The day after Intel had its Technology and Manufacturing expo in China, GLOBALFOUNDRIES kicks off their own version of the event and has made a significant number of announcements concerning upcoming and next generation process technologies. GF (GLOBALFOUNDRIES) had been the manufacturing arm of AMD until it was spun off as its own entity in 2009. Since then GF has been open to providing fabless semiconductor firms a viable alternative to TSMC and other foundries. Their current 14nm process is licensed from Samsung, as GF had some significant issues getting their own version of that technology into production. GF looks to be moving past their process hiccups in getting to FinFET technologies as well as offering other more unique process nodes that will serve upcoming mobile technologies very well.
The big announcement today was the existence of the 12LP process. This is a "12 nm" process that looks to be based off of their previous 14nm work. It is a highly optimized variant that offers around 15% better density and 10% better performance than current 14/16nm processes from competing firms. Some time back GF announced that it would be skipping the 10nm node and going directly to 7nm, but it seems that market forces have pushed them to further optimize 14nm and offer another step. Regular process improvement cadences are important to fabless partners as they lay out their roadmaps for future products.
12FP is also on track to be Automotive Grade 2 Certified by Q4 2017, which opens it up to a variety of automotive applications. Self-driving cars are the hot topic these days and it appears as though GF will be working with multiple manufacturers including Tesla. The process also has an RF component that can be utilized for those designs.
There had been some questions before this about what GF would do between 14nm and their expected 7nm offering. AMD had previously shown a roadmap with the first generation Zen being offered on 14nm and a rather nebulous sounding 14nm+ process. We now know that 12LP is going to be the process that AMD leverages for Zen and Vega refreshes next year. GF is opening up risk production in 1H 2018 for early adopters. This typically means that tuning is still going on with the process, and wafer agreements tend to not hinge on "per good die". Essentially, just as the wording suggest, the monetary risks of production fall more on the partner rather than the foundry. I would expect the Zen/Vega refreshes to start rolling out mid-Summer 2018 if all goes well with 12LP.
RF is getting a lot of attention these days. In the past I had talked quite a bit about FD-SOI and the slow adoption of that technology. In the 5G world that we are heading to, RF is becoming far more important. Currently GF has their 28FDX and 22FDX processes which utilize FD-SOI (Fully Depleted Silicon On Insulator). 22FDX is a dual purpose node that can handle both low-leakage ASICs as well as RF enabled products (think cell-phone modems). GF has also announced a new RF centric process node called 8SW SOI. This is a 300mm wafer based technology at Fab 10 located in East Fishkill, NY. This was once an IBM fab, but was eventually "given" to GF for a variety of reasons. The East Fishkill campus is also a center for testing and advanced process development.
22FDX is not limited to ASIC and RF production. GF is announcing that it is offering eMRAM (embedded magnetoresistive non-volatile memory) support. GF claims that ic an retain data through a 260C solder reflow while retaining data for more than 10 years at 125C. These products were developed through a partnership with Everspin Technologies. 1Gb DDR MRAM chips have been sampled and 256Mb DDR MRAM chips are currently available through Everspin. This technology is not limited to standalone chips and can be integrated into SOC designs utilizing eFlash and SRAM interface options.
GLOBALFOUNDRIES has had a rocky start since it was spun off from AMD. Due to aggressive financing from multiple sources it has acquired other pure play foundries and garnered loyal partners like AMD who have kept revenue flowing. If GF can execute on these new technologies they will be on a far more even standing with TSMC and attract new customers. GF has the fab space to handle a lot of wafers, but these above mentioned processes could be some of their first truly breakthrough products that differentiates itself from the competition.
Currently GF has their 28FDX
GloFo does not have 28FDX. It does have 22FDX and 12FDX. Although GloFo licenced 28nm FD-SOI from STMicroelectronics, they decided to start with 22nm.
Also see: Globalfoundries’
Also see: Globalfoundries’ CEO: Why FD-SOI and Why Now
What about IBM’s technology
What about IBM’s technology that GF uses/licenses from IBM to provide for IBM’s z14 fab process node needs(FinFET and SOI in 14HP Process) Technology. That z14 is running its CPU cores at 5.2 GHz.
“Each IBM z14 SC CPU consists of 6.1 billion transistors, runs at 5.2 GHz and contains 10 cores with dedicated 6 MB L2 per core (2MB L2 for instructions, 4MB L2 for data) and 128 MB shared L3. Meanwhile, the system control (SC) chip consists of 9.7 billion transistors and features 672 MB of L4 and interconnects to ensure coherency between CPUs. Needless to say, that both CP and SC are extremely complex and benefit from manufacturing technologies with small feature sizes.”(1)
(1)
“GlobalFoundries Weds FinFET and SOI in 14HP Process Tech for IBM z14 CPUs”
https://www.anandtech.com/show/11862/globalfoundries-weds-finfet-and-soi-in-14hp-process-tech-for-ibm-z14-cpus
So AMD is going to be
So AMD is going to be updating Its Zeppelin die production to GF’s “12nm” LP(“Leading Performance”) Node(1) that appears to be very close in size to Samsung’s new 11nm node.
