NVIDIA-ийн удахгүй гарах Rubin серверүүдэд ашиглагдах CMX технологи нь NAND хадгалах төхөөрөмжийн зах зээлд асар их ачаалал үүсгэх төлөвтэй байна.
NVIDIA-ийн CMX буюу “context memory storage” технологи нь нэг серверт 576 ширхэг SSD багтааж, нийт 9,600 TB багтаамж олгох хүчин чадалтай юм. Салбарын шинжээчдийн тооцоолсноор, CMX-д шаардлагатай NAND-ийн эрэлт энэ онд 35 сая TB байсан бол ирэх жил 100 сая TB-аас давж огцом өсөх хандлагатай байна. Шинжээч Жукан энэхүү эрэлтийг Apple-ийн хэмжээний томоохон зах зээлийн эрэлттэй дүйцэхүйц хэмжээний нөлөө үзүүлнэ хэмээн онцолжээ.
Энэхүү шинэ технологи нь AI хурдасгуурын HBM (high bandwidth memory) болон ердийн хадгалах төхөөрөмжүүдийн хооронд байрлах бөгөөд ‘Spectrum-X Ethernet’-ээр дамжуулан өгөгдлийг илүү хурдан хуваалцах боломжийг олгодог. Samsung зэрэг үйлдвэрлэгчид энэхүү өсөн нэмэгдэж буй эрэлтийг хангах гол нийлүүлэгч болохоор ажиллаж байна.
Гэвч энэхүү өөрчлөлт нь энгийн хэрэглэгчдийн хувьд сөрөг нөлөөтэй байж болзошгүй юм. AI салбарын томоохон гэрээнүүд нь NAND болон DRAM-ийн нийлүүлэлтийг тэргүүн ээлжинд авах тул жирийн хэрэглэгчдэд зориулсан бүрэлдэхүүн хэсгүүдийн үнэ болон олдоцонд хүндрэл үүсэх эрсдэлтэй байна.
Дэлгэрэнгүйг эх сурвалжаас харах
↓Эх сурвалжийг нээх ↓
Any time there’s potentially positive news about storage and memory supply—such as increasing supply from China—I can’t help but think ‘sure, but the AI industry will guzzle through it regardless.’ Not only is demand seemingly insatiable, but sometimes new developments that encourage even more memory use come along. Case in point here with Nvidia Rubin.
As SE Daily reports (machine translated), Nvidia’s new ‘context memory storage’ (CMX) tech, which should be here with its new Rubin servers later this year, is going to chomp through a whole bunch of NAND.
“Nvidia’s CMX is equipped with 576 SSDs, providing a total storage capacity of 9,600 TB. The industry anticipates that the demand for NAND required for CMX will surge from 35 million TB this year to over 100 million TB next year.”
Analyst Jukan puts these figures into context on X: “Nvidia’s Rubin CMX will quite literally soak up NAND supply like a sponge absorbs water. To put 100 million TB into perspective, it is roughly equivalent to adding another Apple-sized source of demand to the NAND market.”
Samsung is apparently attempting to establish itself as a key supplier to Nvidia for this new kind of demand.
“Nvidia’s CMX is equipped with 576 SSDs, providing a total storage capacity of 9,600 TB. Industry estimates suggest that NAND demand for CMX will surge from 35 million TB this year to more than 100 million TB next year.”Nvidia’s Rubin CMX will quite literally soak up NAND… https://t.co/fNJGMhNJNv pic.twitter.com/mYCMfhr1oqJuly 20, 2026
CXM essentially exists somewhere between an AI accelerator’s high bandwidth memory (HBM) and its regular masses of storage. You can think of its relation to HBM as a bit similar to RAM’s relation to CPU cache and SSD storage. This CXM, which can be shared across different AI accelerators, is accessible much more quickly than standard long-term storage because it connects over ‘Spectrum-X Ethernet’.
It’s good news for NAND makers like Samsung, of course, as there is yet more demand for storage. But it’s probably not so great for consumers like you and I, at least not in the short and medium term. More NAND or DRAM demand from the AI industry, with their big contracts, seems to usually be prioritised over the consumer market.
There are some higher-ups in the memory industry that seem at least somewhat concerned about consumer prices, but I’m sceptical as to whether those sentiments will do anything to redirect any actual supply away from AI servers and into home consumers’ hands.
