Unmarked dark monolithic slab on a plain concrete floor, illustration for measured memory controller plateau article

68.7 Percent at One Thread, 84.4 at Two: What a Measured Memory Controller Ceiling Says About the HBM Base Die

What happened

Samsung has orders for the entire planned output of its second fab in Taylor, Texas, and most of that volume is for AMD. The fab does not exist yet. Site preparation and ground leveling are underway. Full-scale construction is scheduled to begin in early November 2026. Equipment installation is expected to start in February 2029. Mass production is targeted for 2030. That is a fully committed order book on an empty field, four years before the first wafer.

The part being booked is not memory. Taylor Fab 2 is a 2nm logic fab, and what it produces is HBM base dies, the logic chip at the bottom of a memory stack. Samsung plans to combine Texas-made base dies with DRAM core dies produced in South Korea and ship them as finished HBM. The site is about 2.7 million square feet, roughly 1.5 times the size of Taylor Fab 1, which has a contract with Tesla. Samsung declined to comment on customer details and has confirmed no customer.

A base die is worth booking four years early because it holds the memory controller. That is the component that decides whether a published pin rate becomes a delivered bandwidth figure.

Why the base die is where the decision gets made

Standard HBM puts the memory controller on the compute die, which consumes silicon area that could otherwise hold compute. NVIDIA’s NVHBM moves the controller into the 3D stack. The published figures for that change, from NVIDIA’s own August 26, 2026 announcement, are up to 30 percent more memory bandwidth, 15 percent lower HBM power consumption, and up to 25 percent more usable area on the XPU compute die compared with standard HBM4E. Amazon’s Annapurna Labs is named as the first partner, starting with Trainium4.

The catch is which fab builds that base die. NVIDIA has asked all three memory makers, Samsung, SK hynix and Micron, to produce NVHBM base dies on TSMC’s 3nm node, even though Samsung runs its own 3nm lines. The stated reason is yield. Industry estimates put TSMC’s 3nm above 90 percent, against roughly 60 percent for Samsung Foundry’s 2nm process. Samsung is validating its 2nm line as a fallback if TSMC capacity falls short, which is a way of admitting the fallback is worse than the thing it is falling back from.

TSMC flagged tight 3nm capacity at its October 15 earnings call. That node accounted for 33 percent of third-quarter wafer sales, its largest share of any process node, and a further 3nm fab in Tainan is slated for mass production in the first half of 2028. So the practical shape of the supply chain is this: memory makers produce core dies, TSMC produces the base die and packages the accelerator. The memory vendors lose the base-die initiative, including the one that built its HBM4 reputation on making its own base dies on 4nm.

Samsung is still allocating half of its 4nm capacity to HBM4 base dies at Pyeongtaek, and it announced a new advanced foundry fab at its Giheung campus for 2nm and 4nm work aimed at base-die capacity for HBM5. The Giheung decision is framed as catching base-die output up to core-die output at Hwaseong and Pyeongtaek. The bottleneck is not DRAM. It is the logic layer underneath it.

Pin rate is not delivered bandwidth

Here is the arithmetic on the memory side, using Samsung’s own published pin rates. The third column is my own derivation from the pin rate and the pin count, not a published figure.

Generation Pin rate Per-stack bandwidth at 2,048 pins (my arithmetic) Figure Samsung publishes
HBM4, in mass production 11.7 Gbps 3.00 TB/s not published
HBM4E, shipping samples since late May 14.0 Gbps 3.58 TB/s 3.6 TB/s
HBM4E, confirmed preparation 16.0 Gbps 4.10 TB/s 4 TB/s

My numbers agree with Samsung’s published ones, which is the point. The pin rate is a channel, and the channel is not the delivered rate. What turns 4.10 TB/s of channel into something a decode loop can use is the controller, and the controller is being moved onto a different die at a different fab on a four-year schedule.

What I measured on my own hardware

The part of this that transfers to a local build is not the fab. It is the gap between a memory specification and the rate a platform actually delivers. I measured that gap on the box that serves my models.

Hardware: AMD Ryzen 9 5900X, 12 cores, 64 MB of L3, 2×32 GB G.Skill F4-4000C18-32GTZR. The memory facts came from Win32_PhysicalMemory on that machine, not from the product page. Both modules report Speed and ConfiguredClockSpeed of 2667 MT/s. The kit is sold as DDR4-4000 CL18. The platform runs it at 2667, which is 66.7 percent of the rated speed.

Channel arithmetic from the machine as configured: 2667 MT/s, 8 bytes per transfer, 2 channels, 42.7 GB/s. As labelled on the kit: 4000 MT/s, 8 bytes, 2 channels, 64.0 GB/s. The difference between those two numbers is not a benchmark result. It is what the memory controller was able to hold.

Method: memmove copy passes, 256 MB working set per thread. That is unambiguously past the 64 MB of L3, so every row below is DRAM-tier traffic, not cache. Two warm-up passes, 11 timed passes per thread, median per row. The whole sweep was run five times inside one script execution over one connection, and the range column is the spread across those five sweeps. Copy bandwidth counts bytes moved. Memory traffic counts the read and the write, so it is double. These are not the same measurement and I am not presenting them as one.

