Samsung boosts LPDDR5X to 10.7 Gbps, ups efficiency and capacity for mobile and servers
Guess what it's great for? Go on, have a guess... 20 points if you muttered an abbreviation starting with A and ending in I
Samsung has revealed its upgraded LPDDR5X memory modules, which features improved performance, capacity, and efficiency.
The improvements were achieved thanks to moving to the semiconductor giant's 12 nm node process, as Samsung's original batch of LPDDR5X RAM was made on 14nm. Upgrading to 12nm, despite not being a particularly big step, made for a near 26 percent hike in performance, from 8.5 Gbps to 10.7 Gbps.
Samsung claims the kit's power efficiency is sizeably increased by 25 percent thanks to more optimal power variation depending on the workload (basically how modern CPUs and GPUs work) and longer periods of operating at low-power mode. Of course, moving to a slightly smaller node probably also played a role in Samsung's newer LPDDR5X modules becoming more efficient.
The new LPDDR5X chips bring big changes to capacity too, which is now 32 GB per chip. When Samsung launched its very first LPDDR5X modules in late 2021, the maximum capacity was 2 GB, and this limit was eventually raised to 16 GB by 2023. Samsung never pursued 24 GB chips, unlike Sk Hynix which released its own 24 GB LPDDR5X in August.
Devices are currently limited to a maximum of 32 LPDDR5X chips (or at least, we haven't seen more than 32 used at once), so the maximum memory capacity with LPDDR5X has been lifted from 512 GB to a full terabyte. This might seem like a nonsensical use case given that LPDDR has traditionally been aimed at laptops and smartphones, but LPDDR5X has already been used for Nvidia's high-profile Grace-Hopper GH200, which uses 30 LPDDR5X chips to provide the Grace CPU with 480 GB of memory.
- Samsung reveals buzzword-compliant DRAM ready for 5G, AI, edge, and metaverses
- Micron says it's first to QLC NAND with over 200 layers
- Intel's effort to build a foundry biz is costing far more – and taking longer – than expected
- Samsung snags $6.4B in CHIPS Act funds for Texas fabs
A future CPU-GPU device from Nvidia could technically go up to 1 TB, though the actual, usable amount of memory is likely to be 960 GB as GH200 has 32 memory chips, but two are disabled, likely for redundancy purposes.
Samsung's second generation LPDDR5X is a little far away, though. Per its own statement, mass production is only slated to begin in the second half of the year, which probably means this memory will be MIA for the entirety of 2024, or at best might debut close to the end of the year. ®