First 3D Nanotube and RRAM ICs Come Out of Foundry
Here's something you don't see very often at government-sponsored technology meetings—spontaneous applause. It happened at DARPA's Electronics Resurgence Initiative Summit this week when MIT assistant professor Max Shulaker held up a silicon wafer that is the first step in proving DARPA's plan to turn a trailing edge foundry into something that can produce chips that can compete—even in a limited sense—with the world's leading edge foundries.
"This wafer was made just last Friday... and it's the first monolithic 3D IC ever fabricated within a foundry," he told the crowd of several hundred engineers Tuesday in Detroit. On the wafer were multiple chips made of a layer of CMOS carbon nanotube transistors and a layer of RRAM memory cells built atop one another and linked together vertically with a dense array of connectors called vias. The idea behind the DARPA-funded project, called 3DSoC, is that chips made with multiple layers of both would have a 50-fold performance advantage over today's 7-nanometer chips. That's especially ambitious given that the lithographic process the new chips are based on (the 90-nanometer node) was last cutting-edge back in 2004.
The project is only about a year old, but by the end of its 3.5-year run, DARPA wants a foundry technology that makes chips with 50-million logic gates, 4 gigabytes of nonvolatile memory, and 9 million interconnects per square millimeter between the layers that can transmit 50 terabits per second while consuming less than 2 picojoules per bit.
What Shulaker showed on Tuesday can't do all that yet, of course. But it's a key milestone in that journey. Together with SkyWater Technology Foundry and other partners "we've completely reinvented how we manufacture this technology, transforming it from a technology that only worked in our academic labs to a technology that can and is already today working inside a commercial fabrication facility within a U.S. foundry," he said.
Here's the paper I've linked a dozen times in the last year.
(Score: 2) by Alfred on Wednesday July 24 2019, @03:22PM (4 children)
Like any specs they will be misleading
What do they mean by transmit? I thought they meant moving data in the chip, maybe thats not it. 50Tbit(10E12)*2pJ(10E-12) = 100 Joules/sec = 100 watts. Just to move data? That seems off. I doubt my desktop CPU, without the advantage of on die communication, uses that much to talk to RAM. So how much more when we add CPU consumption? Maybe they are looking to that as a burst rate and not continuous.
Either way, specs are misleading
(Score: 2) by takyon on Wednesday July 24 2019, @03:31PM (3 children)
That refers to memory bandwidth between the CPU and integrated DRAM. See page 5:
Memory Access Parameter / 2D / 3D TSV Package / 3DSoC
Total I/O / 512 / 1K / 33K
Max Bandwidth (Gb/s) / 400 / 1K / 46K
Memory access energy (pJ/bit) / 52 / 32 / 1.5
VDD (Volts System) / 1.6 / 1.2 / 0.6
46K Gb/s = 46 terabits per second
The advantage here is that by shortening the distance data has to travel from the CPU to the DRAM, you can get incredible performance increases and reduce the energy needed.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 2) by Alfred on Wednesday July 24 2019, @08:08PM (2 children)
(Score: 2) by takyon on Wednesday July 24 2019, @08:47PM
They say that it "dissipates < 500mW of average operating power" so that's what I'm going to go with.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 2) by sgleysti on Thursday July 25 2019, @12:07AM
Since they call the 2pJ/bit metric "memory access energy", I'm guessing this is the energy to read one bit of the nonvolatile memory. Since this is specified at 4GB = 32Gb, it would take 2pJ/bit * 32Gb = 64mJ to read the entire memory on the chip once.
If that's all that metric means, it is orthogonal to the power required for the 50Tb/s transmission bandwidth, and the summary is just confusing. This is my best guess.