The need

Every business is digital, and rely on digital resources to capture, transform and analyse data : Understand customers, profile supply chain, optimise production, automate processes, etc. At the same time, internet players need to scale their infrastructure to support new users, and improve economical efficiency of their services. On server front, Big Data and other data-intensive applications are hitting the “Memory wall”: The overall performance is limited by the memory throughput, and the wall is getting higher with DRAM reaching 1TB per server.

In the datacenter, a new class of data-centric processing solutions is required to turn massive data computings into a fast and cost efficient operations.

Technology

A pool of programmable co-processors to accelerate data processing

UPMEM very compact RISC processor can be implemented in ASIC designs to complement main processor with a set of programmable engines, optimized for data processing, and exposing a unified interface. 

UPMEM lead use case is Processing-In-Memory for the datacenter, the combination of UPMEM processor with DRAM :

Data driven

 

UPMEM processing elements are integrated into the DRAM chips to leverage best-of-art memory density, faster access, and an undisputed bandwidth.

easy-integrationEasy integration

 

UPMEM solution is packaged on DIMM modules to be smoothly integrated into standard servers, and delivered along with an end-to-end SDK.

truly-scalableTruly Scalable

 

Hundreds of processing elements can be combined to compute hundreds of Giga bytes in parallel, in a single or multi-tenant architecture.

Applications

real-timeReal Time Analytics

Placing ads on a website; Detecting credit card frauds; Analysing incoming News; and more widely On-Line Analytical Processing and On-Line Transaction Processing rely on massive data sets, and require a responsive computing architecture.

 

puzzlesPattern matching

Comparing DNA sequences against a bank of genes; Associating incoming network packets with flows; Scanning files to match a regular expression; When the size of reference data is much bigger than inputs, memory bandwidth is the dimensioning criteria.

 

 

data-baseDatabase

Database server is probably the most common element at enterprise datacenter. Innovative technologies such as In-Memory DataBase and In-Memory DataGrid, are addressing the performance issue, but they need an appropriate hardware to maintain performance while scaling.

 

 

neuronesArtificial Intelligence

Neural Networks are good at interpreting real life information. This technology is successful to identify objects in pictures, read  handwritten text, understand natural language, activate characters in video games. AI has a natural fit with the distributed computing architecture proposed by UPMEM.

 

Company

About us

UPMUPMEM - profile iconEM was founded in January 2015 with a vision to boost data-intensive applications in the datacenter. We are developing a cost efficient and massively scalable Processing-In-Memory solution to address the Memory Wall issue, without disrupting the ecosystem.

We love to think forward and build value through innovations.We are a small and agile team with experiences at companies like STMicroelectronics, VMware, Elsys Design and Trango Virtual Processors. We love those who think out of the box and push us beyond our limits. Our domain of expertise is at the hardware/software interface level. Anything above is Open to our partners and customers through APIs : We are building a means to an end.

Partners

 

CEA tech

Dolphin Integration

Inria

University of Kaiserslautern

Awards

 

i-Lab Award

HP Enterprise

Netva

Tremplin 2016 Award

Contact

Want to know more ? Don't be shy, share your thoughts and questions!