Wednesday, September 24th 2025

Micron's 12-Hi HBM4 Delivers 2.8 TB/s Bandwidth, 11 Gb/s Per-Pin Transfer Rate

Yesterday, Micron reported its fourth-quarter and full-year fiscal 2025 results, beating the consensus and providing interesting comments on its upcoming HBM technology. During the earnings call, Sanjay Mehrotra, CEO of Micron, confirmed that HBM4—the company's next-generation memory—will be available next year with several enhancements over the base HBM4 JEDEC specification. "Micron Technology's HBM4 12-high remains on track to support customer platform ramps, even as the performance requirements for HBM4 bandwidth and pin speeds have increased. We have recently shipped customer samples of our HBM4 with industry-leading bandwidth exceeding 2.8 TB/s and pin speeds over 11 Gbps." The JEDEC specification for HBM4 is 2 TB/s bandwidth and 8 Gb/s pin transfer across a 2048-bit interface. Micron is pushing this to 11 Gb/s, resulting in 40% higher bandwidth at 2.8 TB/s.

There is a simple reason why this is the case. Customers like NVIDIA are asking HBM makers to create a more performant HBM memory to keep up with the exponential growth that their compute dies are experiencing. To satisfy these requirements, Micron is moving beyond JEDEC specifications. "We believe Micron Technology's HBM4 outperforms all competing HBM4 products, delivering industry-leading performance as well as best-in-class power efficiency. Our proven 1-gamma DRAM, innovative and power-efficient HBM4 design, in-house advanced CMOS base die, and advanced packaging innovations are key differentiators enabling this best-in-class product," said Sanjay.
"For HBM4E, Micron Technology will offer standard products as well as the option for customization of the base logic die," confirmed Micron CEO. As we reported earlier, NVIDIA and AMD accelerators will utilize a customized HBM memory for the first time. In an HBM stack, consisting of stacked DRAM dies up to a 12-high, connected with TSVs, there is the possibility of embedding an optional base die with customized logic/accelerator circuitry tailored to a specific need. It appears that NVIDIA and AMD are not only exploring this option to gain more performance, but are also actively working on an implementation with a custom base that will push these accelerators ahead of any third-party ASIC solution. The custom base die will likely be a data processing/logic die that helps route data packets more efficiently, cutting latency and improving performance.
Source: Micron Earnings Call
Add your own comment

5 Comments on Micron's 12-Hi HBM4 Delivers 2.8 TB/s Bandwidth, 11 Gb/s Per-Pin Transfer Rate

#1
Philaphlous
Is there really any advantage to gddr7 over hbm? With these speeds I'd think there'd be 0 bottleneck for video memory... Overclocking memory results in some pretty decent frame rate bumps but if we had memory this fast, heck who knows how amazing gpu's would be. Nvidia could refresh the same Blackwell chips with hbm and there'd be like a 10-20% performance bump at minimum...
Posted on Reply
#2
LastDudeALive
PhilaphlousIs there really any advantage to gddr7 over hbm? With these speeds I'd think there'd be 0 bottleneck for video memory... Overclocking memory results in some pretty decent frame rate bumps but if we had memory this fast, heck who knows how amazing gpu's would be. Nvidia could refresh the same Blackwell chips with hbm and there'd be like a 10-20% performance bump at minimum...
HBM is insanely expensive, and requires a completely different integrated packaging design. It's not as simple as just dropping HBM chips where the GDDR7 chips are now. It requires completely new chip to manage the memory, and it requires integrated packaging instead of a PCB because there's too many lanes. So the entire GPU would have to be redesigned and remade. Not many people would pay $800 for a 5060 with HBM.
Posted on Reply
#3
Wirko
LastDudeALiveHBM is insanely expensive, and requires a completely different integrated packaging design. It's not as simple as just dropping HBM chips where the GDDR7 chips are now. It requires completely new chip to manage the memory, and it requires integrated packaging instead of a PCB because there's too many lanes. So the entire GPU would have to be redesigned and remade. Not many people would pay $800 for a 5060 with HBM.
To add to that, it requires a 1024-bit memory controller per stack (twice as much for HBM4). Also I don't remember ever seeing a processor with a single stack, there are always at least two - the probable reason is that a single stack isn't economically justified.
Posted on Reply
#4
mechtech
"bandwidth exceeding 2.8 TB/s"

nice

almost enough to keep up with the wife's credit card transactions ;)

Be good for an on board mobo igpu chip
Posted on Reply
#5
BoxCutter
PhilaphlousIs there really any advantage to gddr7 over hbm? With these speeds I'd think there'd be 0 bottleneck for video memory... Overclocking memory results in some pretty decent frame rate bumps but if we had memory this fast, heck who knows how amazing gpu's would be. Nvidia could refresh the same Blackwell chips with hbm and there'd be like a 10-20% performance bump at minimum...
Should anything go wrong with a memory chip, it's over(for that specific chip at least) putting these on an interposer and on a substrate is FAB type procedure, meanwhile desoldering and resoldering a memory chip to a PCB while takes some finesse, is really not that difficult. And yeah, they are pretty expensive on top of that.
Posted on Reply
Nov 22nd, 2026 15:49 CST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts