HBM2 has had a tough 12 months. The reminiscence commonplace hasn’t adopted the trail of its predecessors. If it had, we must always’ve already seen widespread rollouts from AMD and Nvidia throughout midrange and high-end merchandise. As a substitute, AMD’s Vega 56 and 64 are at the moment the one shopper GPUs to hold the reminiscence commonplace in any respect.
One of many rumors that dogged HBM2 final 12 months was that clock speeds had been effectively under what was focused Certainly, neither of AMD’s GPUs hit the anticipated 2Gbps per-pin sign charges (Hynix added 2.0Gbps speeds to its HBM2 choices in August 2016, then eliminated them in 2017). Samsung, nonetheless, might have discovered an answer to this drawback, and plans to introduce new HBM2 clocked as excessive as 2.4Gbps per pin.
The brand new HBM2, codenamed Aquabolt, can be clocked 1.2x increased than the earlier HBM2 commonplace and 1.5x increased than the RAM on AMD’s RX Vega 56. The opposite main benefit for increased per-pin bandwidth is it theoretically reduces the necessity for added HBM2 stacks.
This functionality might be extra vital than you may assume. Rumors have recommended that the manufacturing difficulties with HBM2 have centered across the interposer — particularly, the problem of aligning and connecting the through-silicon vias that run by way of the reminiscence stacks and into the interposer itself earlier than connecting with the GPU. A few of you might bear in mind the next doc from AMD’s numerous discussions of HBM.
That picture solely exhibits one HBM stack and its relation to an related SoC die. AMD’s GPUs discipline two HBM2 stacks, and as much as 4 are doable. Samsung’s 2.4Gbps per-pin signaling isn’t going to be sufficient for a single-stack Vega 56 in contrast with the present mannequin, however a single 1024-bit HBM2 hyperlink @ 2.4Gbps would supply 307.2GB/s of reminiscence bandwidth — 1.2x greater than the present bandwidth 256GB/s AMD’s RX 580 delivers.
Samsung has additionally managed to hit these elevated clock velocity with out growing HBM2’s voltage. The corporate’s first-generation HBM2 may run at 1.6Gbps at 1.2v or 2.0Gbps at 1.35v. These second-generation chips can run at 2.4Gbps at 1.2v. Samsung’s new chips used unspecified methods associated to TSV design and thermal management to attach the 5,000 TSVs per die and restrict clock skew. It additionally added extra thermal bumps between the HBM2 dies to enhance warmth dissipation and now provides a protecting layer on the backside of the HBM2 stack to enhance bodily energy.
Samsung didn’t give a ship date for its new Aquabolt chips, however did word Aquabolt is already in mass manufacturing. It’ll take some time for brand spanking new GPU designs to ship that use HBM2, however it’s doable the beleaguered commonplace has lastly turned the nook.