Imec Develops Efficient Processor In Memory Technique for GloFo
Imec and GlobalFoundries have demonstrated a processor-in-memory chip that can achieve energy efficiency up to 2900 TOPS/W, approximately two orders of magnitude above today's commercial processor-in-memory chips. The chip uses an established idea, analog computing, implemented in SRAM in GlobalFoundries' 22nm fully-depleted silicon-on-insulator (FD-SOI) process technology. Imec's analog in-memory compute (AiMC) will be available to GlobalFoundries customers as a feature that can be implemented on the company's 22FDX platform.
Since a neural network model may have tens or hundreds of millions of weights, sending data back and forth between the memory and the processor is inefficient. Analog computing uses a memory array to store the weights and also perform multiply-accumulate (MAC) operations, so there is no memory-to-processor transfer needed. Each memristor element (perhaps a ReRAM cell) has its conductance programmed to an analog level which is proportional to the required weight.
[...] Imec has built a test chip, called analog inference accelerator (AnIA), based on GlobalFoundries' 22nm FD-SOI process. AnIA's 512k array of SRAM cells plus digital infrastructure including 1024 DACs and 512 ADCs takes up 4mm2. It can perform around half a million computations per operation cycle based on 6-bit (plus sign bit) input activations, ternary weights (-1, 0, +1) and 6-bit outputs.
[...] Imec showed accuracy results for object recognition inference on the CIFAR 10 dataset which dropped only one percentage point compared to a similarly quantised baseline. With a supply voltage of 0.8 V, AnIA's energy efficiency is between 1050 and 1500 TOPS/W at 23.5 TOPS. For 0.6 V supply voltage, AnIA achieved 5.8 TOPS at around 1800-2900 TOPS/W.
Promising application: edge computing facial recognition cameras for the surveillance state.
Also at Wccftech.
See also: Week In Review: Auto, Security, Pervasive Computing
Previously: IBM Reduces Neural Network Energy Consumption Using Analog Memory and Non-Von Neumann Architecture
Related: "3nm" Test Chip Taped Out by Imec and Cadence
GlobalFoundries Abandons "7nm LP" Node, TSMC and Samsung to Pick Up the Slack - "The manufacturer will continue to cooperate with IMEC, which works on a broader set of technologies that will be useful for GF's upcoming specialized fabrication processes..."
Radar for Your Wrist
(Score: 2, Interesting) by Anonymous Coward on Sunday July 12 2020, @05:20PM (4 children)
How do these processors-in-memory compare to traditional processor/memory combo in neural network implementation?
(Score: 4, Funny) by takyon on Sunday July 12 2020, @06:31PM (3 children)
I'm not sure, but I believe Intel Xeon Phi and Nvidia GPUs are around the 0.5-2 TOPS/W range. An article about the Groq TSP [techspot.com] gives its performance at 1000 TOPS / 300 Watts, vs. 250 TOPS / 300 Watts or 130 TOPS / 70 Watts for two Nvidia products.
PIM Techniques Boost AI Inference to 8.8 TOPS/W [eetimes.com]
This is also described as processor-in-memory, but only 8.8 TOPS/W.
Gyrfalcon’s New Chip Raises Bar (12.6 TOPS/W) on High Performance Edge AI with Lower Power Use [gyrfalcontech.ai]
Edge AI chip forgoes multiply-accumulate array to reach 55 TOPS/W [embedded.com]
1050-2900 TOPS/W, or 1-3 exaOPS/Watt, is just incredibly efficient, and they are apparently aiming for 10 exaOPS/Watt in a future version (the article says "10 TOPS below 100 mW" but I think they meant "below 1 mW"). The actual performance of 5.8 or 23.5 TOPS is not bad if it can be used by a single camera, for instance. That 55 TOPS/W chip mentioned above "can run the equivalent of 4 TOPS" in order to do "YOLOv3 at 30fps" "in under 20 mW". YOLOv3 is a real-time object detection algorithm.
So you can see how power sipping inference technology would be extremely useful for a surveillance state. Just make tiny, cheap cameras that store facial signatures or facial snapshots with timestamps indefinitely, instead of recording a limited number of hours/days of video. Power the camera and AI chip combo with solar panels or ambient energy [networkworld.com]. Deploy millions of them everywhere, even in small towns, download from them as needed. Full video footage will be more useful in some situations, but this could be a cheap way to track the movements of every single person.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 1, Touché) by Anonymous Coward on Sunday July 12 2020, @06:56PM (2 children)
"Deploy millions of them everywhere, even in small towns"
Hey, thanks for giving them ideas.
(Score: 3, Insightful) by takyon on Sunday July 12 2020, @07:27PM (1 child)
It's just the natural progression of the CCTV craze. Less human effort needed to make it useful, with more real-time AI features. Maybe all participating cameras within a 100 mile search radius can be sent your signature and ping back only if you (or someone else on the shit list) are seen. Cameras in sensitive high-traffic areas will gain features like "emotion recognition" [standard.co.uk].
Ultra low power capabilities will make it easier to deploy in small towns or rural areas, possibly with no need for access to the power grid. So villages in China, Alaska, etc. highway markers, wherever needed. Nobody will be immune. And they should definitely be attached to all 5G microcells.
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 3, Insightful) by turgid on Sunday July 12 2020, @10:03PM
We are the dead.
I refuse to engage in a battle of wits with an unarmed opponent [wikipedia.org].