Recent advances in artificial intelligence may appear revolutionary, but JEDEC is keeping an evolutionary approach for Graphics Double Data Rate (GDDR) standards, even as it’s being increasingly used for AI applications.

In a briefing with EE Times, Michael Litt, chair of the JEDEC Solid State Technology Association’s GDDR subcommittee, said the GDDR7 standard continues the generation-to-generation tradition of double the bandwidth and double the capacity while keeping a lid on power consumption. “We did look at evolutionary and revolutionary approaches,” he said. “That’s always the first thing we do.”

Litt said the path JEDEC chooses is heavily influenced by industry participants.

“We were looking to really maintain an evolutionary approach,” he said. “We have to support the markets that it supports today.” Graphics and gaming continue to be important segments for GDDR7, as well as increased use by AI applications, he said.

Even if a revolutionary jump ahead is possible, the industry would not likely be able to take advantage of it. In the meantime, Litt said, there have been walls that needed to be overcome as GDDR has evolved, including scaling the interface. “The success of JEDEC specs is they don’t try to bite off more than they can chew,” he said. “We try to stay as evolutionary as possible.”

The latest iteration of GDDR offers twice the bandwidth of its predecessor, reaching up to 192GB/s per device. GDDR7 doubles the number of independent channels, from two in GDDR6 to four.

GDDR adds PAM

It’s also the first JEDEC standard DRAM to use the pulse-amplitude modulation (PAM) interface for high-frequency operations. Using a PAM3 interface improves the signal-to-noise ratio for high-frequency operation while enhancing energy efficiency. PAM3 also offers a higher data transmission rate per cycle, resulting in improved performance versus the traditional non-return-to-zero (NRZ) interface.

A new interface made sense for GDDR7, Litt said. “We were looking for something that had more headroom.”

GDDR7 addresses the industry’s need for reliability, availability and serviceability (RAS) by incorporating the latest data integrity features, including on-die error-correction coding with real-time reporting, data poison, Error check and Scrub, and command address parity with command blocking (CAPARBLK).

Litt said the RAS features also encompass security, which is also top of mind. “It’s going to become even more important.”

He said the primary focus is to keep the economics and the ecosystem that exist today, although a more HBM-like, on-package approach was considered.

Despite being in its third iteration, HBM remains a complex memory to implement, and it’s not cheap. And with companies like Micron Technology selling out of HBM3, GDDR can be a viable alternative for some AI workloads.

“We’re seeing more and more customers that are interested in evaluating their options,” Lou Ternullo, senior director of product marketing for silicon IP at Rambus, said in an interview with EE Times. GPU makers, AI workloads drive GDDR capabilities

AI demands are shaping the evolution of GDDR, Ternullo added. “The likes of Nvidia have taught us that a graphics processor can be used for much more than graphics.” One of the reasons GDDR has found uses beyond its initial target market is its ability to do matrix algebra, which helps GPUs handle AI workloads and computer-generated special effects.

GPU maker Nvidia wanted a faster, more reliable memory—hence, the adoption of PAM. Ternullo said transmitting data at super-fast rates means channel integrity becomes a bigger concern. “That’s where reliability comes in.”

Just as it does with HBM, Rambus provides the necessary IP that supports the GDDR ecosystem. The company recently announced what it said is the industry’s first GDDR7 Memory Controller IP. It supports 40-Gbps operation, providing 160-GB/s throughput for a GDDR7 memory device.

Ternullo said the reason its GDDR7 memory controller IP is available so close to the publishing of the JEDEC standard is because Rambus started working early with some key customers in the market.

He said the rapid pace of AI has customers wanting to make sure their system will perform over the long term, which includes evaluating whether HBM or GDDR7 will meet their bandwidth requirements throughout a product’s life cycle.

Ternullo said HBM is a difficult technology to wrap your head around, and as an IP provider, Rambus can help customers evaluate their options. “GDDR7 is another tool in our toolbox and gives customers another tool in their toolbox to figure out how are they going to meet the challenges and requirements of their systems,” he said.

By: DocMemory
Copyright © 2023 CST, Inc. All Rights Reserved