An HMB3 able to processing over 665 GB per second at 5.2 Gbps I / O pace.
SK Hynix up to date a devoted HBM2E (Excessive Bandwidth Reminiscence) web page on their website, the place the corporate slipped info on the following era, the HMB3. This guarantees a bandwidth enhance of round 44% in comparison with HBM2E.
From the outset, allow us to specify that the JEDEC has not but formalized a normal regarding HMB3; moreover, SK Hynix specifies that this era remains to be underneath growth. However as talked about within the picture above, the agency expects a rise in bandwidth of round 44%: the bandwidth per die would go from 460 Gb / s to 665 Gb / s and the pace for the I / O from 3.6 Gbit / s to five.2 Gbit / s. These values should not very completely different from these talked about in 2016: on the time, in an article dedicated to the HMB3, we reported rumors suggesting a bandwidth of 512 GB / s per die.
OpenFive designed a RISC-V SoC engraved in 5 nm and outfitted with HBM3 reminiscence at 7.2 Gbit / s
Reminiscence reserved for the HPC sector
At present, HBME2 primarily issues bandwidth-hungry gadgets comparable to very high-end GPUs or FPGAs: NVIDIA’s Ampere A100 GPU, for instance, has 80 GB of HBM2E on board. These gadgets usually use 4-6 dies of HBM2E reminiscence; this offers a theoretical bandwidth of 1.84 to 2.76 TB / s. In absolute phrases, with HBM3, it could be prone to attain 2.66 to three.99 TB / s.
SK Hynix has not communicated any date relating to the advertising of the HBM3.