HBM4 is the memory generation defined by JEDEC's JESD270-4 standard, announced on 16 April 2025. It doubles HBM3E's interface from 1,024 to 2,048 bits, according to Samsung's 12 February 2026 comparison, with JEDEC specifying up to 2 TB/s per stack. Early HBM4 GPUs include NVIDIA Rubin, whose Vera Rubin production shipments began in August 2026 according to NVIDIA's 26 August earnings call, and AMD's MI455X, launched with the Helios rack on 23 July 2026, with volume deployments expected in the second half of 2026 according to AMD's product page as of 1 October 2026.
For a rental decision, keep three things separate: the standard, the memory supplier's product rating, and the accelerator's specification. Only the last tells you the capacity and bandwidth of the GPU you are evaluating. A faster memory stack is a useful ingredient, but it is not an inference benchmark.
HBM4 vs HBM3E: the standard and the product rating
The standard comparison below deliberately identifies where HBM3E figures come from a product. Micron's HBM3E page, as of 28 September 2026, specifies speeds above the original HBM3 standard. Treating that product rating as a JEDEC limit would blur the distinction this comparison needs. For the physical basics, see HBM versus GDDR.
| Property | HBM3E comparison point and source | HBM4 standard and source |
|---|---|---|
| Interface width per stack | 1,024 I/Os; Micron HBM3E page, as of 28 September 2026 | 2,048 bits; JEDEC, 16 April 2025 |
| Independent channels | 16 in the original HBM3 standard; JEDEC, 27 January 2022 | 32, each with two pseudo-channels; JEDEC, 16 April 2025 |
| Pin speed | VENDOR CLAIM: greater than 9.2 Gb/s; Micron HBM3E page, as of 28 September 2026 | Up to 8 Gb/s; JEDEC, 16 April 2025 |
| Bandwidth per stack | VENDOR CLAIM: more than 1.2 TB/s; Micron HBM3E page, as of 28 September 2026 | Up to 2 TB/s; JEDEC, 16 April 2025 |
| Stack heights | Micron products: 8-high and 12-high, as of 28 September 2026; original HBM3 standard: 4-, 8-, 12-high, with provision for future 16-high, JEDEC, 27 January 2022 | 4-, 8-, 12- and 16-high; JEDEC, 16 April 2025 |
| Capacity per stack | Micron products: 24 GB at 8-high, 36 GB at 12-high; page as of 28 September 2026 | Up to 64 GB using sixteen 32 Gb dies; JEDEC, 16 April 2025 |
The lower pin-speed number in the HBM4 standard column is not a mistake. Width and pin speed are separate inputs. By our arithmetic, 2,048 divided by 1,024 gives a 2.0-fold wider interface. At equal pin speed, raw throughput doubles. That calculation does not predict how much faster a particular GPU will run your model.
Use the units as a quick screening test. A pin-speed claim belongs in Gb/s, while the stack and GPU figures here use TB/s. Before comparing two announcements, write down what each figure measures. Ask whether it describes one memory stack, one GPU package or a larger system. Then preserve qualifiers such as peak, theoretical and up to. Those words define the claim you are evaluating.
Capacity needs its own check. Micron's 16 March 2026 announcement specified 36 GB for its volume-shipping 12-high HBM4. Its HBM3E page, as of 28 September 2026, also listed a 36 GB 12-high product. In that comparison, changing memory generation increases the bandwidth opportunity without increasing stack capacity. Do not read the standard's maximum capacity as a promise for every accelerator.
SK hynix HBM4, Samsung and Micron HBM4 milestones
Read the milestone before comparing the speed. A sample shipment, a prepared production line and volume shipments answer different procurement questions. SK hynix's 12 September 2025 announcement described completed development and production readiness. Its 29 July 2026 results subsequently said mass shipments began in the second quarter of 2026.
| Maker | Milestone and date | Speed claimed | Base die approach |
|---|---|---|---|
| SK hynix | Mass shipments began Q2 2026; SK hynix results, 29 July 2026 | VENDOR CLAIM: over 10 Gb/s; SK hynix, 12 September 2025 | TSMC advanced logic planned for HBM4; SK hynix partnership announcement, 19 April 2024 |
| Samsung | Mass production begun and commercial products shipped; Samsung, 12 February 2026 | VENDOR CLAIM: consistent 11.7 Gb/s, headroom up to 13 Gb/s, maximum 3.3 TB/s per stack; same announcement | 4nm logic base die; same announcement |
| Micron | Volume shipments of 36 GB 12-high HBM4 began in calendar Q1 2026 for Vera Rubin; Micron, 16 March 2026 | VENDOR CLAIM: over 11 Gb/s and greater than 2.8 TB/s per stack; same announcement | Internally developed and manufactured advanced CMOS for HBM4; Micron remarks, 25 June 2025 |
Reported qualification is a separate category. Reuters on 26 January 2026 relayed Korea Economic Daily's report that Samsung passed NVIDIA and AMD HBM4 tests, citing industry sources. Samsung declined comment on its planned NVIDIA supply. That is reported qualification news, not an on-record confirmation from either GPU maker. Samsung's own February shipment announcement is the clearer manufacturing milestone.
