Intel Readies Optane DIMM Roll-out for 2018

[btarunr]

Intel has reportedly slated launch of its Optane DIMM for the second half of 2018. The Optane DIMM marks the biggest change in computer memory in over two decades, and heralds the era of "persistent memory," which combines the best characteristics of DRAM and NAND flash, in that it has the speed and low-latency of DRAM, but the persistence (ability to store data in the absence of power) of NAND flash. Combining the two will be made possible with improvements to the speed and latency of 3D XPoint memory. Intel is currently selling consumer SSDs based on the technology, and has increased production of 3D XPoint chips.

Intel presented the Optane DIMM at the 2017 USB Global Technology Conference. It described Optane DIMM as a primary storage device that will function as a memory-mapped device, but with much higher storage densities than what's possible with current DDR4 DRAM. The enterprise segment, as usual, will have the first take of the technology, with Intel targeting the exascale computing (supercomputers nearing ExaFLOP/s compute throughput) industry, trickling down to other enterprise segments, before finally making its way to the client/consumer segments. This development is also a polite nudge to the DRAM industry to get its act together, and either bring down prices or scale up densities, or miss the bus of change.



View at TechPowerUp Main Site

 

[_Flare]

a strong APU paired with 256GB X-Point DIMM is all my Gaming-Rig needs
... looking forward

 

[xkm1948]

What about backward compatibility? Would I have to use some expensive next gen HEDT motherboard to support this?

 

[cadaveca]

What about backward compatibility? Would I have to use some expensive next gen HEDT motherboard to support this?

That's a really good question. The details are not exactly clear to even me how this is all going to work, but I do know that HEDT platforms right now are lagging behind mainstream platform for drive performance, and to me, that is unacceptable.

 

[natr0n]

Gonna need a lot of dimm slots.

 

[cadaveca]

Gonna need a lot of dimm slots.

Consider X299 with 8 memory slots, and 16 GB DIMMs (which should soon change to 32 GB per DIMM, if not more when DDR5 comes)...

4 for memory @ 128 GB, and 4 for Optane, @ 2TB (8TB total)


with all of it running @ 100000 MB/s.... and with 64-128 CPU cores @ 3 - 4 GHz





We are on the cusp of a huge change in PC performance, IMHO. I don't know that most enthusiasts are really ready for it.

 

[natr0n]

Consider X299 with 8 memory slots, and 16 GB DIMMs (which should soon change to 32 GB per DIMM, if not more when DDR5 comes)...

4 for memory @ 128 GB, and 4 for Optane, @ 2TB (8TB total)


with all of it running @ 100000 MB/s.... and with 64-128 CPU cores @ 3 - 4 GHz





We are on the cusp of a huge change in PC performance, IMHO. I don't know that most enthusiasts are really ready for it.

Yes true, but with memory prices as they are now and predicted to get higher our wallets wont be ready for it.

 

[cadaveca]

Yes true, but with memory prices as they are now and predicted to get higher our wallets wont be ready for it.

True, but...


You have to get out of today, and look far into the future. Money and costs are the least of my concern about this.

 

[StrayKAT]

I was considering a new build soon, but I might as well wait for all of this to roll out. Especially PCIe-4, while I'm at it.

 

[xkm1948]

Consider X299 with 8 memory slots, and 16 GB DIMMs (which should soon change to 32 GB per DIMM, if not more when DDR5 comes)...

4 for memory @ 128 GB, and 4 for Optane, @ 2TB (8TB total)


with all of it running @ 100000 MB/s.... and with 64-128 CPU cores @ 3 - 4 GHz





We are on the cusp of a huge change in PC performance, IMHO. I don't know that most enthusiasts are really ready for it.

 

[First Strike]

But the latency and the bandwidth are problems. XPoint is faster than NAND flash, but still quite a long way from DRAM. So, for bandwidth hungry tasks such as graphics, no, not suitable.

It will benefit server applications and some HPC significantly, enabling what was previously impossible. However for mainstream consumers, all it can do is to offer better $/GB with worse memory performance (anyway, memory performance isn’t that important these days).

 

[RejZoR]

So, my prediction of unified operating and storage space is becoming a reality. 256GB Optane would give you 256GB of RAM and storage space. Since it's all the same storage space, there should be a lot less bottleneck which we now have due to shuffling of data from storage memory devices to operating memory.

Just don't ecxpect these to have same throughput as RAM. Similar or close yes, same, no. So, I'm assuming this won't be ideal for all tasks. Storage heavy operations like servers will likely greatly benefit from it. And compact devices like laptops.

 

[SteveS45]

Consider X299 with 8 memory slots, and 16 GB DIMMs (which should soon change to 32 GB per DIMM, if not more when DDR5 comes)...

4 for memory @ 128 GB, and 4 for Optane, @ 2TB (8TB total)


with all of it running @ 100000 MB/s.... and with 64-128 CPU cores @ 3 - 4 GHz





We are on the cusp of a huge change in PC performance, IMHO. I don't know that most enthusiasts are really ready for it.

I think the ultimate goal is to eliminate the need of RAM as a fast buffer/volatile storage. If we had non-volatile storage as fast as RAM everything could be executed directly from disk.

