A single unmarked dark slab on a plain surface, lit from one side

133 TB/s Is a Card Number: What Measured HBC Silicon Says About Bandwidth per Byte of Pool

What was actually measured

Qualcomm announced on November 2, 2026 that its first-generation High Bandwidth Compute silicon passed hardware validation in its San Diego lab. The test vehicle is a two-layer DRAM stack bonded over a logic die. It ran on industry-standard automated test equipment for more than a month, driven through production-representative data paths with randomized weights and activations, and was measured in both an LPDDR mode and an HBC mode on the same sample.

Two numbers came out of that run. The stack measured 18 times the effective memory bandwidth of external LPDDR5X, and more than 5 times the bandwidth per watt of published HBM4 specifications. Both are Qualcomm’s own measurements on Qualcomm’s own sample, normalized to target operating conditions and compared against published competing specifications. The company’s footnote says exactly that.

Two details matter more than the headline. The validation sample is two DRAM layers. Shipping products will be four. And the efficiency figure is a step down from the 6x Qualcomm projected at its June 2026 investor day. Qualcomm says the result ranges from 5x to 6x depending on normalization, and that nothing in testing changed the June claims. A projection that came in low when it was measured on silicon is still a win, but it is a measured number now, and that changes what it can be used for.

The 133 TB/s is a card number, not a module number

The number attached to the architecture is 133 TB/s. That is the total memory bandwidth of a Dragonfly AI250 accelerator card. It is not the bandwidth of one HBC module, and Qualcomm has not published a per-module figure.

Reading it against the current generation makes the arithmetic legible. Qualcomm rates a Dragonfly AI200 rack at 414 TB/s across 56 cards, which is 7.39 TB/s per card. 133 divided by 7.39 is 18.0. The 18x is not an independent measurement, it is the same claim expressed at card scale. The other published number closes the loop: 133 TB/s times 56 cards is 7,448 TB/s, which is the 7.4 PB/s Qualcomm advertises for an AI250 rack. Both figures are consistent because they are the same figure at two zoom levels. That is fine, as long as nobody reads 133 TB/s as what one stack can do.

The part that is genuinely interesting is what did not change. The AI250 rack carries the same 43 TB of memory as the AI200 rack, and the same 768 GB per card. The pool is identical. Only the rate changed. So the entire claim is about distance: data that travels a few micrometers through a vertical bond instead of a few centimeters across a package.

Bandwidth per byte of pool, not link rate

A link rate tells you what a tier can move. What a workload actually experiences is the rate against the pool behind it, because the pool is what has to be read every token. Here is the same comparison expressed per byte of memory, with the local row measured rather than quoted.

Tier Pool Rate GB/s per GB of pool Source
Qualcomm AI250 card with HBC 768 GB LPDDR5X 133 TB/s 173.4 Qualcomm figures, my arithmetic
Qualcomm AI200 card, no HBC 768 GB LPDDR5X 7.39 TB/s 9.63 Qualcomm rack figure divided by 56 cards
NVIDIA Rubin card, HBM4 288 GB 22 TB/s 76.4 Published spec, my arithmetic
Ryzen AI 9 HX 370 board 32 GB LPDDR5X-8000 128 GB/s 4.00 Published spec, my arithmetic
My serving box: Ryzen 9 5900X, 64 GB DDR4-2667 64 GB 17.37 GB/s measured 0.271 Measured, 7 passes, median

The HBC row is 2.3 times the HBM4 row on a per-byte basis and 18 times its own predecessor. It is also 639 times what a desktop board pulls per byte of its own memory. That ratio is not a speed ranking. It is a statement about how hard each byte is being worked, and it is the honest way to compare a stack bonded onto logic with a DIMM sitting on a motherboard.

What I measured on the serving box

The local row in that table is not a spec. I ran a copy pass and a read-only pass on the machine that serves my models, a Ryzen 9 5900X with 64 GB of DDR4-2667 in two channels, seven passes each on a 256 MB working set, median reported.

