Market for load reduction
UPDATE: added 06/06/2012: info on LRDIMM market
In summary, if you follow the constraints outlined in the other articles here:
– lack of 8Gbit DRAM die (monolithic) for foreseeable future
– 32GB RDIMM only available in 4-rank for foreseeable future (due to lack of 8Gbit DRAM die)
– 4-rank having abysmal speeds at 2 DPC even i.e. not just 3 DPC (due to high memory loading)
– 32GB RDIMM having abysmal speeds at 2 DPC even i.e. not just 3 DPC (because only available in 4-rank)
The conclusion is that you need “load reduction” for 32GB memory modules (to avoid speed slowdown with the 32GB RDIMMs 4-rank) at 3 DPC, 2 DPC and possibly even 1 DPC (see IBM docs which show slowdown for 4-rank even at 1 DPC).
In addition, as examined in the articles which discuss memory speeds on IBM, HP, DELL and SuperMicro servers:
– 16GB RDIMMs (2-rank) experience slowdown at 3 DPC
The conclusion is that you also need “load reduction” for 16GB memory modules at 3 DPC.
Rank multiplication IP
32GB RDIMMs (4-rank) cannot be used at 3 DPC for another reason.
Use at 3 DPC would exceed the “8 ranks per memory channel limit” for current systems.
For this reason 32GB RDIMMs (4-rank) will work at 1 DPC and 2 DPC with the speed slowdowns described above, but will not work at 3 DPC.
Netlist HyperCloud and LRDIMMs both use rank multiplication IP to overcome this problem.
HyperCloud vs. LRDIMMs
HyperCloud trumps LRDIMMs on latency and achievable bandwidth.
And the fact that Netlist owns the IP on load reduction and rank multiplication – and have just won patent reexams at the USPTO where all claims survived reexamination, leading to a difficult situation for Inphi in Netlist vs. Inphi (which is stayed pending the reexams). Inphi is the ONLY maker of buffer chipsets for LRDIMMs.
In addition DDR4 is copying HyperCloud even more (dropping the asymmetrical lines and centralized buffer on the LRDIMM for an even greater following of the HyperCloud’s symmetric lines and decentralized buffer).
DDR4 – low voltage, higher speed
DDR4 will have lower voltage and higher speed, which will increase the need for load reduction.
It is likely that the entire memory space (3 DPC, 2 DPC and 1 DPC) may require load reduction.
So the market for HyperCloud is:
– 3 DPC when using 16GB memory modules (high memory loading market – virtualization, CAD)
– 3 DPC, 2 DPC when using 32GB memory modules (significant market)
– 3 DPC, 2 DPC and 1 DPC for DDR4 (all the market)
UPDATE: added 06/06/2012:
The market for LRDIMMs
I have postulated that there is no market for LRDIMMs.
However marketing and lack of information may allow LRDIMMs to be sold.
– LRDIMMs have high latency issues (asymmetrical lines and centralized buffer chipset)
– LRDIMMs fail to deliver 1333MHz at 3 DPC (on non-tweaked Romley servers when compared against HyperCloud)
– 16GB LRDIMMs are non-viable vs. 16GB RDIMMs 2-rank (cannot deliver faster-than-RDIMM speed at 3 DPC)
– 32GB LRDIMMs outperform 32GB RDIMMs (which will only be available in 4-rank for the foreseeable future) – but then 32GB LRDIMMs cannot outperform 32GB HyperCloud
– Inphi has recently been unable to overturn Netlist IP in load reduction and rank multiplication at the USPTO (Netlist ‘537 and ‘274 patent reexams). Netlist vs. Inphi is stayed pending reexams – patents which survive reexam cannot be challenged by Inphi in court.
– LRDIMMs buffer chipsets are sole-sourced by Inphi – if Inphi is prevented from making LRDIMM buffer chipsets then LRDIMMs may face recall as infringing product, or it may become impossible to replace faulty LRDIMMs with new LRDIMMs (LRDIMMs require that all populated slots ONLY use LRDIMMs)
Industry estimates of attach rates
For industry estimates of attach rates for load reduction:
Market opportunity for load reduction
June 6, 2012