GDDR7 vs GDDR6 Graphics Memory: PAM3 vs NRZ, Bandwidth, Voltage, Reliability, and Compatibility

GDDR7 is the successor generation to GDDR6 graphics DRAM. The major link-level change is signaling: conventional GDDR6 uses NRZ signaling with two signal levels, while GDDR7 uses PAM3 with three levels. Micron describes PAM3 as encoding 1.5 times as much information per signal cycle as GDDR6 NRZ. That helps GDDR7 scale data rate without simply treating the interface as a faster version of the same electrical link.

A memory-generation label alone does not establish gaming FPS, ray-tracing performance, latency, power consumption, or overall GPU speed. A finished graphics processor combines compute resources, caches, memory capacity, bus width, memory-device data rate, firmware and software behavior. This comparison therefore explains the memory technology rather than declaring a GPU winner.

NRZ represents data with two signal levels. PAM3 uses three levels, commonly described as -1, 0 and +1, to carry more information per signal cycle. Micron says this gives PAM3 a 50% higher voltage margin than PAM4 while using lower encoder complexity, helping address signal-loss and frequency-scaling challenges at high graphics-memory data rates.

GDDR6X is a separate technology and should not be used as shorthand for GDDR6. Micron GDDR6X uses PAM4 signaling, whereas ordinary GDDR6 uses NRZ and GDDR7 uses PAM3. A comparison that says only “PAM replaces NRZ” loses an important distinction between these generations and implementations.

Memory vendors commonly state graphics-memory signaling as gigabits per second per pin. Total memory bandwidth is a different quantity and also depends on the width of the GPU memory interface. For a simple theoretical calculation, bandwidth in GB/s equals the per-pin data rate in Gb/s multiplied by bus width in bits and divided by eight.

That is why two GPUs using the same memory generation can have different bandwidth, and why capacity is another separate property. A wider bus or faster selected memory devices can change bandwidth; adding more gigabytes of VRAM does not automatically increase it. The Core Tech Tips VRAM-capacity guide covers capacity pressure separately.

Micron’s current GDDR7 product brief compares specific 16 Gb devices. In that table, its GDDR6 device is listed at a maximum 18 Gb/s per pin and 72 GB/s per component, while its GDDR7 device is listed at 32 Gb/s per pin and 128 GB/s per component. The same comparison lists NRZ versus PAM3 signaling and maximum operating voltage of 1.35 V versus 1.2 V.

Those numbers are Micron product-comparison figures, not universal ceilings for every GDDR6 or GDDR7 device. Samsung, for example, currently lists GDDR6 products reaching 24 Gb/s per pin and its GDDR7 portfolio reaching up to 40 Gb/s. Product generation, vendor, density, speed bin and implementation therefore need to be identified before applying an exact number.

Micron lists 1.2 V operation for the compared GDDR7 device versus 1.35 V for its compared GDDR6 device and advertises more than 50% better power efficiency for its GDDR7 implementation. Samsung has likewise published product-specific GDDR7 efficiency improvements. These are vendor claims about memory devices and stated comparison conditions, not measurements of complete graphics-card board power.

A GPU board also powers the graphics processor, voltage regulation, memory subsystem, fans and other circuitry. Vendors can spend improved memory efficiency on higher data rates or different product targets. It is therefore incorrect to infer a fixed reduction in total board power merely because a card uses GDDR7.

Micron lists expanded reliability, availability and serviceability features for its GDDR7 implementation, including on-die ECC, command/address parity, 9-bit CRC and hard post-package repair. These mechanisms help protect memory-device and link operation as data rates and signaling complexity increase.

Reliability features should not be converted into a claim that one graphics card is universally more reliable than another. Board layout, cooling, component selection, firmware, manufacturing quality and workload conditions remain product-level variables beyond the memory-generation name.

GDDR7 requires a compatible memory controller and a board designed for its signaling, power and physical implementation. Micron explicitly states that GDDR7 is not backward compatible with GDDR6 or GDDR6X. Graphics memory is soldered into a tightly integrated GPU board design rather than exposed as a user-replaceable DIMM standard.

A card designed around GDDR6 therefore cannot be treated like a motherboard where a newer memory module can simply be installed. The GPU memory controller, PCB routing, firmware and selected DRAM devices are engineered together for the shipping product.

When comparing graphics cards, record at least the memory generation, capacity, per-pin data rate and memory-bus width before calculating theoretical bandwidth. Then keep those facts separate from cache architecture and GPU compute capability. For example, Core Tech Tips records the RTX 5090 with GDDR7 and the RX 9070 XT with GDDR6, but that difference alone is not evidence for a cross-vendor performance verdict.

Use workload-specific, properly sourced benchmark evidence when the actual question is performance. Use memory specifications when the question is interface capability, bandwidth, capacity or compatibility. That separation avoids turning a real generational advance in graphics DRAM into an unsupported universal FPS claim.