IFTLE 250 Toshiba Extends TSV Stacks to NAND; UMC to Supply TSV Based Silicon Interposers; Nvidia’s Pascal Coming in 2016 with HBM2
By Dr. Phil Garrou, Contributing Editor
Wow, the big 250!
Thanks to all of you that continue to read my updates and opinions. Thanks to all of those who have supplied me information to pass on to all of you.
I have told many of you that these blogs are my way of keeping myself up to date. I’ll keep doing these as long as I’m still having fun and as long as Extension Media and Pete Singer want me, too.
Before we get to some major 3DIC news, long time readers know I never miss a chance to update you on how my granddaughters are doing. Hannah (11) and Madeline (7) were in NC visiting a few weeks ago, so I will share this shot with you.
and now the news…
Toshiba Extends TSV Memory Stacking to NAND
IFTLE (Insights From the Leading Edge) which started in August of 2007 as Perspectives From the Leading Edge in the now defunct Semiconductor International has mainly focused on advanced packaging with emphasis on 3DIC since that was the hot and coming technology. By blog #12, I announced that Toshiba had would be commercializing CMOS mage Sensors with TSV (albeit one layer devices, but most would conclude, like I do, that this was the commercial start of what we now call 3DIC.
It is therefore somehow fitting that as I hit IFTLE 250 (and IFTLE + PFTLE 378 ) that Toshiba would be the first to announce the extension of TSV technology to NAND. [link] For years, I have been taking the position that NAND could not come till DRAM was commercialized since it is a less costly product. Well, since TSV DRAM has now been announced in many forms by the major DRAM players (Samsung, Hynix and Micron) I guess its now time for the NAND announcement.
Toshiba’s TSV technology achieves an I/O data rate of over 1Gbps which is higher than any other NAND flash memories with a low voltage supply (1.8V to the core circuits and 1.2V to the I/O circuits) and approximately 50% power reduction of write operations, read operations, and I/O data transfers.
They see this stacked NAND flash memory providing the ideal solution for low latency, high bandwidth solutions in flash storage applications, including high-end enterprise SSD.
Sold as 152 pin BGAs, the 8 stack modules are 14 x 18 1.35mm and the 16 stack devices are 14 x 18 x 1.90mm.
UMC Enters HVM on SI Interposers for AMD FIJI
We announced the AMD FII graphics processor would use high bandwidth memory and silicon interposers in IFTLE 240 “AMD Introduces HBM on Fiji R9 390X GPU.”
It has recently been announced that the silicon interposers are being supplied by UMC making them the second entrant into the commercial silicon interposer business (after TSMC and their CoWoS process). The interposers are reportedly being manufactured by UMC in Singapore, at 300mm Fab 12i.AMD FIJI graphics module with 4 stacks of HBM
The GPU package, measuring 50 x 50 mm, will actually reduce the graphics card PCB size, because the memory has been moved to the GPU package, with four 1024-bit HBM1 stacks surrounding the GPU die (n the package see below). 1GB of HBM memory takes up 95 percent less than the same amount of GDDR5 memory, so you free up a lot of room on a circuit board. AMD has yet to announced whether their next gen Greeenland GPUs, which will use HBM2 memory stacks will be made on TSMC’s 16nm or Samsung’s 14nm process.
SK Hynix ships their HBM stacked to UMC who integrates the HBM stacks and GPU onto their silicon interposer for AMD.
Nvidia Pascal Coming in 2016
Not to be outdone by AMD, Nvidia has announced further details on the their Pascal graphics processor due out in 2016.
Nvidia’s Pascal GPU ( GP100) will feature a 4096 bit bus and four HBM2 stacks each up to 8 memory layers. The Pascal chip set will reportedly be manufactured on both TSMC’s 16nm FinFET process [link 1] and Samsungs 14nm FinFET process [link 2] later next year. Rumors are that Nvidia has taped out the GP100 at TSMC and thus should be expecting the first chips from them in Q2 2016. The GPU will be made with either 4 or 8 memory layers in the HBM stacks. They believe the Pascal GPU will be able to achieve 10x better performance compared to Maxwell.Nvidia Pascal GPU Module
OK, Intel, your turn…we’re waiting !
For all the latest on 3DIC and advanced packaging, stay linked to IFTLE…