TDEB Top Bar

Trending
Semiconductor Industry News — The Digital Electronics Blog
Follow Us

TDEB Masthead

The Semiconductor Hub

Semiconductors.  Chips and Systems.  Innovation.

Indexed by leading Search Engines, AI and referenced by Engineers worldwide 500+ Articles 10M+ Pageviews 30+ Reports 50+ Citations

After the SRAM Bitcell Stalled: Sizing Caches and Choosing a Response

Murugavel Ganesan
by
0
Share Facebook X LinkedIn WhatsApp Email
A stacked silicon die sitting on top of a compute die, with a regular grid of memory cells visible at the edge.

TSMC's published high-density SRAM cell shrank from 0.027 µm² at 7nm to 0.021 µm² at 5nm. Then progress all but stopped. More than five years later, reports of its 2nm cell range from 0.021 µm² to about 0.0175 µm², which is at best a modest step. Over part of that stretch, TSMC claimed up to 1.6 times logic density for 3nm against 5nm. Logic kept moving and the cache largely did not.

For a chip architect this turns into a budget problem. Cache takes a growing share of a die that is getting more expensive per square millimeter. This post sets out the bitcell data, explains why the cell has stalled, sizes a 64 MB cache step by step, and weighs the four main responses: stacking, eDRAM, gain cells and MRAM.

Key takeaways

  • The high-density cell has barely moved since TSMC's 5nm. Reports of its 2nm cell disagree, at 0.021 µm² or about 0.0175 µm², and Intel's 18A cell is reported at 0.021 µm².
  • Macro density can still improve without a smaller cell. TSMC's ISSCC 2025 paper title gives 38.1 Mb/mm², and that is the figure to plan with whichever cell size proves correct.
  • Bitcell area badly understates cache area. In our illustrative 64 MB example, 11.27 mm² of raw cells becomes 20.85 mm² of cache.
  • Of the four, stacking is the one we could confirm in current volume products. eDRAM has a product history, while gain cells and SOT-MRAM are still research results.

The idea in plain words

Almost every processor keeps its most-used data in SRAM cache, built from millions of tiny six-transistor memory cells. For decades each new manufacturing node shrank both the logic and those cells, so a chip could get more cache for the same area. That stopped working evenly around 5nm: logic kept shrinking, the SRAM cell barely did.

The result is a budget squeeze. The same cache takes a bigger share of every new chip, and each square millimeter of that chip costs more than before. An architect has a few ways out:

  • Stack the cache on a separate die, made on an older and cheaper process, and bond it to the processor.
  • Use eDRAM, a denser memory with a small capacitor per bit, at the cost of constant refresh.
  • Use gain cells, two-transistor cells that are smaller than SRAM but hold data only for a limited time.
  • Use MRAM, a magnetic memory that keeps data with the power off and leaks very little.

Only stacking is proven in volume products today. The deep dive gives the cell sizes, sizes a 64 MB cache step by step, and compares the four options with their published numbers.

Deep dive: the numbers and details

This part holds the bitcell data, the physics of the stall, a worked 64 MB cache sizing with costs, and the four responses in detail. Skip it if the overview is all you need.

The bitcell data

These are conference figures for the smallest cell on offer. A product may use a larger, faster cell.

ProcessCell areaWhere the figure comes from
TSMC 7nm0.027 µm², high densityWu et al., title of the IEDM 2016 paper
TSMC 5nm0.021 µm², high densityYeap et al., title of the IEDM 2019 paper
TSMC 5nm to 3nm rangeAbout 0.021 µm²; a first 3nm variant is listed at 0.0199 µm² and N3E at 0.021 µm²Stanford's GainSight paper, 2025, which describes this range as a plateau, and Tom's Hardware's comparison table
TSMC N20.021 µm² or about 0.0175 µm²IEEE Spectrum and Mark LaPedus report the first from ISSCC 2025. Tom's Hardware reports the second. We could not read the paper to settle it
Intel 18A0.021 µm² high density, 0.023 µm² high currentIntel's ISSCC 2025 paper, as quoted by Mark LaPedus

The 7nm to 5nm step cut the cell by 22%. Nothing after it comes close to that pace, even on the more generous reading of 2nm. To see why that hurts, compare it with TSMC's December 2022 claim of up to 1.6 times logic density for 3nm against 5nm. Take a die that is half logic and half SRAM by area, an illustrative split, and move it across that step at the best-case logic figure with no SRAM shrink:

  • Logic: 0.50 divided by 1.6 gives 0.3125 of the old die.
  • SRAM: stays at 0.50.
  • New die: 0.8125 of the old area, an 18.75% shrink, where logic alone shrank 37.5%.
  • SRAM share of the die: 0.50 divided by 0.8125, or about 62%, up from 50%.

So each new node gives the same cache a larger share of a costlier die. TSMC says N2 entered volume production in the fourth quarter of 2025, so this is now a live budgeting question and no longer a forecast.

