» Current | 2020 | 2019 | 2018 | Subscribe

Linley Newsletter

Upmem Embeds Processors in DRAM

August 27, 2019

Author: Tom R. Halfhill

Organic brains don’t partition thinking and memory in separate hemispheres, and some computer scientists think electronic brains shouldn’t either. In-memory computing is a frontier technology that relieves the CPU-DRAM bottleneck by integrating both functions on the same chip. The main obstacles are physical: DRAM is difficult to integrate in a conventional logic process, and logic performs poorly in a conventional DRAM process.

French startup Upmem tackles this dilemma by using massive parallelism to overcome the physical limitations of embedding logic in a DRAM chip. Its unique data processing units (DPUs) run at only 500MHz, but each can execute 24 hardware threads. Each chip has eight DPUs and 4Gbits of DRAM. Sixteen chips populate a memory module that’s compatible with industry-standard DIMM slots. Thus, each processor-in-memory (PIM) module has 128 DPUs and 8GB of DRAM, and it can work alongside the usual DIMMs or replace them altogether.

A fully populated Intel Xeon system with six DRAM channels could amass 1,536 DPUs, 36,864 threads, and 96GB of main memory. Even larger configurations are possible using AMD or IBM servers that have eight DRAM channels—or an IBM Power9 scale-up system with 32 channels. Upmem says an AMD or IBM system with 2,048 DPUs, 49,152 threads, and 128GB of main memory can execute one trillion operations per second (TOPS). To verify this claim, it employed simulations and sample chips manufactured at a Taiwanese DRAM fab since 2Q19. The first production-grade modules are scheduled to ship in October, followed by volume production in 1Q20.

Upmem presented its case for in-memory computing at the recent Hot Chips conference. The company claims its DPUs will typically deliver 20x better performance on memory-intensive workloads and 10x better energy efficiency than a standard x86 server. They won’t replace GPUs or FPGAs for most scientific workloads, however, because the DPUs lack floating-point hardware. They focus mainly on pattern matching, such as database index searches, big-data graph analytics, and some networking functions.

Subscribers can view the full article in the Microprocessor Report.

Subscribe to the Microprocessor Report and always get the full story!

Purchase the full article


Linley Spring Processor Conference 2020
Coming April 7-8, 2020
Hyatt Regency, Santa Clara, CA
Register Now!
Linley Fall Processor Conference 2020
Coming October 28-29, 2020
Hyatt Regency, Santa Clara, CA
More Events »


Linley Newsletter
Analysis of new developments in microprocessors and other semiconductor products
Subscribe to our Newsletter »