GDDR7 — the next generation of video memory. Prospects and features of the technology
GDDR7 is a new generation of video memory for graphics adapters, such as Nvidia Blackwell RTX 50. In the next few years, this technology will be used in modern products from various manufacturers, gradually replacing GDDR6 and GDDR6X. The standard has many interesting features, but you should not expect new video cards with such memory to appear quickly. Details are under the cut.
Faster and more productive
First, let's remember the technologies that are used now.
GDDR6 Introduced in 2018, it was first used in Nvidia's RTX 20-series and AMD's RX 5000-series GPUs. There's also GDDR6X memory, used exclusively by Nvidia on its higher-end RTX 30- and 40-series GPUs.
GDDR7 SDRAM (Graphics Double Data Rate 7 Synchronous Dynamic Random-Access Memory) — memory type graphics card (SGRAM), specified in the JEDEC high-bandwidth standard. It uses a “double data rate” interface designed for use in video cards, game consoles, and high-performance computing. It is a type of GDDR SDRAM (graphics DDR SDRAM), the successor to GDDR6(X/W).
This type of video memory was first announced by Samsung back in 2022. The first prototypes also appeared then. But the main work on bringing the technology to market was already underway in 2023 and 2024:
July 18, 2023 – Samsung announced the first generation of GDDR7, which can reach speeds of up to 32 Gbps per pin (33% higher bandwidth per pin compared to GDDR6's 24 Gbps per pin). It offers 40% more bandwidth (1.5 TB/s) than GDDR6 (1.1 TB/s) and is 20% more power efficient. The company will use epoxy molding compound (EMC) as a packaging material, along with optimizations to the chip architecture, which will reduce thermal resistance by 70%. Later in the Q&A, Samsung mentioned that the memory will be manufactured using the D1z node (15-14 nm equivalent) and will run at 1.2V. A lower-clocked 1.1V version will also be available at some point in the future after the 1.2V version is released.
March 5, 2024 – JEDEC officially published the GDDR7 graphics memory standard.
It is known that GDDR7 SDRAM is based on the PAM-3 signaling scheme instead of NRZ (PAM-2). PAM-3 uses three bits per two cycles, while NRZ only uses one bit per cycle. GDDR7 SDRAM will also be manufactured using the 1ß node (equivalent to the 12-10 nm technology node). This will be the last DRAM manufacturing process to use deep ultraviolet (DUV) lithography tools.
Where will GDDR7 be used?
So far, there are not many official statements from graphic adapter manufacturers. The only thing that is known is that Nvidia Corporation will be one of the first to use the new technology. In particular, for the Nvidia Blackwell model range – the first representatives should appear in the fall of 2024.
AMD's RDNA 4 graphics cards were expected to use the new memory type as well. But there are now indications that RDNA 4 will stick with GDDR6 memory.
Be that as it may, Micron believes that the transition to the new video memory will provide a 30% increase in gaming performance, both with raster graphics and ray tracing. The company compares it with the previous generation – GDDR6, including GDDR6X. According to the company's representatives, such a significant increase is possible due to the increased speed of memory chips. Thus, 32 Gbps per contact is expected, which is 12 Gbps faster than regular GDDR6, which provides a speed of 20 Gbps. If this is true, the gaming industry can reach a new level. Although, of course, we need to wait and see if the words of Micron representatives are confirmed.
At the same time, the company brings other impressive numbers. For example, GDDR7 has a 60% higher bandwidth than GDDR6. It is also 50% more energy efficient than the previous generation of memory. In addition, memory latency is 20% lower, which gives a boost to machine learning and AI image generation.
According to Micron engineers, the test platform with 12 GDDR7 chips connected via a 384-bit bus demonstrates a throughput of more than 1.5 TB/s. But the GeForce RTX 4090 with GDDR6X chips with a speed of 21 Gbps per contact offers a throughput of 1.08 TB/s.
What's the bottom line? First, memory will be produced with a data transfer rate of 28 and 32 GT/s, then this figure will be increased to 40 GT/s. Micron, Samsung, and SK hynix are expected to supply the memory to the market.
