
HBM4 is the memory that the next wave of AI accelerators is built around. It doubles the interface of every stack to 2,048 bits, moves the bottom die of the stack onto a logic process, and pushes stacks to 16 dies inside a package that has to stay under a fixed height. All three memory makers announced HBM4 volume production or shipments during the first half of 2026, and NVIDIA's Rubin and AMD's MI400 series are the first large GPUs designed around it.
This special report explains HBM4 itself: what the JEDEC standard says, how shipping parts differ from it, how the stacks are built, where SK hynix, Samsung and Micron stand as of November 13, 2026, which AI chips use it, and what it changes for engineers in 2027. Every number comes from JEDEC, a company or named trade press, with its date. Forecasts belong to whoever made them.
Key takeaways
- The JEDEC HBM4 standard, JESD270-4 (April 16, 2025), doubles the interface to 2,048 bits and the channel count to 32, specifies up to 8 Gb/s per pin for up to 2 TB/s per stack, and allows 4-high to 16-high stacks of up to 64 GB.
- Shipping parts run well above the standard's headline rate. Samsung quotes 11.7 Gb/s for its HBM4 and Micron over 11 Gb/s, which is roughly 2.8 to 3.3 TB/s per stack on the companies' own figures.
- The base die is now a logic die. SK hynix uses TSMC's logic process, Samsung its own 4 nm foundry process, and Micron describes an advanced CMOS base logic die. That opens the way to custom base dies with HBM4E.
- Stacking has not moved to hybrid bonding. JEDEC relaxed the package height to 775 micrometers, and SK hynix and Samsung kept their existing bonding for HBM4, according to 2026 reports.
- HBM4E samples are out from Samsung (14 Gb/s) and SK hynix (up to 16 Gb/s), both 12-high, and Micron expects to ramp HBM4E volume in calendar 2027. We found no published JEDEC HBM4E standard.
HBM in one paragraph
High Bandwidth Memory is DRAM built as a vertical stack. Several thin DRAM core dies sit on top of a base die, and through-silicon vias (TSVs) carry signals and power straight down through the stack. The stack sits beside the processor on a silicon interposer or similar substrate, joined by thousands of short wires. The idea is a very wide interface at a moderate per-pin rate, rather than a narrow, fast one such as GDDR, paid for with a more expensive package. For a refresher on the interposer itself, see our explainer on silicon interposers.
HBM4 in plain words
An AI accelerator spends much of its time waiting for data. The model weights and the working state of every request live in memory, and the chip can only compute as fast as that memory can feed it. HBM is the answer the industry settled on: put the DRAM right next to the processor and connect the two with a very wide bus, so a lot of data moves at once without each wire having to run very fast.
HBM4 keeps that idea and changes three things.
- The bus gets twice as wide. Each stack now has twice as many connections to the processor as HBM3 had. More wires at a similar speed means roughly twice the data per second from each stack.
- The bottom die gets smarter. The die at the base of the stack used to be a simple relay. In HBM4 it is built on a logic process, which makes the fast interface easier to drive and leaves room, in later parts, for functions such as the memory controller.
- Stacks get taller. Up to 16 memory dies can sit in one stack, but the stack still has to fit inside a package of fixed height, so each die is ground thinner.
Why it matters: the size of the model a chip can hold, and how fast it can produce each answer, depend directly on how much HBM sits beside it and how fast that HBM runs. That is why the largest GPUs of 2026 and 2027 are designed around HBM4, and why memory supply may decide which configurations ship.
The costs are real too. A wider bus means more wires through the silicon interposer that carries the stack, more power and heat in a tall pile of thin dies, and a more expensive package. A cheaper variant, SPHBM4, trades some of that width for speed so the stack can sit on an ordinary organic substrate. And where capacity per dollar matters more than bandwidth, some designs still choose GDDR memory instead.
Who makes it and who uses it
The three memory makers in this market are SK hynix, Samsung and Micron. All three reached volume production or shipments of HBM4 in the first half of 2026, and each has taken a different route for the base die, from its own process to a foundry's logic process. The first large GPUs designed around HBM4 are NVIDIA's Rubin and AMD's Instinct MI400 series. The next step, HBM4E, is already sampling, and custom base dies built for one customer's chip are expected to follow.
If you only need the outline, that is it: wider, smarter, taller, and harder to package. The rest of this report gives the standard's numbers, the vendors' parts, how the stacks are built, who supplies them and what each change means for the engineers who design around them.