Threads Median copy GB/s Range across 5 sweeps Median per-stream GB/s Percent of configured channel as memory traffic Percent as data moved Gain vs 1 thread
1 14.65 12.36 to 16.84 14.67 68.7 34.3 —
2 18.01 16.90 to 18.03 9.02 84.4 42.2 +22.9%
4 16.81 15.94 to 16.92 4.32 78.8 39.4 +14.7%
6 16.54 16.42 to 16.64 2.78 77.5 38.8 +12.9%
8 17.16 15.12 to 17.44 2.20 80.4 40.2 +17.1%
12 17.17 16.37 to 17.17 1.46 80.5 40.2 +17.2%
16 16.06 15.90 to 16.80 1.03 75.3 37.6 +9.6%

The table says something simple. The channel reaches its ceiling at two threads. One thread to two is a 22.9 percent gain. Fourteen more threads after that do not add bandwidth; they subtract it. Sixteen threads deliver 16.06 GB/s against 18.01 GB/s at two threads, a 10.8 percent loss, and the sixteen-thread range of 15.90 to 16.80 does not overlap the two-thread range of 16.90 to 18.03. Extra cores do not add bandwidth. They add queue pressure, and the queue is managed by the controller.

That is the same decision the base die makes, one scale down. A memory controller is a queue manager. The pin rate is the channel, and the controller is what gets close to it. On my machine the controller reaches 84.4 percent of the configured channel as memory traffic at two threads and 68.7 percent at one. On an HBM stack, moving that controller onto the base die is what NVIDIA claims buys 30 percent. Both numbers describe the same physical problem. A channel is only worth what the logic layer in front of it can drain.

The per-stream column also shows why a single agent loop does not get the headline number. At eight threads the aggregate is 17.16 GB/s but each stream sees 2.20 GB/s. A decode loop that runs one request at a time operates at the one-thread row, not the eight-thread row. The ceiling on local decode is set by the single-stream column, and the aggregate column only matters when requests are overlapped.

What it means for a local build

The price ladder for 64 GB of desktop memory is now a form-factor and binning story, not a bandwidth story. Prices below were read from the live Amazon listings on November 10, 2026. The theoretical channel column is my own arithmetic: 8 bytes per transfer per 64-bit DIMM path, doubled for two DIMMs.

Kit ASIN Price (Nov 10, 2026) Price per GB Theoretical channel (my arithmetic)
G.SKILL Trident Z RGB 64GB 2×32 DDR4-3200 CL16 desktop UDIMM B087T2B287 $279.99 $4.37 51.2 GB/s
G.SKILL Trident Z5 Neo RGB 64GB 2×32 DDR5-6000 CL36 EXPO and XMP desktop UDIMM B0DN3WFS52 $1,079.99 $16.87 96.0 GB/s
Crucial Pro 64GB 2×32 DDR5-5600 CL46 desktop UDIMM B0C79H54TQ $1,178.77 $18.42 89.6 GB/s
G.SKILL Trident Z5 Neo RGB 64GB 2×32 DDR5-6000 CL30 EXPO desktop UDIMM B0BJNTLJ5X $1,279.99 $20.00 96.0 GB/s

Price scales 4.57 times from the cheapest kit to the dearest. The theoretical channel scales 1.875 times. Those two ratios disagree by a factor of 2.44, and the disagreement is what you are actually buying: latency bins, EXPO profiles, heatspreaders, and a controller that may or may not hold the rated speed on your board.

The spot data explains why the DDR4 kit sits at $279.99 while the same capacity in DDR5 sits above $1,000. TrendForce’s November 4 report has mainstream DDR4 1Gx8 3200MT/s spot at $46.82, down 0.09 percent from $46.86 on October 28, with spot prices above suppliers’ official quotes and trading volumes shrinking. Buyers are taking a wait-and-see stance. 512Gb TLC wafer spot fell 4.33 percent to $16.152. Contract prices are still climbing while spot drifts down, which is why HBM is eating your RAM at the retail level. The capacity that would have gone to a desktop kit is being allocated to stacks with a custom logic layer attached.

The practical rule from my own measurement: a kit rated above what the platform holds delivers the platform’s rate, not its own. My 4000 kit runs at 2667, so it produces the same channel a 3200 kit would produce at 2667. Paying for the higher bin only pays off when the memory controller can hold it, and with two dual-rank 32 GB modules on a consumer board that is not guaranteed. Read what the machine reports. The sticker is a claim about the modules. The configured clock is a claim about the system, and only one of them becomes bandwidth.

Two things follow for anyone sizing a build. First, capacity and bandwidth are separate purchases. A 64 GB kit at $4.37 per GB and a 64 GB kit at $20.00 per GB hold the same model. The second one buys a wider channel, and a wider channel only helps a workload that can fill it, which a single request at a time cannot. Second, the power side of the base-die argument is not decoration. NVIDIA’s 15 percent HBM power claim is a claim about the controller, not the DRAM, and it lands on the same curve as what power capping costs a local build. A controller that drains the channel at 68.7 percent is burning watts to move fewer bytes.

For the buying decision itself, the tier ladder is in the best hardware for running local LLMs guide, and I am not restating it here. The version specific to this article: if your workload is one agent loop at a time, buy the capacity you need at the cheapest bin your board can actually hold, and stop paying for a rated speed your memory controller will not honour.

A four-year order book on an empty Texas field is the same fact at industrial scale. Nobody is buying DRAM. Everybody is buying the logic layer that decides what the DRAM is worth.

3 Comments

Leave a Reply

Your email address will not be published. Required fields are marked *