 

[londiste]

So, my prediction of unified operating and storage space is becoming a reality. 256GB Optane would give you 256GB of RAM and storage space. Since it's all the same storage space, there should be a lot less bottleneck which we now have due to shuffling of data from storage memory devices to operating memory.

Just don't ecxpect these to have same throughput as RAM. Similar or close yes, same, no. So, I'm assuming this won't be ideal for all tasks. Storage heavy operations like servers will likely greatly benefit from it. And compact devices like laptops.

This is likely to get combined with on-die/package HBM-type RAM for cache when it becomes easily enough available and financially profitable. All-in-all, the limits between operating memory and storage will become more and more blurred.

The main reason for moving XPoint to DIMM is bandwidth-latency though. In its current SSD applications, it does seem to be constrained by both controller and interface, both of which can be considerably improved on within the current standards and paradigm by moving to DIMM.

There will still be shuffling data in and out of different levels of memory. That will not change and depending on how it will get implemented on all hardware-firmware-software, it might actually get more data being shuffled around if the end result justifies it.

 

[cadaveca]

I think the ultimate goal is to eliminate the need of RAM as a fast buffer/volatile storage. If we had non-volatile storage as fast as RAM everything could be executed directly from disk.

Optane isn't a replacement.. it's an add-on. You are headed in the right direction, but that's TOO far into the future.....

 

[jabbadap]

Consider X299 with 8 memory slots, and 16 GB DIMMs (which should soon change to 32 GB per DIMM, if not more when DDR5 comes)...

4 for memory @ 128 GB, and 4 for Optane, @ 2TB (8TB total)


with all of it running @ 100000 MB/s.... and with 64-128 CPU cores @ 3 - 4 GHz





We are on the cusp of a huge change in PC performance, IMHO. I don't know that most enthusiasts are really ready for it.

So dual channel for memory and dual channel for optane nvdimms. On server side that would be even more, Skylake X cpus have six memory channels(i.e. quad memory+dual optane).

 

[Hardware Geek]

That's a really good question. The details are not exactly clear to even me how this is all going to work, but I do know that HEDT platforms right now are lagging behind mainstream platform for drive performance, and to me, that is unacceptable.

It's my understanding that most new technologies come to HEDT first and find their way to typical consumer boards when prices come down enough to justify supporting it.

 

[StrayKAT]

I think the ultimate goal is to eliminate the need of RAM as a fast buffer/volatile storage. If we had non-volatile storage as fast as RAM everything could be executed directly from disk.

I don't know about eliminating the need for RAM, but I'd be happy with it as better caching than what we have now. I'd want it to be faster and as intelligent as SRT.. but more than the 64gb limit (or 32 with optane M.2). I imagine it'd cost an arm and a leg though.

 

[Basard]

I like how they write in their slide that RAM is more expensive than NAND.... Yeah, per GB maybe.

But how much is an Optane DIMM vs RAM+SSD?

 

[Vya Domus]

It's my understanding that most new technologies come to HEDT first and find their way to typical consumer boards when prices come down enough to justify supporting it.

Not really.

 

[cadaveca]

It's my understanding that most new technologies come to HEDT first and find their way to typical consumer boards when prices come down enough to justify supporting it.

With Intel's HEDT, this is most definitely not the case, and the CPU core design lags a full generation behind the mainstream desktop parts.

So dual channel for memory and dual channel for optane nvdimms. On server side that would be even more, Skylake X cpus have six memory channels(i.e. quad memory+dual optane).

I actually think that this will only support a single Optane device, not multiple (as current Optane drives work this way). You might also note that the projected launch is 2nd half next year, which could mean this time next year for all we know, so details are going to be a bit lean on what exactly is going on for quite some time yet.​

 

[levish]

Consider X299 with 8 memory slots, and 16 GB DIMMs (which should soon change to 32 GB per DIMM, if not more when DDR5 comes)...

4 for memory @ 128 GB, and 4 for Optane, @ 2TB (8TB total)


with all of it running @ 100000 MB/s.... and with 64-128 CPU cores @ 3 - 4 GHz





We are on the cusp of a huge change in PC performance, IMHO. I don't know that most enthusiasts are really ready for it.

Could be that Real-Time ray tracing is very near

 

[cadaveca]

Could be that Real-Time ray tracing is very near

HDR and VR plus other technologies are pushing bandwidth requirements for all devices, yes. We're still a fair way away from that though, IMHO, given the huge increase in pixels that 4K requires compared to 1080P, never mind 8K or 10K. Finding suitable solutions for true "realtime" rendering at high resolutions is a neat idea though. Augmented Reality stuff really needs a huge boost in visual quality, and today, we're still kind of stuck with what to me, seems like NES-equivalents (I mean the original Nintendo ).

 

[levish]

There's also Voxel based software that is typically CPU/Memory intensive. Exciting stuff!

 

[geon2k2]

So, my prediction of unified operating and storage space is becoming a reality. 256GB Optane would give you 256GB of RAM and storage space. Since it's all the same storage space, there should be a lot less bottleneck which we now have due to shuffling of data from storage memory devices to operating memory.

Just don't ecxpect these to have same throughput as RAM. Similar or close yes, same, no. So, I'm assuming this won't be ideal for all tasks. Storage heavy operations like servers will likely greatly benefit from it. And compact devices like laptops.

Have fun rebooting if an application hangs