Deep dive: the numbers and details
This part holds the figures, equations and worked examples. Skip it if the overview is all you need.
What HBM4 changes versus HBM3E
HBM3E was never a separate JEDEC standard. It is the industry's name for faster, larger parts built on the HBM3 interface (JESD238, published January 27, 2022). So the fair comparison is between the HBM3 standard, HBM3E as shipped, and the HBM4 standard alongside the HBM4 and HBM4E parts now shipping or sampling.
| Generation | Interface per stack | Channels (pseudo-channels) | Per-pin rate | Bandwidth per stack | Stack height and capacity | Source |
|---|---|---|---|---|---|---|
| HBM2E | 1,024 bits | 8 of 128 bits (16 of 64 bits) | Up to 3.6 Gb/s peak | About 461 GB/s at 3.6 Gb/s (our arithmetic) | Up to 12-high, up to 24 GB | Synopsys technical bulletin, April 16, 2022 |
| HBM3 (JEDEC JESD238) | 1,024 bits | 16 of 64 bits (32 of 32 bits) | Up to 6.4 Gb/s | 819 GB/s | 4-, 8- and 12-high, 16-high provided for; 4 GB to 64 GB | JEDEC release, January 27, 2022 |
| HBM3E (vendor parts) | 1,024 bits | As HBM3 | 9.2 Gb/s (Micron comparison point); 9.6 Gb/s maximum (cited by Samsung) | Up to 1,280 GB/s (Samsung 12-high) | 8-high 24 GB, 12-high 36 GB | Samsung, February 27, 2024 and February 12, 2026; Micron, March 16, 2026 |
| HBM4 (JEDEC JESD270-4) | 2,048 bits | 32 (64) | Up to 8 Gb/s | Up to 2 TB/s | 4-, 8-, 12- and 16-high; 24 Gb or 32 Gb dies; up to 64 GB | JEDEC release, April 16, 2025 |
| HBM4 (vendor parts) | 2,048 bits | As HBM4 | Over 10 Gb/s (SK hynix); over 11 Gb/s (Micron); 11.7 Gb/s, up to 13 Gb/s (Samsung) | Over 2.8 TB/s (Micron); up to 3.3 TB/s (Samsung) | 12-high 36 GB in volume; 16-high 48 GB sampling or shown | SK hynix, September 12, 2025; Samsung, February 12, 2026; Micron, March 16, 2026 |
| HBM4E (vendor samples) | 2,048 bits | Not published | 14 Gb/s, scalable to 16 Gb/s (Samsung); up to 16 Gb/s (SK hynix) | Up to 3.6 TB/s (Samsung) | 12-high 48 GB sampling; Samsung plans 8-high 32 GB and 16-high 64 GB | Samsung, June 1, 2026 (UK newsroom); SK hynix, June 18, 2026 |
Vendor figures are the companies' own, measured under their own conditions, so they are not directly comparable. The HBM2E row comes from a Synopsys bulletin, not from the JEDEC release, which we could not open.
Interface width and channels
The headline change is width. JEDEC's HBM4 release gives 32 independent channels per stack, up from 16 in HBM3, with two pseudo-channels per channel. Dividing 2,048 bits by 32 channels gives 64 bits per channel, the same channel width as HBM3, so HBM4 is best read as twice as many HBM3-sized channels rather than a new channel design. JEDEC also says the interface is backward compatible with HBM3 controllers, so one controller can support both.
Doubling width lets bandwidth double without a large jump in per-pin speed, as TrendForce noted on May 22, 2025. The price is more bumps, more TSVs, a larger die and twice as many wires through the interposer.
Per-pin rates: the standard and the parts
The speed target moved quickly. In September 2024, before the standard was final, Rambus described the JEDEC HBM4 rate as 6.4 Gb/s. The published standard set the top rate at 8 Gb/s. Then the parts went past it: SK hynix said on September 12, 2025 that its HBM4 runs at over 10 Gb/s, Micron said on March 16, 2026 that it achieves over 11 Gb/s, and Samsung said on February 12, 2026 that its HBM4 runs at a consistent 11.7 Gb/s with a maximum of 13 Gb/s.
The arithmetic is simple. Bandwidth per stack in gigabytes per second is the per-pin rate times 2,048, divided by 8. At 8 Gb/s that is 2,048 GB/s. At 11 Gb/s it is about 2.8 TB/s, and at 13 Gb/s about 3.3 TB/s, which matches the vendors' stated figures.
