Embracing the exponential growth of data: towards the breakthrough of memory scaling
- 175 zettabytes of worldwide data are expected by 2025, while today’s memory technologies are not ready to handle this exponential growth.
- Intel acquires Habana Labs ($2B, Dec 2019), Nvidia acquires Mellanox ($7B, H1 2020), Facebook acquires Sonics (March 2019) to solve the problem in-house.
- Grovf offers revolutionary memory-scaling technology on top of Xilinx Accelerator cards, aiming to resolve modern computing drawbacks.
The greater majority of worldwide data has been created in the last few years, making data growth the most consistent and noteworthy tech trend. IDC predicts that by 2025, worldwide data will grow to 175 zettabytes, “Data growth, analysis, and use remain a hot topic in our lives that drives technological innovation”, says Dan Eaton, Director of Business Development, Database, and Data Analytics at Xilinx.
Shrinking transistors have powered 50 years of advances in computing until 2003 but then improvements in processor performance slowed down, and now processing power doubles every 20 years rather than 1.5 (the end of Moore’s law).
Since server performance was limited by the speed of access to memory (a.k.a. memory wall), the single power-hungry processor was replaced with several energy-efficient ones, as an alternative way to get more computing power. Thereby achieving maximum speedup possible, reconfigurable chips (known as FPGAs, GPUs) were deployed to achieve higher efficiency and accelerated computing.
However, processor-centric architecture with accelerator chips over the last decade brought its own drawbacks — redundant data duplication, prolonged communication among PCIs, highly limited memory per node, and thus reduced speed of access to memory and inefficient computing.
Nowadays, when data growth is outpacing the advancements of processors, memories barely keep up with neither of them and eventually become the main bottleneck of the computational system.
All these limitations of today’s computer design make us deeply rethink how we build compute and storage technologies, laying at the foundation of the IT industry for the past four decades.
Beyond processor-centric resolutions
Modern computer systems need to undergo a paradigm shift away from processor-centric architecture to the memory-centric one where memory plays the key role and literally everything revolves around it.
“In this distributed non-traditional architecture, memory stops being directly attached to the core processor and turns into a shared global resource, equally accessible by all compute parties”, says Millind Mittal, Fellow/VP at Xilinx. “Advances in this horizontal scaling approach make shared global memory accessible at lower latency and offer expanded capacity for a lot more memory support”.
Currently, various off-chip interconnects like CCIX, GenZ, OpenCAPI, UPI are essential components for a better deployment of the emerging memory technologies and ensure optimized communication between processors and accelerators.
These protocols in particular vary by processor compatibility (CCIX supported by ARM, OpenCAPI by IBM, UPI by Intel, etc.) as well as scalability and latency prioritization. For instance, deployment of CCIX makes the most of latency results though encounters processor scaling limitations, whereas with Gen-Z the latency is sacrificed in favor of scaling capacities.
With a critical need to efficiently handle vast amounts of data, many corporations seek to improve their technology performance by switching to memory-centric architecture.
In the first quarter of 2019, Facebook acquired Sonics to solve memory expansion and interconnect problems, while Intel acquired AI chipmaker Habana Labs for $2B. The 2020 acquisition season has officially been kicked off by NVIDIA with a $7B deal of Mellanox, for HPC scaling in data centers.
Grovf and Xilinx join for the next memory-centric architecture
Grovf and Xilinx are collaborating to develop Grovf MonetX — a technology that resolves the bottlenecks both in scalability and speed. The goal is to address the challenges of memories and off-chip interconnects. Thanks to its flexible and cost-effective architecture, this memory scaling technology is expected to be an attractive solution for corporate Big Data management.
MonetX combines CCIX specifications with a proprietary fast network protocol allowing to expand the memory by tens of TBs per server while successfully accelerating in-memory computing workloads at exceptionally low latencies.
The smart memory expansion platform is built on top of Xilinx’s Alveo U280 acceleration card and can be implemented in Arm systems through CCIX interconnect from one end and memory coherent low-latency network on the other end. This introduces a new NIC (Network Interface Card) concept that Xilinx calls Memory NIC or MNIC.
“Over the past few years, memory-centric architecture was the focus of interest in the sector, and now we finally see the first practical implementations. Through our joint solution with Xilinx, we will derive maximum from memory disaggregation, allowing servers to better support modern applications, run faster and be more efficient in their use of memory.’’, says Khachik Sahakyan, CEO of Grovf.
Currently, a few hyperscalers in the market are likely to pioneer in adopting the Grovf MonetX MNIC first-of-a-kind solution, and some details on this product PoC will be apparent soon.
Grovf is a memory expansion company building the world’s first Memory Network Card to coherently scale server memory to hundreds of TBs. The company helps to achieve increased memory utilization and thereby accelerate Big Data compute.
Xilinx is an American technology company that invented the field-programmable gate array (FPGA), programmable SoCs, and now, the ACAP. Their highly-flexible programmable solutions deliver the most dynamic processing technology across a wide span of industries and technologies — from consumers to cars to the cloud. The company enables rapid innovation with its adaptable, intelligent computing.