Blueshift's memory module to be 1000 times faster than conventional memory

[lihas]

theiet.org/new-form-of-computer-memory-is-1-000-times-faster/

According to this article, this new startup called "blueshift memory" have developed a new kind of memory module which could be upto 1000 times faster than the ones we use today.

They give analogy with a city where everything - shops, houses, etc. are constantly changing. This refers to data being moved around in a conventional memory module.

There memory module, using the same analogy , in some ways gives each place a sort of permanent address.

The thing that is mind boggling is how can fixing placement of data in RAM speed up memory?

At best it can improve processor's cache performance due to spatial locality of data.

I am not able to think anything beyond.

There is nothing given about the internal workings of the memory.

I want to know what's happening inside; even a discussion on rough possibilities will do.

It feels like showing a bone to a dog, but then never really giving it to him.

 

Their website - https://www.blueshiftmemory.com/

 

[Bouzoo]

21 minutes ago, lihas said:

new startup called "blueshift memory" have developed a new kind of memory module which could be upto 1000 times faster than the ones we use today.

We've heard so many promises from startups coming forth with such allegedly revolutionary products that I am expecting us to be able to build lightsabers in next few years.

Seriously, if it sounds too good to be true it probably is.

[samcool55]

Yeeeaa I dunno but this sounds a bit weird to me.

right now RAM is fast enough, especially for servers.

If we do run into limitations it is usually that the RAM is just too far away from the CPU (basically the trace lay-out limits speed)

And 1000 times faster, compared to what and what exactly is faster? Are they talking about latency or just raw speed?

There is no point in insane fast memory if it takes like a second to respond or memory with very low latency but is slow as balls.

[williamcll]

Cool but is it compatible with mainstream processors? 

[Dabombinable]

1000x faster than EDO?

[porina]

3 hours ago, samcool55 said:

right now RAM is fast enough, especially for servers.

Ram is a major bottleneck and the gap is growing bigger, not smaller, since we started getting more cores but not faster ram. Look at server systems, you have 6 or 8 ram channels per socket. Tricks like large L3 caches are only a workaround until we get fast enough ram to keep up with the cores.

[The King of the Undead]

for some reason I read the companys name as bull$h!t.

So Will ddr4 be cheap now?

[Anomnomnomaly]

Back in the mid to late 90's I read a really, really exciting article about how CPU's would soon be built using crystals and light to process data.

 

That was around 96-97.... it's 2019... Where are my crystal CPU's dammit.

[CarlBar]

Yeah this is either major overhype or major bull.

 

1 hour ago, porina said:

Ram is a major bottleneck and the gap is growing bigger, not smaller, since we started getting more cores but not faster ram. Look at server systems, you have 6 or 8 ram channels per socket. Tricks like large L3 caches are only a workaround until we get fast enough ram to keep up with the cores.

 

Like you said lots of ram channels. If that wasn't working adequately in most cases we would have seen more substantial changes.

 

 

Where entirely capable of building higher bandwidth interconnects, but doing so would mean new tech and thats costly enough it's not a serious proposition.

[WereCat]

Hmmmm

 

 

it better not cause a resonance cascade while going so fast

[Franck]

1 hour ago, porina said:

Ram is a major bottleneck and the gap is growing bigger, not smaller, since we started getting more cores but not faster ram. Look at server systems, you have 6 or 8 ram channels per socket. Tricks like large L3 caches are only a workaround until we get fast enough ram to keep up with the cores.

I totally agree with that. And my database servers that take advantage of RAM tables are with me on this.

[samcool55]

3 hours ago, porina said:

Ram is a major bottleneck and the gap is growing bigger, not smaller, since we started getting more cores but not faster ram. Look at server systems, you have 6 or 8 ram channels per socket. Tricks like large L3 caches are only a workaround until we get fast enough ram to keep up with the cores.

Tbh that is not the fault of ram but of the CPU's not supporting fast ram (and for that reason no fast ram with ECC exists). We have ram that can easily do 3000+Mhz, even the 4000Mhz barrier can be smashed by some ram kits, but that's way out of spec for most server cpu's.

 

Taking EPYC as an example, that maxes out at 2666Mhz, that is really slow compared to what we get on the desktop if we OC it. It would help a lot if server CPU's would support faster ram because it exists and it does help (zen for example performs a lot better with faster ram) but they don't sooo yea...

 

I'm not saying faster ram magically solves everything, but with the current tech we have it seems possible we can make some server stuff that's not as badly bottlenecked by ram as most servers are right now.

[Doobeedoo]

It's always very interesting to see memory and storage advances, even being far in the future. Maybe one day eventually everything being like light based storage and chips. 

[porina]

36 minutes ago, samcool55 said:

Tbh that is not the fault of ram but of the CPU's not supporting fast ram (and for that reason no fast ram with ECC exists). We have ram that can easily do 3000+Mhz, even the 4000Mhz barrier can be smashed by some ram kits, but that's way out of spec for most server cpu's.

There's not a fault to be applied but merely the reality we're not in a good balance right now and it is getting worse, not better. Servers buy in quantity and they care about reliability. The fast ram kits you can get are relatively exotic, specially binned and hardly mass produced. Compatibility is also far from 100%. DDR5 will eventually help. For now we're left with trying to cram in more channels or using caches to relieve the pressure. I'm almost surprised we haven't gone back to on package L4 caches yet. A local high speed interconnect could make that very interesting.

[Meandme]

Background:

https://www.techopedia.com/definition/14630/von-neumann-bottleneck

 

Trends:

http://blog.appliedmaterials.com/tackling-big-questions-memory

 

With usage of faster memory you could integrate more cores into CPUs/GPUs.

 

Blueshift Memory has been created to boost special Big Data problems.

Some O(n) problems are running on O(logn) or even O(c).

Test videos from last year:

 

 

 

 

 

[Beskamir]

9 hours ago, Bouzoo said:

We've heard so many promises from startups coming forth with such allegedly revolutionary products that I am expecting us to be able to build lightsabers in next few years.

Seriously, if it sounds too good to be true it probably is.

Yea, if they want people to take them seriously they need to show their shit works and give it to neutral parties (reviewers) to verify.

[Froody129]

If you're going to try advertise your startup... at least do it properly