Looks like there will be more of a GF 12nm Zeppelin die refresh/newer stepping coming from AMD with maybe that 7nm not so needed as soon. That is only if AMD is able to, at 12nm, get those Zen/core’s clock speeds up there using that GF 12nm process to get nearer to Intel’s latest CPU Coffe/Whatever designs’ clock speeds.
And I’d rather AMD come out sooner with a 12nm Zen/Zeppelin Die/stepping refresh than wait longer for that 7nm Zen2 version to arrive a later(?) Time! Because at 7nm things are a bit more complicated to get to market in any advertised time frame, and this is an AMD time frame we are speaking about. So even more so I’d rather that AMD try and get a Zen(First generation) and new Zeppelin die stepping out the door ASAP and get those Zen, Ryzen/Threadripper/Epyc clocks speeds higher to be more competative with Intel’s CPU cores.
(1)
“Some of AMD’s next chips will arrive on GloFo’s new 12LP process”
http://techreport.com/news/32579/some-of-amd-next-chips-will-arrive-on-glofo-new-12lp-process
I’m hoping that 12LP provides
I’m hoping that 12LP provides for AMD’s version of a “tick”, although it sounds like their “tock” will also arrive on a new process so perhaps the analogy breaks down there.
Now there appears to be
Now there appears to be rumors over at guru3d that AMD Might Replace it’s RX 500 Cards with RX Vega 28 and RX Vega 32 variants based on the Vega 11 die configuration and use 4GB of HBM2.
So maybe that 12nm(Leading Performance) Node from GF may come in handy for these rumored Vega 11 based variants and we can finally see what Vega’s HBCC/HBC(HBM2) technology will bring to any RX 500 series replacments that will be Vega 11 die based and Discrere Mobile SKUs that will only come with 4GB of HBM2.
So there can be better thermal/power effencies using GF’s 12nm LP IP and that Vega HBCC IP in its using even 4GB of HBM2 as a last level VRAM Cache to a larger amount of virtual VRAM paged out to regular system DDR4 DIMM based DRAM or even SDD/Hard-Drive. And any Vega 11 SKUs that only offer 4GB of HBM2 still performing like they have 8GB or More Gigabytes of effective virtual VRAM via that HBCC/HBC(HBM2) IP that Vega has to offer.
So even though Vega’s HBCC is not taken advantage of much for gaming on any current Vega SKUs with 8GB of actual HBM2, because the games are already optimized mostly for 8GB of actual VRAM, those rumored Vega 11 based variants will only have 4GB of HBM2 and that can be compensated for by Vega’s HBCC useing the 4GB of actual HBM2 as a last level cache for the GPU to leverage regular system DRAM as virtual VRAM above the only actual 4GB of HBM2 that these Vega 11 based RX 500 replacments will offer for gaming.
So maybe there can be Vega 11 variants, RX Vega 28 and RX Vega 32 with higher clocks and better thermal performance because of that new GF 12nm “Half Node” improvment over the older 14nm node that GF licenses from Samsung. And there will also be Vega’s HBCC IP that will help make up for some of that 4GB of HBM2’s smaller size by Vega’s HBCC using the 4GB of HBM2 provided as a last level HBM2/HBC to a larger amount of Virtual VRAM out on regular system DDR4 DRAM.
Vega 11 will really have some features that will allow it to be competative with Nvidia’s GTX 1060 and maybe even some GTX 1070 SKUs. For sure Vega 11’s lesser amounts of shaders on less CUs will use less power and if AMD can keep from removing too much ROP/TMU units in Vega 11 at 12nm that may just force Nvidia’s hand and make Nvidia bring on more improvments sooner in order to compete.
P.S. Hopefully by the time any 12nm Vega 11 products are ready SK Hynix will hopefully have their competing HBM2 supply problems resolved and AMD will be able to have a more plentiful second source supply of HBM2. This should force Samsung to lower their HBM2 pricing a little so HBM2 can become more affordable to use in both mainstream and flagship GPU offerings.
I think that AMD’s RTG may have forseen HBM2 higher costs and that’s partally why RTG/Raja created that HBCC/HBC IP for Vega based systems with smaller amounts of HBM2(2GB-4GB). So Vega’s HBCC/HBC IP can really shine on any mainstream/mobile Vega GPU variants that have the ability to make the HBM2 act like a last level VRAM cache. That HBCC/HBC IP is even used for Compute/AI workloads where in memory datasets are much larger than even 16GB or 32GB of HBM2 and even for professional graphics non-gaming rendering where animation scenes can be very large(textures/meah data) sizes also above even 16GB or 32GB of HBM2 VRAM size.