Stack heights and capacity
JEDEC allows 4, 8, 12 and 16 DRAM dies per stack, with 24 Gb or 32 Gb dies. So twelve 24 Gb dies make 36 GB, sixteen 24 Gb dies or twelve 32 Gb dies make 48 GB, and sixteen 32 Gb dies make the 64 GB maximum.
Micron's volume HBM4 is 36 GB 12-high, and it is sampling 48 GB 16-high stacks to customers. Samsung lists 24 GB to 36 GB in 12 layers, with 48 GB 16-high planned to customer timelines. SK hynix displayed a 16-layer 48 GB HBM4 at COMPUTEX 2026, according to its newsroom on June 2, 2026. The first HBM4E samples from SK hynix and Samsung reach 48 GB in only 12 layers, which on JEDEC's die densities points to 32 Gb dies, although neither company states the die density in the releases we read.
Power
JEDEC gives no power figure for HBM4. It does give voltage options: VDDQ of 0.7, 0.75, 0.8 or 0.9 V and VDDC of 1.0 or 1.05 V, both vendor-specific. For comparison, JEDEC's HBM3 release gave 1.1 V operation and 0.4 V low-swing signaling on the host interface.
The vendors give relative gains only, each against its own baseline. SK hynix claims power efficiency improved by more than 40% over the previous generation. Samsung also claims 40% over HBM3E, which it attributes to low-voltage TSV technology and power distribution network optimization. Micron claims more than 20% better power efficiency than its HBM3E 12-high at 9.2 Gb/s, using its internal power calculator. None of these is an absolute figure a designer can budget with, and better energy per bit does not mean lower stack power: a stack moving twice the data can still draw more in total.
Reliability features
HBM3 already brought most of the reliability machinery. The Synopsys bulletin on HBM3 lists on-die ECC with 304-bit codewords, real-time error severity reporting, error check and scrub, and refresh management. HBM4 adds directed refresh management (DRFM), which JEDEC says improves row-hammer mitigation and reliability, availability and serviceability.
The base die: from pass-through to logic
Every HBM stack has a base die at the bottom. Through HBM3E, it was made on the memory maker's own DRAM-oriented process. TrendForce described the HBM3E base die in May 2025 as memory-only, acting as a simple signal pass-through, and the HBM4 base die as logic-based, allowing tighter integration with the host chip, lower latency and better data path efficiency.
The JEDEC release does not prescribe a process for the base die. The three makers have chosen differently:
- SK hynix said on April 19, 2024, when it signed a memorandum of understanding with TSMC, that it would adopt TSMC's advanced logic process for the HBM4 base die after using its own technology up to HBM3E, so that more functions fit in limited space, as reported by TrendForce on April 19, 2024.
- Samsung builds its HBM4 logic base die on its own 4 nm foundry process, with the core DRAM on its sixth-generation 10 nm-class (1c) process, and its HBM4E samples keep a Samsung Foundry 4 nm base die.
- Micron says in its fiscal 2026 annual report that its HBM4 uses advanced CMOS technology on the base logic die and the DRAM core dies.
Why move to logic at all? An HBM4 base die has to drive 2,048 data signals at well above 8 Gb/s while handling test and repair, which is easier with logic transistors and metal. A logic base die can also hold functions that used to sit in the host chip, starting with the memory controller. Rambus's HBM4E controller release of March 4, 2026 already offers the controller for use either in an AI SoC or in a custom base die.
That is where standard and custom HBM part company. Samsung said in February 2026 that it expects custom HBM samples to reach customers in 2027, and Micron said on September 30, 2026 that it is working with NVIDIA on what it called the industry's first custom HBM4E, named NVHBM. Who designs a custom base die, who pays for it and who owns its yield is a separate question, and we covered it in Custom HBM: who designs the base die, who pays and who owns yield. This report stays with the standard part.
SPHBM4: the same stack on a cheaper package
JEDEC added a second branch to the family on July 13, 2026, when it published JESD330-4, Standard Package High Bandwidth Memory (SPHBM4). It uses the same DRAM core dies as HBM4, so capacity per stack is the same, but replaces the base die with a new interface die that has a 512-bit interface instead of 2,048 signals. It runs that narrower interface at a higher frequency, with 4:1 serialization, to reach the same aggregate throughput as HBM4. The coarser bump pitch lets the stack mount on an organic substrate rather than silicon, and JEDEC says organic routing supports longer channels from the SoC to memory, which could allow more stacks per chip.
