Simply when Nvidia is prepping its Blackwell GPUs to make the most of HBM3e reminiscence modules, the JEDEC Strong State Expertise Affiliation has introduced that the following model, HBM4, is close to completion. HBM3e, an enhanced variant of the prevailing HBM3 reminiscence, tops out at 9.8 Gbps, however HBM4 is prone to attain the double-digit 10+ Gbps pace.
HBM4, additionally an evolutionary step past the present HBM3 normal, additional enhances knowledge processing charges whereas sustaining important options corresponding to increased bandwidth, decrease energy consumption, and elevated capability per die and/or stack. Its options and capabilities are essential in functions that require environment friendly dealing with of huge datasets and sophisticated calculations, together with generative synthetic intelligence (AI), high-performance computing (HPC), high-end graphics playing cards, and servers.
For a begin, HBM4 comes with a bigger bodily footprint because it introduces a doubled channel rely per stack in comparison with HBM3. It additionally options completely different configurations that require numerous interposers to accommodate the differing footprints. Subsequent, it can specify 24-Gb and 32-Gb layers with choices for supporting 4-high, 8-high, 12-high and 16-high TSV stacks.
There are media studies about JEDEC having eased reminiscence configurations by decreasing thickness of HBM4 to 775 µm for 12-layer, 16-layer HBM4 resulting from rising complexity at increased thickness ranges. Nonetheless, whereas HBM producers like Micron, SK hynix, and Samsung had been poised to make use of hybrid bonding expertise, the HBM4 design committee is reportedly of the view that hybrid bonding would improve pricing. That, in flip, will make HBM4-powered AI processors costlier.
Hybrid bonding allows reminiscence chip designers so as to add extra stacks compactly with out the necessity for through-silicon-via (TSV), which makes use of filler bumps to attach a number of stacks. Nonetheless, with a thickness of 775 µm, hybrid bonding will not be wanted for the HBM4 type issue.
For compatibility, the brand new spec will be certain that a single controller can work with each HBM3 and HBM4 if wanted. The designers of the HBM4 spec have additionally reached an preliminary settlement on pace bins as much as 6.4 Gbps with dialogue ongoing for increased frequencies.
Associated Content material
- SK Hynix Speeds HBM Roadmap as AI Demand Soars
- HBM reminiscence chips: The unsung hero of the AI revolution
- SK Hynix Boosting HBM Output with $14.6 Billion Funding
- A sneak peek at HBM chilly warfare between Samsung and SK hynix
- Excessive-bandwidth reminiscence (HBM) choices for demanding compute
<!–
VIDEO AD
–>
<!–
div-gpt-ad-inread
–>
googletag.cmd.push(operate() { googletag.show(‘div-gpt-ad-inread’); });
googletag.cmd.push(operate() { googletag.show(‘div-gpt-ad-native’); });
–>
The submit How will HBM4 affect the AI-centric reminiscence panorama? appeared first on EDN.