Why the cell has nearly stopped shrinking

A 6T cell stores a bit in two cross-coupled inverters and depends on those transistors being well matched. A 2025 Stanford paper, "The Future of Memory: Limits and Opportunities", puts the limit plainly: as transistors approach atomic scales, manufacturing tolerances limit matching, and that reduces signal margin. Logic escapes this because each gate restores the signal for the next stage. A bitcell has no such help. It has to hold state, survive a read and accept a write using very small devices.

The same paper goes further, stating that SRAM's cost per byte has flattened with no roadmap to a significant reduction within five years. That is an academic view, not a foundry commitment, but it is a sensible planning assumption until a shrink appears in a product you can buy.

What has kept moving is everything around the cell. IEEE Spectrum reported that TSMC's N2 design doubles the bit line from 256 to 512 cells, which improves density by close to 10%. Longer bit lines mean fewer sense amplifiers, so less periphery per bit. So density in Mb/mm² can rise even when the cell barely moves.

Standby power adds to the pressure. imec names both slowing bit density and standby power as reasons to look past SRAM for last-level cache. Density and leakage also shape other uses of the cell, as our post on SRAM-based in-memory compute discusses.

A worked example: how big is a 64 MB cache?

The cell areas and macro densities below are sourced. Everything else is an illustrative assumption, not data from any foundry or product. Replace those inputs with your own.

InputValueStatus
Data capacity64 MB, or 536,870,912 bitsIllustrative choice
High-density cell, 2nm class0.021 µm²Sourced, see table above
Macro density, 2nm class38.1 Mb/mm²Sourced, TSMC ISSCC 2025 paper title
Macro density, 7nm29.2 Mb/mm²Sourced, Yokoyama et al., VLSI Circuits 2020 paper title
ECC12.5% extra bits (8 per 64)Illustrative
Line size and tag64-byte lines, 30 tag and state bits per lineIllustrative
Control, routing and floorplan factor1.25 times array areaIllustrative

Step 1, raw cells. 536,870,912 bits at 0.021 µm² is 11.27 mm².

Step 2, the macro overhead. A cell of 0.021 µm² implies 47.6 Mb/mm² if nothing else existed, treating Mb as a million bits. The titled macro reaches 38.1 Mb/mm², which is 80% of that, and only 67% if the cell is really 0.0175 µm². Either way a large share goes on periphery and layout overhead, even in a block built to show density. The steps below use the macro figure, so the cell size dispute does not change the result.

Step 3, the extra bits. ECC adds 67,108,864 bits. There are 1,048,576 lines, so tags add 31,457,280 bits. The total is 635,437,056 bits, or 18.4% over the data alone.

Step 4, array area. 635.44 Mb divided by 38.1 Mb/mm² is 16.68 mm².

Step 5, the finished cache. 16.68 mm² times 1.25 is 20.85 mm². That is 1.85 times the raw cell figure.

Now use the 7nm macro density, standing in for an older node you might stack. Steps 4 and 5 give 21.76 mm² and then 27.20 mm², about 30% more area. Here is a rough cost comparison with illustrative wafer prices of $30,000 for the leading node and $10,000 for the older one. A 300 mm wafer has 70,686 mm², ignoring edge loss and yield.

On the leading-node dieOn a stacked older-node die
Cache area20.85 mm²27.20 mm²
Illustrative cost per mm²$0.424$0.141
Illustrative silicon cost$8.85$3.85

The gap is $5.00, so in this model stacking wins if bonding, test, yield loss and thermal work together cost less than that per cache. The answer depends entirely on the wafer prices, so keep the method and replace the dollars.

A real part gives a useful check. AMD's slides list its first 3D V-Cache die at 41 mm² on TSMC 7nm for 64 MB, about 1.5 times our 27.20 mm². The material we read does not break that area down, and the die carries through-silicon vias, so this is no more than a rough check. It does suggest a first-pass estimate is more likely to be too small than too large.

Four architecture responses

OptionDensity evidenceMain catchStatus
Stacked SRAMSame cell, moved to a cheaper or separate dieBonding cost, heat, a costlier hop than on-die wiresIn volume products
eDRAM (1T1C)0.029 µm² at Intel 22nm, 17.5 Mb/mm²Refresh, plus a capacitor built into the metal stackPast products
Gain cell35% to 69% of 6T SRAM cell area in one Stanford studyRefresh, lower speed, single-ended readResearch
SOT-MRAMNo numeric cell area in the imec material we readExtra access transistors, array-level maturityResearch

Cache stacking

This response leaves the cell alone and changes where it sits. AMD's ISSCC 2022 paper by Wuu et al. describes a hybrid-bonded 64 MB stacked cache for a 7nm CPU, and AMD's slides list a 9 µm bond pitch on an 81 mm² compute die that carries 32 MB of its own L3. With its 2024 Ryzen 7 9800X3D, AMD moved the 64 MB cache below the processor, bringing the compute die nearer the cooler.

