LPDDR6: what the new memory changes for l'IA

#What Samsung announced

Samsung published the details of its LPDDR6 memory on September 23, 2026: up to 114 GB/s of memory bandwidth, achieved through improved I/O speed per pin and a wider I/O architecture. The company also announced up to a 21% improvement in energy efficiency over the previous generation, LPDDR5X, thanks to Dynamic Voltage and Frequency Scaling (DVFS) and Dynamic Efficiency Mode (DEM).

Notable fact: Samsung validated 16 GB LPDDR6 running at 10.7 Gbit/s per pin on Qualcomm’s brand-new Snapdragon 8 Elite Extreme Gen 6—a mobile-platform industry first according to Samsung. More details are announced for the Snapdragon Summit 2026, which follows this page’s publication.

#Capacity and bandwidth: two distinct things

114 GB/s is a bandwidth figure, not a capacity figure: it describes how fast data moves between memory and the processor, not how much memory is available. The module tested by Samsung is 16 GB—a modest capacity, consistent with mobile use, and unrelated to the 128 GB of unified memory already found in some current PC APUs. Don't confuse the two: faster memory with limited capacity can't load a larger model; it only speeds up access to data already in memory.

#Where does the 114 GB/s figure come from?

The JEDEC JESD209-6 standard, which officially defines LPDDR6, sets a data rate of 10,667 to 14,400 MT/s, equivalent to effective bandwidth of approximately 28.5 to 38.4 GB/s according to Tom's Hardware; the article does not specify the width to which this figure refers. The standard also changes the channel width from 16 bits (LPDDR5X) to four 24-bit subchannels, for more concurrent access and lower latency.

A comparison is warranted, but this is our deduction, which Samsung does not confirm: interpreted as throughput per channel, the 28.5 GB/s floor multiplied by 4 channels gives exactly 114 GB/s, the figure highlighted by Samsung. If this interpretation is correct, the announcement does not exceed the standard's baseline: it adds together multiple channels, a configuration already common in mobile memory. The upper end of the JEDEC range (38.4 GB/s per channel) would suggest up to 153.6 GB/s in a 4-channel configuration—a theoretical figure that no source confirms for an actual product to date.

This table properly puts the Samsung figure in context: at the JEDEC multi-channel floor, far below the 256 to 819 GB/s of already-commercialized PC or Mac platforms with unified memory. It also shows that an AMD rumor already points to a concrete PC use for LPDDR6 (Medusa Halo, unconfirmed), with a figure well above Samsung’s validated 16 GB mobile module—because it would be a wider-bus configuration, not the same chip.

#LPDDR6 for local AI: the current state of play

#Why bandwidth matters for an LLM

Samsung states this explicitly in its own document: during inference, the system continuously accesses the model weights and KV cache data to generate each response. As models grow and conversations get longer, moving data between memory and the processor becomes increasingly decisive. This is exactly the factor limiting the generation throughput of a local LLM: the theoretical ceiling is estimated by dividing bandwidth (GB/s) by the size of the active weights (GB)—an estimate, never a guaranteed measurement.

Higher memory bandwidth, without changing capacity, mainly benefits devices where memory is already the primary bottleneck—typically configurations without a dedicated GPU or with unified memory, where the CPU and model compete for the same memory channel.

#When will LPDDR6 arrive on a PC or Mac?

No Samsung or Qualcomm source consulted announces a timeline for a PC or Mac platform equipped with LPDDR6. The cited validation concerns a Qualcomm mobile SoC, not a CPU or Apple Silicon chip. Only one PC lead exists to date, and it remains an unconfirmed leak from the manufacturer involved: Tom's Hardware reports that an upcoming AMD variant, “Medusa Halo” (2027), could adopt LPDDR6 on a 256-bit bus at 14,400 MT/s, for approximately 460.8 GB/s — a figure far higher than Samsung's 16 GB module because it would be a wide PC-class bus, not a mobile chip. AMD has confirmed neither the Medusa Halo name nor this memory choice.

This same normalization process by JEDEC ensures that multiple manufacturers (Samsung, as well as SK hynix and Micron) produce memory that interoperates with the same memory controller once the standard is published and adopted by chip manufacturers.

#Should you wait for LPDDR6 before building an LLM machine?

Need an LLM machine now

LPDDR6 is currently used in no PC or Mac. The unified-memory and dedicated-GPU options already available remain the only measurable choices.

Buying a high-end smartphone in 2027

LPDDR6 could appear there first, benefiting lightweight embedded AI more than large local LLMs.

Tracking a future PC platform (Medusa, Nova Lake…)

Check at the time of each platform's official announcement whether it actually adopts LPDDR6, without assuming it based solely on this Samsung announcement.

Decision based on the figure 114 GB/s

This is the bandwidth figure advertised by Samsung, not the throughput of a complete machine, a tokens/s measurement on a real LLM, or usable memory capacity for large models.

• VRAM: what is it, and how much do you need for AI?
• Ryzen AI Max+ 395: unified memory already available
• Local LLMs without a GPU: what a CPU can do today
• VRAM/memory calculator for a model
• Source: Samsung LPDDR6 technical announcement
• Source: JEDEC, the memory standards organization
• Source: Tom’s Hardware on the JEDEC LPDDR6 standard