Keep the speed qualifiers intact. Samsung's February announcement distinguishes consistent operation from headroom. Micron's March announcement attributes its speed result to internal testing and confidential customer test-vehicle validation. These are VENDOR CLAIMS about memory products. They do not establish the operating speed of every GPU that uses those products.
The base die is now part of the comparison
SK hynix's 19 April 2024 announcement places the base die beneath the DRAM stack, connected to the GPU. The company planned to replace its HBM3E-era proprietary process with TSMC advanced logic for HBM4. Its stated benefit was room for more functions and customer-specific performance and power optimization: a VENDOR CLAIM, not a measured application gain.
Foundry-made logic therefore belongs in the HBM discussion. Samsung's 12 February 2026 announcement identifies a 4nm logic base die. But do not assume every supplier uses the same route. Micron's 25 June 2025 remarks described an internally manufactured HBM4 base die; its 23 September 2025 roadmap assigned TSMC manufacturing to standard and customized HBM4E base logic dies. Keep HBM4 and that later roadmap step separate.
HBM4E is already sampling
The next step is already in customers' hands as samples. Samsung shipped 12-layer, 48 GB HBM4E samples on 29 May 2026; VENDOR CLAIM: a stable 14 Gb/s per pin, scalable to 16 Gb/s, and up to 3.6 TB/s per stack. Samsung tied mass production to customer schedules without a fixed date. SK hynix shipped its own 12-layer, 48 GB HBM4E samples on 18 June 2026, with a VENDOR CLAIM of up to 16 Gb/s per pin. ANALYST ESTIMATE: TrendForce's 29 September 2026 outlook expects HBM4E to ramp gradually in the second half of 2027. For a renter, HBM4E matters from 2027, when Rubin Ultra is planned; until then the choice is between the HBM3E GPUs in the tables below and early HBM4 systems.
Which accelerators specify HBM4 memory
Read every capacity below per GPU package. NVIDIA's 21 July 2026 architecture article describes Rubin as two compute dies unified into one GPU. Do not double its published memory figure because the package contains two dies. The broader Vera Rubin platform guide covers the system around it.
| Accelerator | Memory per GPU | Bandwidth per GPU | Status and source date |
|---|---|---|---|
| NVIDIA Rubin | Up to 288 GB HBM4; NVIDIA architecture article, 21 July 2026 | VENDOR CLAIM: up to 22 TB/s peak in that article; 19.2 TB/s on NVIDIA's platform page as of 1 October 2026 | Vera Rubin production shipments began August 2026; NVIDIA earnings call, 26 August 2026 |
| NVIDIA Rubin Ultra | Demonstrated with 1 TB HBM4E; Tom's Hardware, 17 March 2026; not a shipping configuration | No settled figure | Demonstrated as a planned 2027 product; ANALYST ESTIMATE: TrendForce reported the final memory specification undecided and alternatives under evaluation, 4 August 2026 |
| AMD MI400 preview | 432 GB HBM4 engineering projection; AMD Advancing AI 2025 deck, as of 1 October 2026 | VENDOR CLAIM: projected 19.6 TB/s; same deck | Engineering preview for 2026, not a shipment confirmation |
| AMD MI450 Series description | Up to 432 GB HBM4; AMD Helios blog, 14 October 2025 | VENDOR CLAIM: 19.6 TB/s; same blog | Dated series description; distinguish it from the current named part |
| AMD MI455X | 432 GB HBM4; AMD MI400 Series page, as of 1 October 2026 | VENDOR CLAIM: up to 23.3 TB/s peak theoretical; same page | AMD launched the MI400 series and Helios on 23 July 2026 and called Helios in production; AMD's page, as of 1 October 2026, expects volume deployments in the second half of 2026 |
Two discrepancies matter. NVIDIA's July architecture article and its platform page as of 1 October 2026 give different Rubin bandwidth figures. AMD's Helios page as of 1 October 2026 also retains 19.6 TB/s in its compute-tray description while its FAQ gives up to 23.3 TB/s; both are VENDOR CLAIMS. Use the specification attached to the system being offered; do not silently choose the larger number.
Rubin Ultra needs an additional qualification. Tom's Hardware's 17 March 2026 report described a demonstration, while TrendForce's 4 August 2026 report said the final memory specification remained undecided. ANALYST ESTIMATE: TrendForce reported evaluation of HBM4E and HBM4 alternatives. The demonstrated capacity belongs to that demonstration, not a settled shipping configuration. Follow the NVIDIA GPU roadmap for that distinction.
Nor should other accelerators be classified from bandwidth alone. Google's 22 April 2026 TPU 8t/8i comparison and Amazon's 27 February 2026 Trainium4 announcement do not identify an HBM generation. They are not confirmed HBM4 entries in this comparison.
