Intelligent Memory with Memory Expansion and NDP 

MX1

Intelligent Memory with 

Memory Expansion and NDP

MX1

CXL-Based
Memory Scalability 

MX1 supports CXL 3.0, allowing main memory expansion up to 1TB by connecting four channels of 256GB DDR5 DIMMs without increasing the number of CPU memory channels. 

This reduces unnecessary data replication and movement, enhancing overall system efficiency for applications requiring large-scale data processing. 

Additionally, it supports bandwidth expansion up to 128GB/s through the PCIe Gen6 interface.

Enhanced Data Processing Efficiency with Near Data Processing Technology 

MX1 is equipped with a multi-core processor,
performing parallel offloading tasks (Near Data Processing, NDP) near the data, thus improving data processing speed and efficiency.

NDP technology minimizes latency from data movement across interfaces and significantly reduces TCO for applications requiring large-scale data processing.

Proven Performance through FPGA Prototypes 

MX1 FPGA prototype has demonstrated a 46% reduction in query processing time compared to server CPUs in database acceleration, with potential reductions up to 95% in ASIC
(based on TPC-H benchmarks) 

Maximized Performance
with Cutting-Edge Process Technology 

MX1 utilizes Samsung Foundry's advanced 4nm process, maximizing power efficiency and performance. 

Strong RAS feature 

- Double die Chipkill Correction 

* RAS : Reliability, Availability, Serviceabiltiy

It supports multi-bit/multi-die DRAM ECC (Error Correction Code) and Chipkill, preventing critical system errors caused by various issues, and includes SSD RAID functionality for enhanced reliability. 

Applications

LLM VECTOR DATABASES

  • Recent LLMs utilize vector databases to retrieve updated information after training.
  • To curb the rapid increase in model size, vector databases are expected to be utilized more intensively.
  • The acceleration of vector databases in memory can play a crucial role in the advancement of LLMs. 

SCALE-OUT DATABASES

  • A large volume of data needs to be processed to create value from it even before AI training/interference.
  • Scale-out database clusters like Spark, Databricks, Snowflake are extensively used in ETL. These clusters typically consist of numerous servers.
  • By offloading the analytics query engine to computational memory, we could significantly reduce the cluster size. 

GRAPH DATABASES

  • Graph databases are extensively used in social networks handling enormous amounts of data based on nodes and relationships.
  • Graph algorithms mostly involve traversing the relationships between nodes. The key is to traverse pointers in parallel.
  • Many small cores with memory-optimized architecture are much more suitable for handling pointer traversing than CPUs. 

CXL-Based
Memory Scalability

MX1 supports CXL 3.0, allowing main memory expansion up to 1TB
by connecting four channels of 256GB DDR5 DIMMs
without increasing the number of CPU memory channels. 

This reduces unnecessary data replication and movement,
enhancing overall system efficiency for applications
requiring large-scale data processing. 

Additionally, it supports bandwidth expansion up to 128GB/s
through the PCIe Gen6 interface. 

Enhanced Data 

Processing Efficiency 

with Near Data Processing Technology 

MX1 is equipped with a multi-core processor,
performing parallel offloading tasks (Near Data Processing, NDP)
near the data, thus improving data processing speed and efficiency. 

NDP technology minimizes latency from data movement
across interfaces and significantly reduces TCO for applications
requiring large-scale data processing. 

Proven Performance 

through FPGA Prototypes 

MX1 FPGA prototype has demonstrated a 46% reduction
in query processing time compared to server CPUs in database acceleration, with potential reductions up to 95% in ASIC
(based on TPC-H benchmarks) 

Maximized Performance
with Cutting-Edge Process Technology

MX1 utilizes Samsung Foundry's advanced 4nm process, maximizing power efficiency and performance. 

Strong RAS feature - Double die Chipkill Correction  

* RAS : Reliability, Availability, Serviceabiltiy

It supports multi-bit/multi-die DRAM ECC (Error Correction Code) and Chipkill,
preventing critical system errors caused by various issues, and includes SSD RAID functionality for enhanced reliability. 

Applications

LLM VECTOR DATABASES

  • Recent LLMs utilize vector databases to retrieve updated information after training.
  • To curb the rapid increase in model size, vector databases are expected to be utilized more intensively.
  • The acceleration of vector databases in memory can play a crucial role in the advancement of LLMs. 

SCALE-OUT DATABASES 

  • A large volume of data needs to be processed to create value from it even before AI training/inference.
  • Scale-out database clusters like Spark, Databricks, Snowflake are extensively used in ETL. These clusters typically consist of numerous servers.
  • By offloading the analytics query engine to computational memory, we could significantly reduce the cluster size. 

GRAPH DATABASES 

  • Graph databases are extensively used in social networks handling enormous amounts of data based on nodes and relationships.
  • Graph algorithms mostly involve traversing the relationships between nodes. The key is to traverse pointers in parallel.
  • Many small cores with memory-optimized architecture are much more suitable for handling pointer traversing than CPUs. 

Xcelerate Your
Intelligence


Company Registration Number : 710-81-02837

Address : 20, Pangyoyeok-ro 241beon-gil, Bundang-gu,
                 Seongnam-si, Gyeonggi-do, Republic of Korea

CEO : Jin Kim

© 2024 MetisX | All Rights Reserved

Xcelerate
Your Intelligence


Company Registration Number : 710-81-02837

Address : 20, Pangyoyeok-ro 241beon-gil, Bundang-gu,
                 Seongnam-si, Gyeonggi-do, Republic of Korea

CEO : Jin Kim

© 2024 MetisX | All Rights Reserved