Intel planning to get rid of DRAM on PCs (semiaccurate.com)
Initially I didn't like the idea. But thinking it better it might be actually good. The big PC bottleneck nowadays is DRAM. A big on-die/stacked memory at many times faster bandwidth and lower latency could be a game changer.
And anyway, nowadays the usual advice to clients is to max out the RAM on purchase since it's a mess to replace (be it a non-technical user or a datacenter cluster with critical uptime) and the specific configuration of RAM will become pricer per GB in the mid term (because the market moves constantly forward to different versions).
And also note one of the main reasons some algorithms run faster on GPUs is the memory bandwidth bottleneck. If Intel gets rid of the BW problem it's a level game againsta Nvidia.
This move would allow Intel to have some custom performance tricks. For example, memset multiple GB in a fraction of the time.
24 comments
[ 3.0 ms ] story [ 60.6 ms ] threadAnd anyway, nowadays the usual advice to clients is to max out the RAM on purchase since it's a mess to replace (be it a non-technical user or a datacenter cluster with critical uptime) and the specific configuration of RAM will become pricer per GB in the mid term (because the market moves constantly forward to different versions).
And also note one of the main reasons some algorithms run faster on GPUs is the memory bandwidth bottleneck. If Intel gets rid of the BW problem it's a level game againsta Nvidia.
This move would allow Intel to have some custom performance tricks. For example, memset multiple GB in a fraction of the time.
EDIT: from the comments I get the impression that they want to move the memory closer to the CPU.
The description of "getting rid of DRAM" sounds misleading if this is what they are deploying.
In other HN curiosities, I lost the ability to flag a few weeks ago.
Does anyone know more? Are they moving the RAM or replacing with some flash derivative? Something else?
More details in this LWN article:
http://lwn.net/Articles/498283/ and here: http://en.wikipedia.org/wiki/Phase-change_memory
http://blogs.intel.com/research/2011/09/15/hmc/
1 point by alecco 2 hours ago | link | parent [dead]
Initially I didn't like the idea. But thinking it better it might be actually good. The big PC bottleneck nowadays is DRAM. A big on-die/stacked memory at many times faster bandwidth and lower latency could be a game changer. And anyway, nowadays the usual advice to clients is to max out the RAM on purchase since it's a mess to replace (be it a non-technical user or a datacenter cluster with critical uptime) and the specific configuration of RAM will become pricer per GB in the mid term (because the market moves constantly forward to different versions).
And also note one of the main reasons some algorithms run faster on GPUs is the memory bandwidth bottleneck. If Intel gets rid of the BW problem it's a level game againsta Nvidia.
This move would allow Intel to have some custom performance tricks. For example, memset multiple GB in a fraction of the time.
As SSD gets faster every year, we need less memory to cache off the content where we used to need for performance reason from a HDD. With Windows 8 i have been using less then 4GB of Ram including standby, with Virtual Memory turned off. ( I have 16GB of Memory ). The future CPU die could easily stack 4 1Gbit DRAM with 8 layers of up to 4GB of Memory. Coupled with a SSD that goes up to 2GB/s transfer.