In 2021, Google and Antmicro introduced a platform for testing DRAM memory chips against the unfortunate side effect of the physical shrinking of memory chips—the Rowhammer vulnerability. The platform was developed to propose a radical improvement over the “security through obscurity” approach that was predominant in the industry; as both Antmicro and Google believe that the open source approach to mitigating security threats is a way towards accelerating developments in the field.
The framework was originally developed in the context of securing consumer-facing devices, using off-the-shelf Digilent Arty (DDR3, Xilinx Series7 FPGA) and Xilinx ZCU104 (DDR4, Xilinx UltraScale+ FPGA) boards, then followed by a dedicated open hardware board from Antmicro that allowed work on custom LPDDR4 modules. The framework has since helped discover a new attack method named Blacksmith and continues to provide valuable insights into how the security of both edge device and data center memory can be improved.
In constant development since then, the project has welcomed two more major elements to the ecosystem in order to enable testing of DDR4 Registered Dual In-Line Memory Modules (RDIMM)—commonly used in data centers as well as the newer DDR5 standard and continues to provide useful data.
Memory testing for data center use cases
To extend the Rowhammer tester support from consumer-facing devices to shared-compute data center infrastructure, Antmicro developed the data center DRAM tester board. We adapted this open source hardware-test platform from the original LPDDR4 board to enable Rowhammer and other memory security experiments with DDR4 RDIMMs using a fully configurable, open source FPGA-based DDR controller.
The data center DRAM Xilinx Kintex-7 FPGA based test board features:
- DDR4 RDIMM connector
- 676 pins FPGA (compared to the 484 for the LPDDR version)
- RJ45 Gigabit Ethernet
- Micro-USB console
- HDMI output connector
- JTAG programming connector
- MicroSD card slot
- 12 MBytes QSPI Flash memory
- HyperRAM—external DRAM memory that can be used as an FPGA cache
It’s worth mentioning that the RDIMM DDR4 memory (as opposed to the custom LPDDR4 modules designed for the original project) are generic and available off-the-shelf. This makes it easier for security researchers to get started with data center memory security research compared to edge devices using LPDDR.
The Data Center DRAM Tester board design has now been upgraded into revision 1.2, which brings new features for implementing even more complex DRAM testing scenarios. The 1.2 boards support a Power over Ethernet (PoE) supply option so the board can act as a standalone network device with data exchange and power-cycling done over a single Ethernet cable. This simplifies integration of the board in DRAM testing clusters and custom runners capable of doing hardware-in-the-loop testing.
The new revision of the board will support hot-swapping of the DRAM module under test, which should speed up testing of multiple DRAM modules without the need to power-cycle the tester. Finally, the new revision of the board will include power-measurement circuitry so it will be possible to compare the peak and average power consumption of DRAM while working with different DRAM refresh scenarios.
We are also working on a custom enclosure design suitable for desktop and networked installations.
Extending open source testing to DDR5
With DDR5 quickly becoming the new standard for data center memory, Antmicro and Google’s Platforms teams also set out to develop a platform capable of interfacing with DDR5 memories, again directly from a low-cost FPGA without a dedicated hard block. The resulting DDR5 tester platform follows the structure of the data center DDR4 tester, while expanding on functionality of the Serial Presence Detection, which monitors the power supply states and system health, or adjusting the circuitry for a nominal IO voltage of 1.1V.
Data center DRAM testing is part of Google’s and Antmicro’s belief in security through transparency. Both hyperscalers and a growing number of organizations who operate their own data centers increasingly embrace this perspective, and there is great value in providing them with a scalable, customizable, commercially supported open source platform that will help in collaborative research and mitigation of emerging security issues.
Rowhammer attacks, security threats, and countermeasures remain an active research area. With Google, Antmicro continues to adjust the Rowhammer test platform to most recent developments, opening the way for researchers and memory vendors to more sophisticated testing methods to enable testing of state-of-the-art memories used in data centers. This work stems from and complements other open source activities the companies jointly lead as members of RISC-V International and CHIPS Alliance, aimed at making the hardware ecosystem more open, secure and collaborative. If you’re interested in open source solutions for DRAM security testing and memory controller development, or more broadly, FPGA and ASIC design and verification, don’t hesitate to reach out to Antmicro at contact@antmicro.com.
By Michael Gielda – Antmicro