Manoj Roge, VP of Strategic Planning & Business Development
November 12, 2018
Micron Technology announced today that its high-performance GDDR6 SDRAM (in volume production since June) will be the high-speed memory of choice for Achronix’s yet-to-be-announced, next-generation FPGA family, making Achronix the world’s first FPGA vendor with announced GDDR6 support. Achronix has worked closely with Micron on GDDR6 compatibility and this announcement of Achronix’s next-generation FPGA family support is proof of this extremely close partnership. Next-generation Achronix FPGAs will be fabricated on TSMC’s industry-leading 7nm process node.
Achronix’s next-generation FPGAs will fully exploit the GDDR6 SDRAM’s extreme memory bandwidth because they will incorporate as many as eight hard GDDR6 memory interfaces — each with as many as thirty-two 16 Gbps lanes — delivering a maximum, aggregate memory bandwidth of 4 Tbps. That data rate is 5 times faster per lane than the fastest DDR4 SDRAM, which today is 3.2 giga-transfers per second. To achieve 4 Tbps of memory bandwidth with DDR4 ×8 (8-bit) memories running at 3.2 Gbps would require 160 memories, or with 72-bit DIMMs, a total of 18 ×72 interfaces, which is not practical.
Previously, that much memory bandwidth is only available from a pair of HBM 2 memory stacks mounted on a 2.5D interposer along with an FPGA die. However, eight 16 Gb GDDR6 SDRAMs (mounted on a conventional PCB assembly) deliver more total system memory capacity than two HBM 2 memory stacks, avoiding an expensive 2.5D or 3D assembly processes.
System designers can easily scale FPGA-based system performance by using from one to eight of Achronix’s next-generation FPGA’s GDDR6 memory interfaces — the performance of a suitably designed system can be scaled over a very wide performance range simply by adding GDDR6 SDRAMs. From a system design perspective, this capability gives engineers significant flexibility to develop systems that have exactly the right amount of memory capacity needed for their target application.
For example, a single 32-bit GDDR6 interface can be paired with one 8 Gb or 16 Gb GDDR6 memory, resulting in a memory bandwidth of 512 Gbps. Using all eight of the FPGA’s 32-bit GDDR6 interfaces creates a GDDR6 memory bank with either 64 or 128 Gb of total capacity (using 8 Gb or 16 Gb memories respectively), with a resulting memory bandwidth of 4 Tbps.
Although the “G” in GDDR6 stands for graphics, and the initial use for GDDR6 memory has been in video cards and graphics processors, the GDDR6 memory’s 16 Gbps/lane data bandwidth and the 4 Tbps memory bandwidth enabled by the FPGA’s eight GDDR6 interfaces make high-speed GDDR6 memory an excellent choice for many FPGA-based applications that require vastly increased memory bandwidth — applications including artificial intelligence and machine learning (AI/ML), blockchain, networking and automotive.
Recently, Micron’s Director of its Global Graphics Memory Business posted a corporate commitment on Micron’s Memory blog about their goals: “…to continue establishing GDDR6 memory as the DRAM solution for other high-performance applications other than gaming.” Today’s announcement by Micron underscores that commitment — and Achronix’s commitment as well.