JEDEC did not say which products will use SPHBM4 or when. It gives engineers a second option: standard HBM4 for the highest bandwidth density on an interposer, SPHBM4 where package cost or reach matters more.
How it is built: TSVs, bonding and the 775 micrometer limit
An HBM stack is a pile of thinned DRAM dies with thousands of TSVs each, joined by micro-bumps with an underfill or film between them. Two bonding methods dominate.
- MR-MUF (mass reflow molded underfill) is SK hynix's method. SK hynix describes it as injecting a protective material between the stacked chips. SK hynix uses an advanced version for HBM4 and for its 12-high HBM4E samples, which it says reach 48 GB while keeping the structure stable.
- TC-NCF (thermo-compression with non-conductive film) is Samsung's method. Each die is pressed and heated with a thin film between layers. In February 2024 Samsung said its advanced TC-NCF reached a 7 micrometer gap between chips for 12-high HBM3E, kept the 12-high stack at the same height as an 8-high one, and used smaller bumps for signals and larger ones where heat must escape.
Height is the binding constraint, because the stack must fit the same package as the logic die beside it, and JEDEC fixes its thickness. Through HBM3E the standard thickness was 720 micrometers, according to a TrendForce report of July 7, 2026. For HBM4, JEDEC set a relaxed nominal thickness of 775 micrometers for both 12-high and 16-high stacks, as TrendForce reported on April 21, 2025, which it said reduced the immediate need for hybrid bonding.
That headroom is still tight for 16 dies. TrendForce, citing EE Times, reported on January 13, 2026 that SK hynix's 16-high HBM4 shown at CES 2026 thins each DRAM wafer to 30 micrometers to fit under 775. Sixteen core dies at 30 micrometers is 480 micrometers of DRAM silicon, which leaves under 300 micrometers for the base die, 16 bond layers and the mold on top. Our arithmetic, but it shows why every micrometer of bond line matters.
Hybrid bonding, which joins copper pads directly with no bump, would remove most of that bond-line height. It was once expected with HBM4. TrendForce reported on July 7, 2026, citing ZDNet Korea, that Samsung and SK hynix are reconsidering when to adopt it, that both kept conventional bonding for HBM4, and that 16-high HBM4E may be the earliest adoption point. The January report said SK hynix believes hybrid bonding becomes necessary above 20 layers, and the July report said JEDEC is discussing a limit of around 900 to 1,000 micrometers for HBM5. These are press reports, not company statements. The physics and the pitch roadmap of hybrid bonding are covered in our post Hybrid bonding from first principles.
Suppliers: where SK hynix, Samsung and Micron stand
All three makers reached volume HBM4 in the first half of 2026. The table lists only what each company has said publicly, with the date.
| Milestone | SK hynix | Samsung | Micron |
|---|---|---|---|
| HBM4 development or samples | Development complete and mass production readied, September 12, 2025 (over 10 Gb/s, 1b DRAM node, Advanced MR-MUF) | Not covered in the Samsung sources we read | Not covered in the Micron sources we read |
| HBM4 volume | Mass shipments began in Q2 2026, ramping in the second half (July 29, 2026); shipments "began in earnest" in Q3 (October 28, 2026) | Mass production and first commercial shipments, February 12, 2026 (11.7 Gb/s, 1c DRAM, 4 nm base die) | Volume shipment of 36 GB 12-high began in Q1 calendar 2026 (March 16, 2026) |
| 16-high HBM4 | 16-layer 48 GB displayed at COMPUTEX 2026 (newsroom, June 2, 2026) | 48 GB 16-high planned to customer timelines (February 12, 2026) | 48 GB 16-high samples shipping to customers (March 16, 2026) |
| HBM4E | 12-high 48 GB samples shipped, up to 16 Gb/s (June 18, 2026) | 12-high 48 GB samples, 14 Gb/s scalable to 16 Gb/s, up to 3.6 TB/s (UK newsroom, June 1, 2026) | Development well underway on 1-gamma DRAM; expects to ramp volume in calendar 2027 (fiscal 2026 annual report) |
| Custom HBM | Not addressed in the releases we read | Custom HBM samples expected to reach customers in 2027 (February 12, 2026) | Custom HBM4E (NVHBM) with NVIDIA (September 30, 2026) |
A few points behind the table:
- SK hynix said on October 28, 2026 that it expects HBM4 to surpass HBM3E in both shipment volume and revenue in the fourth quarter of 2026, and that it will ramp its M15X fab to maximum capacity by the end of 2026 to respond to HBM demand. Those are company expectations.
