GDDR6 vs GDDR6X: What the Memory Architecture Difference Actually Means

GDDR6X achieves higher bandwidth per pin than GDDR6 by using a fundamentally different signaling method. That bandwidth advantage comes with measurable tradeoffs in heat and power consumption that affect which cards run hot at the memory junction even under normal loads.

When comparing GPUs, GDDR6 and GDDR6X appear in spec sheets as if they are minor revisions of the same thing. They are not. GDDR6X uses a different electrical signaling approach called PAM4 that lets each physical pin carry more data per clock cycle, increasing effective bandwidth without requiring a wider memory bus or higher clock frequency. The penalty is higher power consumption per chip and elevated junction temperatures under load. Understanding the difference helps explain why two cards with similar spec-sheet bandwidth numbers behave differently in thermal monitors.

How GDDR6 Signals Work (NRZ)

Standard GDDR6 uses Non-Return-to-Zero (NRZ) signaling, where each pin represents one of two voltage states per clock cycle: high or low, encoding a single bit. This is the signaling approach used across virtually all consumer DRAM types including DDR5 and LPDDR5. At 16 Gbps per pin, a GDDR6 chip with a 256-bit bus (32 pins per side, 8 chips) delivers 512 GB/s of theoretical peak bandwidth.

GDDR6 at equivalent clock speeds draws relatively modest power per chip, and junction temperatures in the 80 to 95 degree Celsius range are typical under sustained loads, depending on cooler design and airflow over the VRAM chips.

GDDR6X and PAM4 Signaling

GDDR6X replaces NRZ with Pulse Amplitude Modulation 4 (PAM4). Instead of two voltage states per clock cycle, PAM4 uses four distinct amplitude levels. Each clock cycle therefore encodes two bits instead of one, doubling the data rate per pin without increasing the physical clock frequency. At the same 16 Gbps base clock as GDDR6, a GDDR6X implementation operates at an effective 21 Gbps per pin or higher because the amplitude levels themselves can be pushed.

In practice, GDDR6X as implemented on RTX 30 and RTX 40 series GPUs runs at 19 to 23 Gbps per pin depending on the specific card, versus GDDR6 implementations at 16 to 18 Gbps per pin on comparable AMD and NVIDIA mid-range products. On a 384-bit bus, that difference translates to meaningfully higher peak bandwidth: the RTX 4090 with GDDR6X on a 384-bit bus reaches approximately 1008 GB/s theoretical peak bandwidth.

The Thermal Tradeoff

PAM4 requires finer voltage discrimination between four amplitude levels rather than two. The receive circuitry must distinguish small differences in signal voltage precisely, which demands more power in the analog front-end of each memory chip. GDDR6X chips run at higher junction temperatures than GDDR6 chips at equivalent workloads and clock speeds. This is not a flaw in the design; it is an inherent consequence of the PAM4 approach at current process nodes.

The RTX 3090 Ti and RTX 4090, both using GDDR6X, regularly show memory junction temperatures of 100 to 110 degrees Celsius under sustained full load in well-cooled cases. By contrast, AMD RX 7900 XTX cards using GDDR6 on a 384-bit bus typically show memory junctions in the 80 to 90 degree range. The GDDR6X cards have more total bandwidth, but the GDDR6 cards run cooler at the memory junction for equivalent thermal load.

Which GPUs Use Which Memory Type

NVIDIA has used GDDR6X on its high-end desktop cards since the RTX 3080: RTX 3080, 3080 Ti, 3090, 3090 Ti, 4080, 4080 Super, 4090, and the 4070 Ti Super all use GDDR6X. The RTX 4070, 4060 Ti, and lower segments use standard GDDR6.

AMD uses GDDR6 across its entire RDNA 3 and RDNA 4 desktop lineup. The Radeon RX 7000 and RX 9000 series are uniformly GDDR6. AMD achieves competitive bandwidth primarily through wide bus widths: the RX 7900 XTX runs 384 bits, and the RX 9070 XT runs 256 bits at higher per-pin speeds rather than adopting PAM4.

Laptop GPU implementations typically follow desktop memory type assignments, though power and thermal constraints often result in lower effective clock speeds than the desktop counterparts, reducing the bandwidth gap in practice.

Memory Overclocking Behavior

GDDR6 overclocks differently from GDDR6X. On GDDR6 cards, memory frequency can often be pushed 200 to 500 MHz above stock speeds with acceptable error rates and a manageable temperature increase. On GDDR6X cards, the PAM4 signaling is already operating with tighter amplitude margins, and headroom before errors appear tends to be smaller in absolute terms. Modest GDDR6X memory overclocks of 100 to 200 MHz above stock are common; large overclocks on GDDR6X require more careful stability testing and monitoring of junction temperature, since heat accelerates the error rate at marginal voltage levels.

Practical Takeaways for Card Selection

If you are building a compute-heavy workstation where sustained memory bandwidth matters and thermal management is constrained, a GDDR6 card with a wide bus can deliver similar effective throughput at lower junction temperatures. If peak gaming bandwidth and raster/ray tracing performance per dollar is the priority and your case has good airflow, GDDR6X cards at the high end offer the highest bandwidth available in discrete GPU form. Both memory types are mature and reliable within their operating ranges; the choice should follow the workload and thermal envelope, not a preference for one signaling scheme over another.