Extremely-wide HPC Reminiscence to Attain 8 GT/s
SK Hynix was one of many key builders of the unique HBM reminiscence again in 2014, and the corporate actually hopes to remain forward of the trade with this premium kind of DRAM. On Tuesday, buried in a word about qualifying the corporate’s 1bnm fab course of, the the producer remarked for the primary time that it’s engaged on next-generation HBM3E reminiscence, which can allow speeds of as much as 8 Gbps/pin and shall be obtainable in 2024.
Contemporary HBM3 memory from SK Hynix and different distributors helps knowledge switch charges as much as 6.4Gbps/pin, so HBM3E with an 8 Gbpis/pin switch fee will present a reasonable, 25% bandwidth benefit over present reminiscence units.
To place this in context, with a single HBM stack utilizing a 1024-bit extensive reminiscence bus, this could give a identified good stack die (KGSD) of HBM3E round 1 TB/sec of bandwidth, up from 819.2 GB/sec in case of HBM3 at present. Which, with trendy HPC-class processors using half a dozen stacks (or extra), would work out to a number of TB/sec of bandwidth for these high-end processors.
In line with the corporate’s word, SK Hynix intends to begin sampling its HBM3E reminiscence within the coming month, and provoke quantity manufacturing in 2024. The reminiscence maker didn’t reveal a lot in the best way of particulars about HBM3E (actually, that is the primary public point out of its specs in any respect), so we have no idea whether or not these units shall be drop-in appropriate with present HBM3 controllers and bodily interfaces.
| HBM Reminiscence Comparability | ||||||
| HBM3E | HBM3 | HBM2E | HBM2 | |||
| Max Capability | ? | 24 GB | 16 GB | 8 GB | ||
| Max Bandwidth Per Pin | 8 Gb/s | 6.4 Gb/s | 3.6 Gb/s | 2.0 Gb/s | ||
| Variety of DRAM ICs per Stack | ? | 12 | 8 | 8 | ||
| Efficient Bus Width | 1024-bit | |||||
| Voltage | ? | ? | 1.2 V | 1.2 V | ||
| Bandwidth per Stack | 1 TB/s | 819.2 GB/s | 460.8 GB/s | 256 GB/s | ||
Assuming SK hynix’s HBM3E growth goes in keeping with plan, the corporate ought to have little bother lining up prospects for even quicker reminiscence. Particularly with demand for GPUs going by way of the roof to be used in constructing AI coaching and inference methods, NVIDIA and different processor distributors are greater than willing to pay premium for advanced memory they should produce ever quicker processors throughout this boom period in the industry.
SK Hynix shall be producing HBM3E reminiscence utilizing its 1b nanometer fabrication expertise (fifth Era 10nm-class node), which is at present getting used to make DDR5-6400 reminiscence chips which can be set to be validated for Intel’s subsequent technology Xeon Scalable platform. As well as, the manufacturing expertise shall be used to make LPDDR5T reminiscence chips that can mix excessive efficiency with low energy consumption.
