take some time before it reaches mass production. “We’ve always said there will be a slow deployment of 2.5D,” Splinter said.

The 2.5D chip market is progressing somewhat faster than 3D. Several foundries and IC-packaging houses currently provide interposers and workable manufacturing flows to enable 2.5D designs. There are still some gaps in the technology, however.

“2.5D depends on having a stacked memory solution,” said E. Jan Vardaman, president of TechSearch International, a research firm. “The inability to obtain a memory stack is a gating factor. Some people also say the cost for 2.5D is too expensive.”

Test is also an important but sometimes overlooked part of the flow. “The test challenges for 2.5D are very similar to 3D. For die stacking, it is crucial to have each die pre-tested for KGD,” said Bassilios Petrakis, product marketing director at Cadence Design Systems.

“In the case of the interposer, the question often comes up as to whether it needs to be tested for connectivity upfront prior to bonding with other dies. There is also consideration for how to test partially populated interposers as well as multiple die stacks,” Petrakis said. “An example of that would be a logic die that talks to a Wide I/O DRAM and another logic die on top. If the bottom die of the interposer is the most expensive die, you may only want to attach it to an interposer with all other die attached that have been tested good so far. This may be the most economical way to produce good modules. Finally, all dies on interposers must have some form of a wrapper with boundary scan. We prefer the use of IEEE 1500-style wrappers, but we are also able to accommodate the simpler Wide I/O style boundary scan. Special I/O wrap test before die stacking/bounding can detect possible TSV shorts but not opens.”

Another challenge is to find a suitable manufacturing partner. In general, there are two schools of thought—turnkey versus a hybrid approach. TSMC and Samsung provide a turnkey solution, in which the companies provide both the front- and back-end work. In contrast, GlobalFoundries and UMC are sticking with their hybrid approaches. In that model, the foundries handle the front-end steps, but pass on the back-end work to the IC packaging houses.

Both approaches have their advantages and disadvantages. In the turnkey approach, the foundry can assume the responsibility of the supply chain, thereby keeping costs and quality under strict control. The problem with the turnkey method is that some customers are nervous about handing over their sensitive front-end, assembly and test intellectual-property (IP) to a foundry, said Ajit Manocha, chief executive of GlobalFoundries. “We are not a closed fab,” Manocha said. “Customers prefer to take their proprietary information to the OSATS. We are not going to force customers to do the assembly with us.”

Taking the right path
As it turns out, each customer will choose its own path. To simplify its respective supply chains, Altera and Xilinx are working with a limited set of partners. Most others may end up dealing with a more complex supply chain.

Huawei, for example, is working with separate chipmakers, interposer suppliers, foundries, assembly houses and integrators. At present, Huawei is developing its 2.5D ASIC/FPGA device at IME, a Singapore R&D organization. IME has set up a complete front-end production flow using fab gear from Applied Materials. IME also developed its own interposer technology. IME is a research institute of the Science and Engineering Research Council of the Agency for Science, Technology and Research (A*STAR).

Huawei declined to comment on which foundry it will use once it moves into production, but the challenges are obvious. “This could be a logistics nightmare,” said Ron Leckie, president of Infrastructure Advisors, a consulting firm.

Unfazed by the challenges, Huawei believes it must move in a new and radical direction to address the memory bottleneck in the network. “At one time, when you went to a new node, your gains were pretty sharp,” said Huawei’s Mohammed. “Now, every time we go to the next node, the power becomes a challenge and you have to go with larger and larger die sizes.”

The current line of specialized networking memory chips and other components are unable to keep pace. “Commodity memory cannot handle it,” he said. “Serdes was able to help with the bandwidth at one time. But now, the gains are flatter.”

To solve the problem, the company evaluated several options. “A company like Huawei doesn’t jump into a technology. We have to go through many doors before we decide this is a technology we go after,” he said.

Last year, for example, Huawei looked at combining an ASIC and RLDRAMs in a 64mm x 64mm package, he said. After dropping that idea, the company looked at integrating those devices in larger substrates or smaller packages. Those options were scrapped. Then, it looked at combining a bare die FPGA and packaged memory in a $25 module. “It was not leading-edge technology,” he said. “Any one of our competitors could have picked it up.”

Finally, the company decided on 2.5D. 3D is more suited for mobile applications. “The size of our line cards is constant. We want to put more and more items on the line card to make it more functional and effective. 2.5D is a very powerful enabler for that,” he said. “Initially, this is going to be more expensive. But if you combined enough items, there is a strong potential for cost reduction. It also allows us a faster time to market.”

In Huawei’s proposed design, the FPGA from Altera and the memory stack from Tezzaron are situated on a silicon interposer. “Instead of 10 or 20 DDR DRAMs, all of this can be replaced by one Wide IO memory,” he said. “DDR memory performance is so slow. All of this goes away with Wide IO memory, which is only 12mm x 12mm.”

In total, the company’s proposed 2.5D device occupies less space. The bandwidth per watt is at least 30 times better than conventional approaches, he said.

To realize its design, the world’s largest networking equipment company must overcome some major hurdles, namely the KGD issues, the lack of EDA tools and the supply chain. “Hopefully, we can obtain known good dies and bare dies,” he said. “There is good work going on at Cadence, Mentor and others, but this is still an area of concern. There are also some business concerns like who’s responsible and who’s not responsible?”

Ultimately, to make 2.5D/3D a viable solution in the overall market, Huawei advocates another critical piece to the puzzle–collaboration. “We are advocating pre-competitive collaboration. Let’s makes sure the technology succeeds. When the technology can take care of itself, let’s start competing,” he added.

  • Jose

    so far no more news with this Huawei IME 2.5D joint venture….looks like it was a publicity (power point style) stunt from ASTAR IME that will do everything to have an alliance with Huawei (see their GaN program and the CBS 48h news on Shane Todd )…at least with press releases the IME exec will get their bonus, but no one shall expect any system data from this publicity stunt joint venture.