Samsung Electronics’ SOCAMM2 / Captured from Samsung Electronics official web site
Samsung Electronics is searching for to regain its management within the reminiscence market with its newest low-power server reminiscence module, the small define compression connected reminiscence module 2 (SOCAMM2), by supplying samples to a serious buyer as a part of the ultimate stage earlier than mass manufacturing.
Samsung Electronics lately revealed that it started supplying its SOCAMM2 samples to a serious shopper firm. The corporate didn’t specify the shopper or particulars of the validation progress, however business officers assume the samples are buyer samples for Nvidia’s next-generation Vera Rubin synthetic intelligence (AI) platforms.
Chip samples are usually categorized into three phases: working die, engineering pattern and buyer pattern. A working die refers back to the earliest useful stage of a chip. An engineering pattern is an early preproduction model despatched to clients to confirm compatibility with their programs and merchandise. After reflecting buyer suggestions, a near-final model often called a buyer pattern is delivered for qualification. As soon as the qualification course of is accomplished, the chip goes into mass manufacturing.
SOCAMM2 is a module of low-power double information price (LPDDR) reminiscence chips, aimed toward chopping energy consumption to roughly one-third of standard DDR-based modules. In keeping with a Shinhan Securities report, making use of the most recent LPDDR5X to servers can cut back energy consumption by about 77 % in contrast with DDR5, whereas rising bandwidth by 35 %.
Whereas high-bandwidth reminiscence (HBM) chips are often mounted on the graphics processing unit (GPU) for sooner information processing, SOCAMM2 is often mounted subsequent to the CPU and focuses on dealing with massive information volumes with larger energy effectivity. On this setup, HBM helps computing acceleration, whereas SOCAMM2 is chargeable for enhancing total system-level energy effectivity.
The semiconductor business is now working to arrange a normal for SOCAMM2, and an rising variety of chip builders are being attentive to the module’s energy effectivity as AI service corporations scramble to cut back the large quantity of electrical energy consumed by information facilities.
In keeping with Samsung, its 192-gigabyte SOCAMM2 is 57 % smaller than standard DDR-based server modules and delivers greater than twice the bandwidth whereas consuming at the least 55 % much less energy. The corporate famous that this permits steady efficiency even below intense AI workloads and presents excessive design flexibility, permitting it to be extensively deployed throughout a spread of AI programs.
Not like conventional soldered LPDDR options, SOCAMM2 allows simple reminiscence upgrades or replacements with none mainboard modification, serving to system directors reduce downtime and dramatically cut back the whole value of possession.
To satisfy rising market demand for low-power reminiscence, Samsung Electronics stated it’s searching for to broaden the ecosystem for LPDDR-based server reminiscence, notably by collaboration with Nvidia to optimize SOCAMM2 for Nvidia’s acceleration infrastructure.
Together with Samsung, reminiscence giants SK hynix and Micron are additionally scrambling to provide SOCAMM2 for Nvidia.
SK hynix is reportedly present process pattern validation for its 192-gigabyte SOCAMM2 product at Nvidia, whereas Micron has already delivered its 192-gigabyte SOCAMM2 to the market. Nevertheless, business officers stated Samsung has secured the biggest share within the Nvidia-bound provide to date, as a result of firm’s technical collaboration with Nvidia and its superior provide capability in contrast with its rivals.
“Because the period of AI inference accelerates, system efficiency is prone to be decided not by the GPU’s uncooked computing energy however by how reminiscence bottlenecks are addressed,” an business official stated. “As main cloud service suppliers at the moment are prioritizing energy effectivity for CPUs and reminiscences, SOCAMM2 might be seen as the place to begin of a brand new AI reminiscence cycle.”
