Revision T was made for a number of reasons:
1. Electrical testing of chip I1 revealed the word lines were shorted to VDD. The fix involved removing an unwanted split contact, which affected the thinOx, nImp, pImp, and cut layers. (Actually, the fix was done so as to avoid having to change the nImp and pImp layers, but they ended up having to be changed for another reason; see item 4c below.) (A static connectivity checker would have caught this error. We didn't have one last December when Revision A was made, but the new Core software will have one.)
2. A desk check revealed that the interface between the DecoderLogic cells and *RamTwoBits cells had some 0.5-micron errors in metal2, involving some design rule violations of both minimum width and minimum spacing. (They were not caught earlier because Spinifex ran out of virtual memory trying to check the entire chip, and I didn't think of trying to check some of the larger cells. The CrossRAMCellLibrary was checked by Spinifex and had no DRC errors. The error was introduced because core/patchwork flipped some of the DecoderLogic cells but none of the *RamTwoBits cells. That placed a symmetry requirement on those cells of which I was unaware, so I made them asymmetric.)
3. A bug in the CDMEBES DeNotch function was discovered; it had been undersizing shapes by 0.75 microns per side. The bug was fixed. We think that only the CrossRAM03 used this DeNotch function, and we know only the nImp and pImp layers were affected. (See also item 4c below.)
4. VTI biasing:
a. VTI's latest Tooling Specification, 02-10001, Rev *C, required a change of the thinOx biasing to 0.75 microns per side. (This was the latest in a series of flip-flops between 0.75 microns and 0.9 microns per side. You probably don't want to know the details. But this time we think we've got it right.)
b. We (Bob Allen) decided to take VTI's biasing for the cut layer seriously. It's -0.25 microns per side on the mag tape, and -0.15 microns per side on the mask plate. We had been hoping to use 0.0 microns per side, the same as ICL, so we could share this mask between ICL and VTI. (ICL almost certainly can't use this bias on their P&E 341 projection aligner.)
c. We also decided to take a closer look at VTI's specification for nImp and pImp masks. (VTI's latest spec, 02-10001, Rev *C, changed the specification for nImp, but we hadn't been following the old specification very well either.) Oversimplified, we used to bias both nImp and pImp by 1.5 microns per side, but now we bias nImp by 2.0 microns per side and pImp by 2.5 microns per side for VTI, as they seem to specify. This has the effect of providing better protection of diffusion shapes from unwanted implants, but may risk failing to implant all of the diffusion that should be implanted. (We don't yet know whether ICL can live with this biasing.)
See /nimitz/cedar/users/Hoel.pa/Dragon/Foundry/-
ImplantMaskIssues7Feb86.tioga for details.
d. Note that we still aren't doing everything according to VTI's Tooling Specification. But we think we have the appropriate waivers from VTI.
i. nWell and pWell are negotiated specially, since we want a p+ to n+ spacing of 10 microns and VTI's standard process supports 12 microns.
ii. pImp is generated to protect only ndif rather than everything except pdif. We think that's good for the implanter equipment. We also think it doesn't make any difference to the chip.
iii. cut2 is still made the same for ICL and VTI, at 0.5-micron resolution and no biasing. However, in the special case that the via is drawn 2 x 2 microns and overlaps diffusion, VTI asks for biasing the portion that is over diffusion by 0.25 microns per side. (That requires 0.25-micron resolution.) (Note that VTI's approach could make some non-rectangular vias on the mask plate.)