Intel's route is similar in intent. Its Intel 18A data center white paper describes compute chiplets on an active base tile that can hold memory IP for caching, built on a previous-generation node to lower R&D cost. It gives 9 µm for the first generation of Foveros Direct 3D and 3 µm for the second. Our posts on pitch and yield in hybrid bonding and on how chiplets assemble cover the mechanics.

The hop is not free. A table in the Stanford paper lists about 600 fJ per bit for a hybrid-bonded connection against about 5 fJ per bit on-die. Those are representative figures, not a product measurement, but a ratio of 120 is a good reason to keep the busiest data on the same die as its core.

eDRAM

Intel's 22nm eDRAM process, reported at the 2013 VLSI Technology Symposium, had a 0.029 µm² cell, 17.5 Mb/mm² on a 128 Mb macro, and retention above 100 µs at 95 °C. That 22nm cell is only 38% larger than a 0.021 µm² SRAM cell, yet its macro density is less than half of 38.1 Mb/mm².

Singh et al. list eDRAM's use in IBM's Power series, Intel's Haswell and the Xbox 360, mainly as last-level cache. The catches are a trench capacitor built into the metal stack and constant refresh. Check with your foundry before planning around it.

Gain cells

A gain cell stores charge on a transistor's own gate, using two transistors and no separate capacitor. Stanford's OpenGCRAM paper (February 2026) reports 2T silicon cells at 69% of 6T SRAM cell area, and 2T cells with an oxide-semiconductor write device at 35%. The paper does not state a node for those layouts. It also says these macros run slower, partly because of the single-ended read path.

Retention sets the architecture. The same paper gives microseconds for silicon cells and milliseconds for the oxide hybrid. imec reports far longer retention for capacitor-less IGZO cells: above 1,000 s with write times under 10 ns, and above 100 s at a 14 nm gate length. imec names reliability as the main concern, particularly threshold shifts under positive bias stress, and calls the cell an attractive eDRAM candidate.

The strongest argument is about data lifetime. Stanford's GainSight study, modeled at TSMC N5 with silicon gain cell figures scaled from 16nm arrays, found the following for a GPU:

Cache levelAccesses under 1 µs (silicon gain cell)Accesses under 10 µs (hybrid gain cell)
L164.3%97.9%
L218.4%52.0%

For a systolic array buffer, it reports silicon gain cells at 41.97% of SRAM area and hybrid cells at 22.63%, and it finds mixed SRAM and gain cell arrays can cut active energy by up to 66.8%. Those are simulated results for specific workloads. Earlier work adds a caution: Singh et al. simulated a 2.5 times cell advantage in 28nm FD-SOI that became about 2 times at cache level, and found refresh energy eroding the benefit at 7nm.

MRAM

MRAM has two different stories, and they should not be mixed. STT-MRAM is established as an embedded non-volatile memory. Intel's ISSCC 2019 7 Mb array on 22FFL was reported with 10-year retention at 200 °C and endurance above 10⁶ cycles. That suits flash replacement, but a cache writes far more often.

SOT-MRAM is the cache candidate. imec reports below 100 fJ per bit, endurance beyond 10¹⁵ cycles and switching down to 300 ps, all on single devices, plus a functional array on a 300 mm wafer at VLSI 2024. Its case rests on low standby power and fast writes more than on proven density. A three-terminal cell needs two access transistors per element, and imec's work on sharing one track across several pillars is aimed at this penalty. Treat its area benefit as a goal until array data with a stated cell size appears.

How to choose

This is our reading of the published work. Answer in order:

  1. What share of your die is SRAM at the next node? Rerun the half-and-half sum above with your own numbers. If the share barely moves, stop here.
  2. Is the cache large and last-level, and can the package take it? If both are true, cost a stacked die using the method above. Include the thermal path.
  3. How long does your data really live? Profile real workloads. If lifetimes are short, evaluate a gain cell tier, with refresh modeled honestly.
  4. Is standby power your real problem? Then track SOT-MRAM, as a roadmap item until your foundry shows array-level data.

Our SoC lifecycle on The Semiconductor Hub shows where these decisions fall in a program.

Sources

The worked example and the half-and-half die split use illustrative numbers apart from the sourced cell sizes and macro densities. Vendor and imec figures are the organizations' own claims, as published on the dates shown, and the macro densities come from different designs by different companies.

Murugavel Ganesan
About the Author
Murugavel Ganesan
Product Director | SerDes · IVN · Edge AI · Compute | Ex Intel · Infineon

A semiconductor product leader working across product architecture, definition, strategy, and silicon and software execution, from concept through volume production. He has worked with teams across Asia, Europe and the USA, including at Intel, Infineon and ArrayComm.

His work covers ASICs, SoCs, FPGAs and chiplets for wireless, wireline, automotive, IoT and AI, including PCIe, CXL and DDR interfaces alongside RF and mixed-signal design.

Share Facebook X LinkedIn WhatsApp Email

Post a Comment

0 Comments

Your comments will be moderated before it appears here.

Post a Comment (0)

#buttons=(Ok, Go it!) #days=(20)

Our website uses cookies to enhance your experience. Learn more
Ok, Go it!