- Samsung claims 10% better thermal resistance and 30% better heat dissipation for HBM4 against HBM3E, and 16% better energy efficiency for HBM4E against the previous generation. Other outlets date Samsung's global HBM4E announcement to May 29, 2026; the copy we read is the UK newsroom version of June 1, 2026.
- Micron said on September 30, 2026 that it continues to execute well on its HBM4 ramp and expects initial output from its Singapore HBM advanced packaging facility in early calendar 2027. Pricing and supply are covered in our post on the AI memory squeeze.
"First" claims in this market are company claims, and we have not tried to referee them.
Which AI chips use HBM4
Two large GPU families have published HBM4 configurations.
- NVIDIA Rubin. NVIDIA's Rubin architecture blog (dated July 21, 2026 on the page) says Rubin integrates up to 288 GB of HBM4 per GPU, driven by dedicated HBM controllers and 12-high stacks, with up to 22 TB/s of memory bandwidth, 2.8 times Blackwell. If each 12-high stack is 36 GB, that is eight stacks, and 22 TB/s across eight stacks is 2.75 TB/s each, or roughly 10.7 Gb/s per pin. That inference is ours, not NVIDIA's. For comparison, TrendForce's report of NVIDIA's GTC 2025 keynote (March 19, 2025) gave Rubin 13 TB/s, so the faster HBM4 parts showed up directly in the GPU's specification. NVIDIA launched the Rubin platform on January 5, 2026 and said Rubin-based products would be available from partners in the second half of 2026. On May 31, 2026 it said Vera Rubin was ramping into full production, with production shipments set to begin in the autumn of 2026.
- AMD Instinct MI400 series. AMD launched the MI400 series at Advancing AI on July 23, 2026, with the MI455X powering its Helios rack. AMD's ROCm blog of August 4, 2026 gives 432 GB of HBM4 per MI455X and 31 TB of HBM4 across the 72 GPUs in a Helios rack. Twelve 36 GB stacks would make 432 GB, but the AMD pages we read do not state the stack count.
Not every new accelerator uses HBM4, and the exceptions are telling. NVIDIA's Rubin CPX, a context-processing GPU announced on September 9, 2025, uses 128 GB of GDDR7. NVIDIA said at the time that it expected Rubin CPX to be available at the end of 2026. Where capacity per dollar matters more than bandwidth per watt, GDDR still has a role.
The next step is HBM4E. TrendForce's report of GTC 2025 (March 19, 2025) placed Rubin Ultra, on HBM4E, in the second half of 2027. heise reported from GTC 2026 on March 17, 2026 that Rubin Ultra boards carry 16 HBM4E stacks for about one terabyte per GPU. TrendForce reported on August 4, 2026 that NVIDIA has been evaluating lower-spec options for Rubin Ultra since early in the third quarter of 2026, including 8-high HBM4E, 12-high HBM4 and 8-high HBM4, and that the final specification was undetermined. We treat the Rubin Ultra memory configuration as open.
What HBM4 means for chip and system engineers
For a team designing an accelerator, or a system around one, HBM4 changes the work in six places.
PHY and controller
The PHY now has to handle 2,048 data lines per stack at rates the vendors have taken past 10 Gb/s, while the standard's top bin is 8 Gb/s. That gap matters: the speed you can run in a product is set by the parts you can buy and the channel you can build, not by the JESD270-4 headline. IP vendors have aimed above the standard. Cadence announced on April 17, 2025 an HBM4 PHY and controller at 12.8 Gb/s, with the PHY as a hard macro for TSMC N3 and N2 and the controller as soft RTL, and a reference interposer design validated on a test chip with an HBM4 device. Rambus announced an HBM4E controller on March 4, 2026 at up to 16 Gb/s, or 4.1 TB/s per device. Because JEDEC says HBM4 is backward compatible with HBM3 controllers, a controller that supports both generations is a reasonable hedge where supply is uncertain.
Interposer routing
Eight HBM4 stacks mean 16,384 data signals to route between the host and the memory, against 8,192 for eight HBM3E stacks, before command, address, clock and power. That pushes interposer layer count, line and space, and the shoreline the host die must give up for HBM PHYs. A faster pin rate also tightens the channel budget for crosstalk and loss across the interposer. Our CoWoS versus EMIB decision guide covers the packaging choice that frames this work.
