Lenovo warns era of cheap memory is over as RAMageddon forces enterprise server shifts and NVIDIA hardware adjustments
A report by ComputerBase on a recent industry presentation by Lenovo at the International Supercomputing Conference reveals that the era of cheap memory and storage is over. According to Lenovo executives, despite the rise in global manufacturing capacity, prices in DRAM and NAND solid state storage look set to stay astronomically high long term. The manufacturing industry is entering a new chapter where pricing cycles are broken.
Although the "prices will never go down again"
statement was delivered tongue in cheek on stage, the bullish fundamental outlook for at least the next five years, if not decade or more, suggests a significantly higher pricing floor will be maintained as the new standard. This is borne out by a number of leading memory producers, such as SK Hynix, which has just announced plans to literally triple its total production capacity over the next 10 years. Industry speculation indicates that no semiconductor manufacturer would invest heavily in such large plug figures, unless they were confident that returns will far outweigh potential stock write offs over the "next while"
.
Getting through this extended high cost era called RAMageddon by Lenovo will require enterprise buyers to rethink their system memory configuration entirely. In the old days, server players always hawked their boxes by how much memory they could hold. Now, fully populating all of those memory banks is too expensive for many companies. Evolving workloads to GPU accelerated architectures is a much more attractive solution.
As this market evolution occurs the arrival of server platform, with support for 16 channel memory begins. Along the same lines, a dual socket system of this nature can potentially be fully populated with up to 32 memory modules. The presence of such high memory amounts, is to yield, to the large performance presentation of these solution, thus, also raising the minimum environment configuration to a minimum of 1 TB RAM.
Leading hardware such as NVIDIA's GB200 NVL4 platform have introduced hybrid architectures to tackle these memory limits. This blends a memory resource of close to 1 TB to be shared amongst a bank of LPDDR5 system memory around the 2 central processors and 186 GB of HBM3e high bandwidth memory across the 4 graphics processors averaging around 120 GB per CPU, which is in regards to servers quite sparse.
In response to these staggering costs, reports have begun to emerge that NVIDIA is potentially repackaging the memory capacity on the upcoming Vera processors. The hardware is expected to ship with a scaled back memory configuration option of 768 GB instead of the planned for 1.5 TB. On scale deployments such as an NVL72 rack, this would reduce total memory requirements from 55 TB down to 28 TB, providing a significant cost saving for enterprise cloud providers.
