① Against the backdrop of memory shortages, Google is dismantling decommissioned servers to recover DDR4 components and developing hardware and software solutions to overcome memory bottlenecks; ② Goldman Sachs expects memory prices to rise further in the third quarter, with additional data showing that the spot price for 8GB DDR4 modules rose to $142 in August; ③ Although Google has launched custom chips optimized for memory efficiency, its move to recycle DDR4 indicates that memory shortages remain a critical bottleneck for the industry.
Cailianshe, September 3 (Editor: Ma Lan) – The persistent shortage of memory is forcing major data center operators to adopt unconventional measures.
According to$Alphabet-A (GOOGL.US)$ Nikhil Cherian, Senior Director of Supply Chain Infrastructure, revealed that Google is developing both software and hardware solutions to overcome memory bottlenecks, and is even dismantling decommissioned servers to recover reusable components, thereby establishing an internal recycling supply chain.
Google is reported to have designed special hardware adapters that allow previous-generation memory, such as DDR4, to be connected to its new generation of AI servers. Cherian also acknowledged that Google is importing decommissioned servers to strip their DDR4 modules for recycling.
He further stated that the AI industry has rapidly shifted from being compute-constrained to memory-constrained, with high-performance memory now accounting for approximately 75% of the bill of materials cost for a given AI server.
Signals of Memory Shortages
Last week, Goldman Sachs noted in a report that memory prices would continue to rise in the third quarter, forecasting an 18% to 23% increase in PC DRAM prices and a 13% to 18% increase in server DRAM prices.
Data from Trendforce shows that in the August spot market, prices for 8GB DDR4 and 8GB DDR5 modules rose to $142 and $133, respectively. The discount for DDR5 relative to DDR4 remained at 6%, unchanged from July.
Rising memory prices have compelled the AI industry to take measures to reduce the memory capacity required per unit of computing power. Google’s two TPU ASICs launched this year feature memory-optimized designs, which the company claims could reduce memory requirements to one-sixth of previous levels.
Google’s TPU8i chip adopts a dedicated hierarchical memory design, relying on high-speed DDR5 memory architecture to execute host-level tasks. Additionally, each TPU8i chip is equipped with 288GB of HBM3e high-bandwidth memory.
However, Google's move to reclaim DDR4 memory suggests that DDR5 and HBM3e are insufficient to meet the company's memory requirements for its AI chips, forcing it to rely on previous-generation memory solutions. Memory remains a key bottleneck constraining the growth of the company's computing power.
Editor/KOKO