Power and thermal
Plan for more stack power even with better energy per bit. Samsung said doubling the data I/Os to 2,048 drives power and thermal challenges. Taller stacks hold heat further from the cold plate, and the logic base die adds its own power at the bottom. Power delivery also has to feed the new voltage rails JEDEC allows, which are vendor-specific, so a board or interposer designed for one supplier may need changes for another.
Test and known-good stacks
An HBM stack is only useful if it is known good before it is attached to a large and expensive host die, because a bad stack found after assembly can scrap the whole package. Built-in self-test on both sides of the link helps here: Cadence says its HBM4 IP includes RAS and BIST features for tuning memory performance in the field, and ships lab software for post-silicon bring-up. With a logic base die there is a new question of who tests it and to what coverage, which matters most for custom base dies and is set out in our custom HBM post.
Verification
Verification IP for HBM4 is available commercially. Cadence's release, for example, lists DFI VIP, an HBM4 memory model and a system performance analyzer. Teams should verify against the DRAM vendor's own timing and feature set, not only against the JEDEC baseline, because shipping parts run at bins above 8 Gb/s and include vendor-specific options. Refresh management, including DRFM, and error reporting deserve their own test plan.
Sizing the memory
The first architectural question is still how much memory a model needs and how much bandwidth it can use. The model memory sizer on The Semiconductor Hub estimates how much memory a given model needs, which turns the choice of stack count and stack height into a concrete number. Within a generation, TrendForce noted on August 4, 2026, stack height sets the trade between HBM capacity per GPU and the number of GPUs that can ship from a given supply of DRAM.
What to watch in 2027
These are the open questions, with whoever has made a forecast named.
- HBM4E in shipping products. Micron expects to ramp HBM4E volume in calendar 2027, SK hynix and Samsung have shipped 12-high samples, and Rubin Ultra was placed in the second half of 2027 in NVIDIA's GTC 2025 roadmap. Watch which stack height and per-pin rate reach volume first. TrendForce said on August 4, 2026 that Rubin Ultra I/O could reach 14 to 16 Gb/s if HBM4E meets its schedule, or 11 to 12 Gb/s with optimized HBM4.
- A JEDEC HBM4E standard. We found no published JEDEC HBM4E standard as of November 13, 2026. Vendor parts are running ahead of any standard, as they did with HBM3E.
- Custom base dies. Samsung expects custom HBM samples in 2027 and Micron's NVHBM targets NVIDIA's next GPUs. The test is a custom die shipping in a product.
- 16-high and hybrid bonding. 16-high HBM4 is sampling and 16-high HBM4E is the reported earliest point for hybrid bonding, according to TrendForce's July 7, 2026 report. Watch for any maker announcing hybrid bonding in a qualified HBM product.
- SPHBM4 adoption. A first named design would show whether organic-substrate HBM finds a place below the top accelerators.
- Supply. TrendForce projects HBM bit shipments to grow 50 to 60% in 2027, which it says is still insufficient for demand (August 4, 2026). Supply may decide which configurations ship.
