SRAM
GSI Technology (GSIT): sells standalone SRAM and also develops APU products. Standalone SRAM is not the embedded cache inside every AI chip.
Supplier product source ↗SRAM vs DRAM vs HBM: AI Memory Explained
Memory architecture, supplier exposure, and investment research considerations.
Key distinctions
- SRAM is fast working memory, often used for small caches inside processors.
- DRAM holds working data. HBM is stacked DRAM with a wide interface for high bandwidth—not a separate memory family.
- NAND flash retains data without power. SSDs typically use NAND for storage; they do not replace working memory.
How memory fits around an AI accelerator
Accelerator package
GPU chip
Compute cores + SRAM cache
HBM · DRAM
Working data near the GPU
Server memory · DRAM
For example, DDR5 modules used by the CPU
Storage · NAND / SSD
Keeps files and datasets when powered off
Compare the trade-offs
Bandwidth is data moved per second; latency is the wait for an access; capacity is how much fits. A high-bandwidth product is not automatically the lowest-latency option.
| Technology | Role | Strength | Trade-off |
|---|---|---|---|
| SRAM | Fast working memory often used in small caches close to a processor. | Low access latency; no periodic refresh. | Larger cells mean less capacity per area and higher cost per bit. |
| DRAM | Working memory that temporarily holds data while a computer runs. | More capacity per chip area than SRAM. | Needs refresh; slower access than SRAM. |
| HBMA DRAM subtype | Stacked DRAM built to move large amounts of data to an accelerator. | High bandwidth in a compact package near the accelerator. | Stacking, packaging and cooling add engineering complexity. |
| NAND | Storage chips that keep data when power is off. | Retains data without power; suited to large storage capacity. | Slower than working memory; write endurance is finite. |
SRAM needs no periodic refresh, but it still loses data without power. DRAM—including HBM—also needs power and refresh. Only NAND in this comparison is non-volatile.
Who supplies what?
Selected US-listed examples, not a complete supplier list or a ranking. A technology’s use in AI does not establish how much of a company’s revenue comes from AI.
DRAM
Micron (MU): DRAM is a core product business.
Supplier product source ↗HBM
Micron (MU): sells HBM; not a pure-play HBM company.
Supplier product source ↗NAND
Sandisk (SNDK): flash storage is its core business. Micron (MU) also supplies NAND and SSDs.
Supplier product source ↗NVIDIA designs accelerators that incorporate caches and use external memory. That does not make NVIDIA a merchant supplier of standalone SRAM; nor does more GPU cache automatically create orders for GSI Technology.
Micron business & financials →Sandisk business & financials →
Investment research framework
These are editorial research prompts, not return forecasts. Read the latest dated company disclosures before drawing a company-level conclusion.
- HBM expansion announced? Check customer qualification, usable output and shipment timing. Then compare margins and capital spending; announced capacity is not recognized revenue.
- DRAM or NAND prices rising? Separate price increases from shipment growth. Compare inventories and new supply, and avoid treating one strong quarter as permanent earnings.
- An AI chip uses more SRAM? Identify whether it is embedded cache or a separately purchased component. Confirm a supplier relationship before assuming a listed company benefits.
- Business growth and valuation Compare cash generation, debt, dilution and the valuation already paid. Better technology alone does not establish an attractive stock price.
Technical sources
- IBM · DIMM / SRAM / DRAM — Memory families and server modules.
- Micron · HBM — Stacked DRAM, wide interface and package integration.
- IBM · NAND flash — Persistent storage and write endurance.
- Sandisk · RAM and flash — Working memory versus persistent storage.
- NVIDIA · Hopper architecture — A concrete example of GPU caches and HBM; not a universal GPU layout.