Supply forecasts do not set your rental bill
ANALYST ESTIMATE: TrendForce on 22 May 2025 forecast an HBM4 generational price premium above 30%. That was a forecast, not a disclosed purchase contract. ANALYST ESTIMATE: its 29 September 2026 outlook forecast a 121% year-over-year increase in blended HBM average selling prices for 2027. The latter covers the HBM mix; it is not an HBM4-over-HBM3E premium.
ANALYST ESTIMATE: TrendForce's September outlook also said HBM and conventional DRAM compete for limited advanced-process and wafer capacity, keeping supply tight through 2027. Use that as supply-planning context. It does not provide a conversion from memory-component costs to a cloud GPU bill. Compare actual offers and the work completed within your latency target.
Who supplies the memory is shifting too. ANALYST ESTIMATE: Counterpoint put second-quarter 2026 HBM revenue shares at 50% for SK hynix, 33% for Samsung and 18% for Micron, as reported by Seoul Economic Daily on 3 September 2026; most of that revenue was still HBM3E, so these are not HBM4 shares. Micron said on 30 September 2026 that agreements already covered the vast majority of its calendar-2027 HBM supply, at significant year-over-year price increases it did not quantify.
Rent for capacity first, then measure bandwidth's value
HBM4 GPUs are not in this site's rental tables yet, as of 1 October 2026. That is a coverage limit, not a claim that no cloud has deployed Rubin. CoreWeave announced Vera Rubin availability on its cloud on 30 September 2026; NVIDIA's account that day described access for early-access customers. Treat that dated announcement separately from the live comparison below.
Capacity and bandwidth are central to large-model inference. NVIDIA's 17 November 2023 inference guide explains that transfers of weights, keys, values and activations can dominate decode latency. It also identifies KV-cache growth with sequence length and batch size as a throughput constraint. A capacity check must include the intended context and concurrency, not just model weights.
Bandwidth is not the only constraint. The Scaling Book, as of 28 September 2026, explains that batching can make decode's feed-forward operations compute-bound, while each request retains its own KV cache. Test the workload you intend to serve. A low-concurrency latency test and a heavily batched throughput test answer different questions.
The live specification table compares H200, B200, B300, MI325X and MI355X on capacity, memory type and bandwidth.
The live rental table shows the current comparison for those HBM3E families.
Use the GPU specification reference to check family ranges, then inspect the actual configuration. Start your comparison with the H200 rental page and MI355X rental page. The AMD Instinct comparison covers the AMD options in more detail.
The decision rule: first reject any GPU whose memory cannot hold the weights plus the KV cache at your context length and concurrency. Among those that fit, run the same model, precision and request mix, and keep the cheapest configuration that meets your latency target at the live price for the GPU count you need. If a memory-bound workload still misses its target on the fastest HBM3E part, that measured gap is your reason to evaluate an HBM4 system once you can get one. The memory-generation label alone is not.
Sources
- JEDEC HBM4 standard announcement, 16 April 2025.
- JEDEC HBM3 standard announcement, 27 January 2022.
- Micron HBM3E product page, as of 28 September 2026.
- SK hynix HBM4 development, 12 September 2025.
- SK hynix quarterly results, 29 July 2026.
- SK hynix and TSMC partnership, 19 April 2024.
- Samsung commercial HBM4 announcement, 12 February 2026.
- Reuters via CNA: reported Samsung qualification, 26 January 2026.
- Micron HBM4 volume shipments, 16 March 2026.
- Micron prepared remarks, 25 June 2025.
- Micron prepared remarks, 23 September 2025.
- NVIDIA Rubin architecture, 21 July 2026.
- NVIDIA platform product specifications, as of 1 October 2026.
- NVIDIA earnings call, 26 August 2026.
- Tom's Hardware: Rubin Ultra demonstration, 17 March 2026.
- TrendForce: supply outlook and Rubin Ultra configuration review, 4 August 2026.
- AMD Advancing AI 2025 deck, as of 1 October 2026.
- AMD Helios blog, 14 October 2025.
- AMD: MI400 series and Helios launch, 23 July 2026.
- AMD MI400 Series product page, as of 1 October 2026.
- AMD Helios product page, as of 1 October 2026.
- Google TPU technical comparison, 22 April 2026.
- Amazon and OpenAI announcement, 27 February 2026.
- TrendForce: HBM4 premium forecast, 22 May 2025.
- TrendForce: HBM pricing and capacity outlook, 29 September 2026.
- CoreWeave Rubin cloud announcement, 30 September 2026.
- NVIDIA account of CoreWeave early access, 30 September 2026.
- NVIDIA inference optimization guide, 17 November 2023.
- The Scaling Book: inference, as of 28 September 2026.
- Samsung: HBM4E samples, 29 May 2026.
- SK hynix: 12-layer HBM4E samples, 18 June 2026.
- Seoul Economic Daily: Counterpoint HBM market shares, 3 September 2026.
- Micron: fiscal Q4 2026 prepared remarks (PDF), 30 September 2026.