[FoRK] advances in spintronics

Eugen Leitl eugen at leitl.org
Fri Nov 23 08:31:14 PST 2012


On Fri, Nov 23, 2012 at 08:15:59AM -0800, Gregory Alan Bolcer wrote:
>
>
> On 11/23/2012 8:06 AM, Eugen Leitl wrote:
>> MRAM is actually great for stacked memories, since
>> it's nonvolatile, and has SRAM speed. Flash is at the
>> end of scaling, so this should eat flash eventually.
>
> I don't know if I agree with that.  In a world where they are creating  
> 1Terabyte volatile RAM per slot with high end applications that use  

This is not slot, but stacked memories. Think of kBit wide
bus, to an SRAM-speed nonvolatile memory on top of a 2-3 W die
which can eat >200 GByte/s memory bandwidth for breakfast.
You can currently get thas with e.g. GDDR5 for GPU, and 
this is large, expensive and runs hot. Instead, you want to feed 
each die, or even a whole wafer. No way around this other 
than stacking.

> memcache and run the full software stack in memory, the size roadmap  

Sequential machines are at the end of the line. Your only
model is a sea of nodes on a mesh, and you need to be able
to feed these nodes.

The future looks something very like this
https://s3.amazonaws.com/ksr/projects/287173/posts/325902/image-168911-full.jpg

Not 32 kBytes/core is tight, but you can work well
with 128 kBytes/core, if you know how. But most code
(and OS) will fall flat on its face when it has
to deal with 16 kCore and 1 MByte/core embedded
memory.

I'm not sure many software developers are yet aware that
the future is different from the past. If they don't
adapt, their code stops getting faster, in fact, will
get slower on that kind of hardware.

> becomes very limiting.

Exactly, conventional systems are very limited.

>> You'll notice they put it on an LSI RAID adapter, where it
>> saves a bufering battery.
>
> Ok.  BTW a complete aside, my stupid battery on my LSI board has been  
> charging for 3 days and still isn't charged.  It's supposed to take a  
> half an hour.  I'm assuming it popped out during shipping or else is  
> missing altogether, but I haven't had time to recrack open the box.  
> Stupid LSI. :-)

Some of the boards come with supercaps -- in fact, even some
better SSDs come with supercaps built-in.

>> It doesn't flip, as MRAM is rad-proof.
>
> That's good.
>
>>
>>> refreshes?
>>
>> The real interesting thing is that a slight MRAM cell
>> extension can do basic logic. So that one will eat FPGA eventually.
>
> Good also.

Current nonvolatile FPGA loads SRAM from flash at boot.


More information about the FoRK mailing list