Sources
- JEDEC, "JEDEC and Industry Leaders Collaborate to Release JESD270-4 HBM4 Standard", April 16, 2025, press release copy on Silicon UK: silicon.co.uk
- JEDEC, "JEDEC Publishes HBM3 Update to High Bandwidth Memory (HBM) Standard", January 27, 2022, copy on Design & Reuse: design-reuse.com
- JEDEC, "New JEDEC SPHBM4 Standard Enables HBM4-Class Bandwidth on Organic Substrates", July 13, 2026, copy on 01net: 01net.it
- Synopsys, Brett Murdock, "What Designers Need to Know About HBM3", April 16, 2022: synopsys.com
- SK hynix, "SK hynix Completes World's First HBM4 Development and Readies Mass Production", September 12, 2025: news.skhynix.com
- SK hynix, "SK hynix Ships Samples of 12-Layer Next-Gen 'HBM4E'", June 18, 2026: news.skhynix.com
- SK hynix, "SK hynix Announces 2Q26 Financial Results", July 29, 2026: news.skhynix.com
- SK hynix, third quarter 2026 business results, October 28, 2026: news.skhynix.com
- SK hynix, COMPUTEX 2026 newsroom feature, June 2, 2026: news.skhynix.com
- TrendForce News, "SK hynix Partners with TSMC to Strengthen HBM Technological Leadership", April 19, 2024: trendforce.com
- Samsung Semiconductor, "Samsung Ships Industry-First Commercial HBM4 With Ultimate Performance for AI Computing", February 12, 2026: news.samsungsemiconductor.com
- Samsung Newsroom UK, "Samsung Begins Shipment of Industry-First HBM4E Samples", June 1, 2026: news.samsung.com
- Samsung Semiconductor, "Samsung Develops Industry-First 36GB HBM3E 12H DRAM", February 27, 2024: news.samsungsemiconductor.com
- Micron Technology, "Micron in High-Volume Production of HBM4 Designed for NVIDIA Vera Rubin, PCIe Gen6 SSD and SOCAMM2", March 16, 2026: investors.micron.com
- Micron Technology, annual report on Form 10-K for the fiscal year ended September 3, 2026: sec.gov
- Micron Technology, "Fiscal Q4 2026 Earnings Call Prepared Remarks", September 30, 2026: investors.micron.com (PDF)
- NVIDIA Technical Blog, Eduardo Alvarez, Vishal Mehta and Farshad Ghodsian, "Inside NVIDIA Rubin GPU Architecture: Powering the Era of Agentic AI", page dated July 21, 2026: developer.nvidia.com
- NVIDIA, "NVIDIA Kicks Off the Next Generation of AI With Rubin: Six New Chips, One Incredible AI Supercomputer", January 5, 2026: nvidianews.nvidia.com
- NVIDIA, "NVIDIA Vera Rubin Ramps Into Full Production to Power Agentic AI Factories Worldwide", May 31, 2026: investor.nvidia.com
- NVIDIA, "NVIDIA Unveils Rubin CPX: A New Class of GPU Designed for Massive-Context Inference", September 9, 2025: nvidianews.nvidia.com
- TrendForce News, "GTC 2025: NVIDIA Unveils More on Rubin GPUs, Announces Feynman for 2028 in Roadmap Update", March 19, 2025: trendforce.com
- heise online, Mark Mantel, report on Rubin Ultra boards at GTC 2026 (Kyber rack), March 17, 2026: heise.de
- AMD, "AAI 2026: AMD Launches AMD Instinct MI400 Series GPUs for Frontier AI, HPC", July 23, 2026: newsroom.amd.com
- AMD ROCm Blogs, "Introducing AMD CDNA 5 and the AMD Helios Rackscale Solution", August 4, 2026: rocm.blogs.amd.com
- TrendForce News, "JEDEC Finalizes HBM4 Standard, Potentially Easing Pressure on Hybrid Bonding Adoption", April 21, 2025, citing JEDEC and ZDNet Korea: trendforce.com
- TrendForce News, "SK hynix May Stick With MR-MUF for HBM4 16-High amid Fluxless Bonding Evaluation", January 13, 2026, citing DealSite and EE Times: trendforce.com
- TrendForce News, "Samsung, SK hynix Reportedly Reconsider Hybrid Bonding Timeline; 16-High HBM4E May Be Earliest Adoption", July 7, 2026, citing ZDNet Korea: trendforce.com
- TrendForce, "HBM4 Raises the Bar on Manufacturing Complexity, Premium Expected to Exceed 30%", May 22, 2025: trendforce.com
- TrendForce, "DRAM Supply to Remain Tight in 2027, Prompting NVIDIA to Lower HBM Configurations for Rubin Ultra", August 4, 2026: trendforce.com
- Cadence, "Cadence Enables Next-Gen AI and HPC Systems with Industry's Fastest HBM4 12.8Gbps IP Memory System Solution", April 17, 2025, copy on Silicon UK: silicon.co.uk
- Rambus, "Rambus Announces Industry-First HBM4 Controller IP to Accelerate Next-Generation AI Workloads", September 9, 2024: rambus.com
- Rambus, "Rambus Sets New Benchmark for AI Memory Performance with Industry-Leading HBM4E Controller IP", March 4, 2026: rambus.com
Specifications are as published by JEDEC and the companies on the dates shown, and performance, power and thermal claims are the companies' own, measured on their own baselines. Forecasts belong to the firms named and are not ours. Several items are press reports the companies have not confirmed, and they are marked as such. This report is technical commentary and is not investment advice.



Your comments will be moderated before it appears here.