Copy pass: median 16,589 microseconds for 256 MB, which is 32.36 GB/s counted as data moved and 16.18 GB/s counted as memory traffic. A copy touches memory twice, so the two conventions differ by a factor of two, and which one you quote changes the percentage you get.

Read-only pass: median 15,452 microseconds, 17.37 GB/s. Against the theoretical dual-channel figure for DDR4-2667, which is 42.7 GB/s, that is 40.7 percent of the channel. Counted as memory traffic, the copy pass is 37.9 percent of the channel; counted as data moved, 75.8 percent. A channel spec is not an achievable single-thread rate, and the gap is the point.

What that rate buys in practice. A decode loop reads the active weights once per token. At 17.37 GB/s, a 5.03 GB Q4_K_M file gives an arithmetic ceiling of 3.45 tokens per second, and a 12 GB file gives 1.45. Those are derived from the measured scan rate, not measured decode. A real loop beats them when part of the weights stay in cache, which is why a cache-tier number cannot be expressed as a share of the memory channel at all. The same distinction between what a channel is rated for and what one thread can pull is the subject of a measured cache cliff on this board, and the local row above is the DRAM side of that same picture.

Why the distance is the mechanism

HBM stacks DRAM dies over a base die, and that base die carries almost no compute. It holds memory controllers and repair logic. The reason is physical: the through-silicon vias feeding a DRAM stack consume base-die area and act as thermal pillars, so there is no room for real logic. A Futurum briefing estimated that HBM plus its I/O can consume up to 50 percent of an accelerator’s power budget. That estimate is the cost HBC is trying to remove by putting an accelerator die under the stack instead of beside it.

Qualcomm claims a connection-level innovation that frees that real estate for serious logic. The energy figure follows from the geometry: moving a bit a few micrometers through a vertical bond costs roughly an order of magnitude less than moving it across a package. The 5x bandwidth per watt is what that looks like when it is measured on a working data path.

Thermal behavior is the open half. A base die that computes produces more heat than a base die that only repairs, and it sits under four layers of DRAM. Qualcomm says its validation included thermal behavior and gave no detail. That is the part of this program that has not been answered in public.

The competitive shape is worth naming. Memory vendors responded to the concept by asking to work closer with Qualcomm, and several have announced stacked-DRAM variants of their own, including Samsung’s zHBM. Qualcomm’s own posture is that stacked DRAM and customized HBM will co-evolve. The architecture is not a bet against HBM. It is a bet on where the logic die sits.

What this means for a local build

Nothing ships, and nothing will for a year. AI250 sampling is scheduled for the first half of 2027, with second-generation HBC in the AI300 in 2028, alongside the C1000 data center CPU. Anyone buying hardware today is buying against published bus widths, not against this architecture.

The relevant transfer is that LPDDR stopped being a laptop memory and became a data center memory tier. Qualcomm’s current card runs 768 GB of LPDDR5X per card, and the first server CPU that talks LPDDR is the other half of that story, covered in what AMD EPYC Verano says about memory power. The decode-side case for buying bandwidth instead of capacity is the same argument made about Maia 300 earlier this year, and the buying ladder for a home build is in the hardware guide rather than restated here.

On the machine a reader can actually buy, the per-byte rate is what decides whether a model runs at reading speed or at a crawl. As of November 20, 2026, the tiers on Amazon look like this:

Prices are approximate and read from the listings on the date above. The DDR4 kit is a 288-pin UDIMM desktop kit, not a SODIMM, and the three mini PCs have soldered LPDDR that cannot be upgraded after purchase.

What this article is not

No independent reproduction exists. The 18x and the 5x are Qualcomm measurements on a two-layer test sample, normalized to target conditions, against published competitor specifications. No per-module bandwidth figure has been disclosed. No product has shipped. The thermal answer is not published.

What is real is that a stacked DRAM-on-logic architecture moved from an investor day slide to measured silicon in one quarter, and that the measurement came in below the projection. The next number that matters is a four-layer stack measured by someone else.

Comments

No comments yet. Why don’t you start the discussion?

Leave a Reply

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