Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/threads/more-than-moore-ibm-announces-0-7nm-process-node-significant-logic-sram-scaling-vs-2nm.25369/page-2
        )

    [addOns] => Array
        (
            [DL6/MLTP] => 13
            [Hampel/TimeZoneDebug] => 1000070
            [SV/ChangePostDate] => 2010200
            [SemiWiki/EmailDomainReplace] => 1000010
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000970
            [ThemeHouse/XPress] => 1010570
            [XF] => 2031070
            [XFI] => 1060170
        )

    [wordpress] => /var/www/html
)

More than Moore - IBM Announces 0.7nm Process Node (Significant logic/SRAM scaling vs 2nm)

The CFET (IBM's device) is very different from Huawei's "logic folding", even though, superficially, both involve some kind of stacking. From the design standpoint, the CFET is simply a rearrangement of P and N transistors within the standard cells: the two complementary transistors are just placed on top of one another, rather than side-by-side in the cell, as is currently done. Pin connectivity and intracell interconnects can also be a little more complicated with the CFET, but let's set that aside for now.
From the circuit design perspective (EDA), once a library of CFET standard cells is built, the rest of the design follows in the traditional way: all those cells are still located in the same 2D plane, so standard 2D design concepts can be used.

Logic folding, on the other hand, involves splitting chip designs or macro designs between two different planes of standard cells (actually coming from different wafers, in Huawei's case). For EDA tools, this is much more disruptive and difficult, and this could be Huawei's true breakthrough. Starting from the circuit schematics, the tool has to decide on which plane to place the circuit components, and utilize the two tiers so that both are efficiently packed with transistors.

In terms of process, CFET and Huawei's logic folding are also completely different: the core of the CFET process, as it is currently implemented by all chip makers, is in the front-end-of-line part of the flow (i.e. the part of the flow that deals with the transistor itself), and followed by interconnect patterning. The stacking is done by patterning both N and P transistors right after one another (or even at the same time, as in the case of the monolithic CFET). In Huawei's case, however, we start from two wafers with interconnected transistors, and bond them using bond pads located on the upper metal levels. The new process challenge involves hybrid wafer bonding at a relatively tight pitch (1.5um). Hybrid bonding is not trivial, but still orders of magnitude simpler than creating a manufacturable CFET process!
 
IBM Semiconductor Technology Research (STR) is part of IBM Research and is in an awkward position. After IBM paid $1.5 billion to have GlobalFoundries take over its semiconductor manufacturing division in 2015, IBM became a semiconductor company that has labs but no fabs. IBM no longer has its own commercial fabs to test or co‑develop manufacturing technologies, let alone support high volume production.

Compared with the not‑for‑profit structures of IMEC in Belgium and ITRI in Taiwan, IBM STR operates under a for‑profit corporate umbrella and must keep producing leading edge, high visibility breakthroughs to justify its existence and attract new funding sources. This is challenging because there are only four leading edge semiconductor manufacturers in the world (excluding mainland China): three established incumbents - TSMC, Intel, and Samsung - and one potential newcomer, Japan’s Rapidus. Rapidus is the only company licensed to use IBM’s 2nm technology. TSMC, Samsung, and Intel all appear uninterested in adopting or paying for IBM’s technologies.

GlobalFoundries exited leading edge logic manufacturing in 2018. Worse, the deteriorating relationship between IBM and GlobalFoundries eventually led to lawsuits accusing each other of contract breaches or trade secret misappropriation. They finally settled these disputes in January 2025.

As leading edge semiconductor research and manufacturing become extremely expensive and complex, it is unclear how long the current business model and organizational scale of IBM Semiconductor Technology Research can remain sustainable.

This is the most read blog on SemiWiki:

 
Back
Top