UPMEM, the developer of an ultra-efficient, scalable and programmable silicon-based acceleration solution, released today the first silicon-based PIM benchmarks. UPMEM PIM modules (DDR4-compatible DIMMs embedding UPMEM PIM DRAM chips), while running in standard Intel x86 Xeon servers, populating most of the memory module slots, deliver breakthrough performance for widespread performance-critical applications, being ten times more efficient than standard application servers.
The UPMEM Processing-in-Memory (PIM) allows compute to be off-loaded directly into the memory chips where the data transits or sits for processing data-intensive applications. UPMEM dramatically reduces data movement while leveraging the existing server architecture and memory technology for its PIM accelerators.
Early results on genomics mapping and analysis or index search validate the overwhelming potential of PIM for improving application servers’ speed and energy efficiency by over one order of magnitude. Disruptive efficiency will help the growth and promises of genomics-based healthcare. Offloading the full human genome analysis (mapping and variance analysis) for a GATK compliant analysis proves more than ten times faster if the Xeon server is equipped with UPMEM PIM modules instead of the same amount of standard DRAM. As most of the DNA fragment analysis versus reference genome happens on the PIM-DRAM chips, additional data bandwidth and processing power are fully leveraged, resulting in six times less energy consumption per throughput.
Searching through massive data indexes has always been a race for time and efficiency: adding an order of magnitude of speed and efficiency, as PIM provides, will open tremendous benefits to the service providers and end users. By sending the request to the hundreds of PIM-DRAM units in parallel, instead of relying on the main CPU, a server with PIM modules produces 11 times more throughput per second than the same server with the same amount of DRAM. Smart parallelization of search also allows much better latency (up to tens of ms instead of hundreds) to get the result back on average 35 times faster. The energy consumed at server level is six times less per throughput without any hardware modification, and can be extended.
“The early benchmarks reflect the benefits that PIM can bring when stopping most of the off-chip data movement between the memory and the processing cores, while unleashing the available data bandwidth,” says CTO and Co-Founder Fabrice Devaux. “We are pleased to see the interest we have received from application owners and developers, as well as enterprises, since our launch in 2019. We expect dozens of evaluations from partners and large research labs to show market adoption and potential, in all geographies, with a larger percentage in the U.S. market.”
UPMEM is bringing to market an ultra-efficient, scalable and programmable PIM technology that allows drastic reduction of data movement in the computing node for data-intensive applications in the data center and at the edge. UPMEM was founded in 2015, with headquarters in Grenoble, France, and a network of partners from Asia to the U.S. The team combines both entrepreneurial and technical expertise, ranging from processor architecture, software design, and low-level application workloads. Among the UPMEM investors are Western Digital, Partech, C4 Ventures, Supernova Invest and the French Tech Innovation Agency.