Discussion Optane Client product current and future

[cbn]

Gigabyte has four Z370 boards with "built-in"* 32GB Optane:

https://www.techpowerup.com/243513/...h-exclusive-built-in-32gb-intel-optane-memory

https://www.gigabyte.com/MicroSite/496/optane.html

*It is not really built-in, but rather a regular M.2 2280 32GB Optane stick (3 out of 4 models also have a heatspreader on it).

 

[cbn]

Besides TSVs, I wonder how lithography shrinks might affect the progress towards what is described in the following link from August 2016:

https://news.stanford.edu/press/view/9468

Showing that phase-change materials can be transformed from zero to one by a picosecond excitation suggests that this emerging technology could store data many times faster than silicon RAM for tasks that require memory and processors to work together to perform computations.

 

[IntelUser2000]

The Aorus Optane board is quite good for the value. The difference in price is only $50 cdn, which is less than the average going value of the standalone module at $66-80. The US Newegg site is having an extra sale, so the extra cost for the 32GB bundled module is only $20 US.

 

[cbn]

Another article on Phase Change memory:

http://www.sciencemag.org/news/2017...puter-memory-could-herald-next-generation-ram

That’s because materials scientists in China have recently found a way to speed up—by more than a factor of 10—so-called phase-change random access memory (PCRAM), which can hold onto information even when your computer’s power is off.

The problem has been that the most commonly used phase change material—an alloy of germanium, antimony, and tellurium known as GST—can be an uneven performer, sometimes switching from amorphous to crystalline in 10 nanoseconds, which is just as fast as the DRAM in today’s computers. At other times, it can take hundreds of nanoseconds to write data bits, which is too slow to be competitive. Now, however, Feng Rao, a materials scientist at the Chinese Academy of Sciences’s Shanghai Institute of Microsystem and Information Technology, and colleagues have found a way to write all PCRAM bits quickly, making it faster than most alternatives, including NAND flash—one of the other kinds of memory that’s able to store information absent a power supply.

So the breakthrough (a material science one) involves going from 100's of nanoseconds to 10's of nanoseconds for writes in the worst case scenario. (This compared to a best write speed of 10ns)

However, the article goes to explain the tech still needs to prove itself under real world conditions and manufacturing:

However, Jeong adds, the faster PCRAM must still prove that it can be scaled up, withstand the high temperatures found in standard chip-manufacturing conditions, and still be able to rewrite bits of data many trillions of times to match DRAM’s performance.

(So PCM potentially replacing DRAM)

P.S. Intel does not claim 3DXPoint is phase change memory, but it has been observed by others to be phase change memory:

https://www.pcper.com/reviews/Editorial/How-3D-XPoint-Phase-Change-Memory-Works

https://www.theregister.co.uk/2018/04/26/chinese_scientists_claim_xpoint_is_phase_change_memory/

(So although we can track specs like # of layers, lithography and TSVs, one thing Intel may never dislose is material science or material science changes)

From the April 2018 "theregister.co.uk" link:

"Intel is not disclosing the materials used for 3D XPoint memory media, the underlying memory media used in Intel Optane SSDs and memory, or the manufacturing processes used."

 

[Keljian]

I did a thing:
https://hardforum.com/threads/intel-900p-optane-wow.1958026/page-2#post-1043611550

Tested (on a Ryzen 1700):
Optane 32 gb caching a Samsung 850 Evo and a 950 pro using primocache with 4K block size.

In a nutshell, faster in both scenarios, system is definitely more responsive. Note boot drive is the 950pro, which has a heatsink on it.

 

[cbn]

For Optane memory, I wonder what happens if before Optane memory software set-up the user manually assigns the page file to the Optane device? Does the Optane software use the rest of the drive for cache? Or does the Optane software use the entire drive (thus eliminating the manual set page file)?

I am particularly interested in using this for the 16GB Optane as it is only a block level cache through the Intel Optane memory software.

 

[nosirrahx]

For Optane memory, I wonder what happens if before Optane memory software set-up the user manually assigns the page file to the Optane device? Does the Optane software use the rest of the drive for cache? Or does the Optane software use the entire drive (thus eliminating the manual set page file)?

I am particularly interested in using this for the 16GB Optane as it is only a block level cache through the Intel Optane memory software.

Optane activation blanks the drive no matter what is on it.

 

[Dayman1225]

Intel Persistent Memory is sampling today and revenue is 2018

Announcing for developers, remote access to systems with Optane DC Persistent Memory through Intle Builders Consutrcution Zone

 

[Dayman1225]

Anandtech Persistent Memory Live Thread

 

[cbn]

128GB capacity....(I assume this one has 1 die per package)

.......would like for client.

 

[cbn]

https://www.anandtech.com/show/12828/intel-launches-optane-dimms-up-to-512gb-apache-pass-is-here

Intel is not officially disclosing whether it will be possible to mix and match DRAM and Optane Persistent Memory on the same memory controller channel, but the 192GB DRAM capacity for the development preview systems indicates that they are equipped with a 16GB DRAM DIMM on every memory channel.

The preview systems will feature 192GB of DRAM and 1TB of Optane Persistent Memory, plus SATA and NVMe SSDs.

The preview systems are 2P with the first twelve slots having 16GB DRAM each with the second twelve slots having some some slots fitted with Optane but other slots empty?

As a reference point the Lenovo 2P system had 16 slots. (With the 12 slots filled with DRAM DIMMs and four slots filled with Optane)


https://www.tomshardware.com/news/intel-xd-xpoint-dimm-lenovo-thinksystem,36573.html

 

[cbn]

If that memory scheme works for server, then a dual channel consumer system (with four DIMM slots) could have 2 DRAM DIMMs and one Optane DIMM while maintaining dual channel at the full speed of the DRAM? (This via software (or OS) handling the Optane DIMM differently than it would if it saw a regular DRAM DIMM in that same 3rd slot).

 

[cbn]

Some very old information I found on a widening CPU to Memory gap:

http://slideplayer.com/slide/4926188/

https://www.ece.ucsb.edu/~strukov/ece154bSpring2013/week2.pdf

If this trend still holds true today then I'm guessing this is one very strong reason why Intel is so interested in Phase change memory and its potential.

Fast, big, persistent memory that would open up new design possibilities in future CPUs. (That would be great, but then it further increases the need for fast storage*)

*QLC NAND (increased layers, string stacking, smaller lithography) vs. Multi-actuator (including helium) hard drives on various levels.

 

[IntelUser2000]

Optane SSD 905P to be released in an M.2 form factor. Some articles say summer release. Performance rated identical to 905P.

 

[Brahmzy]

^^ Whoa. Is it a std length or that super long-ass one that no boards support??

 

[IntelUser2000]

It's M2 22110. Yes, its little larger than some boards can support. You should be fine if you have an ATX sized board. My H270 supports up to 110mm length.

 

[IntelUser2000]

Tomshardware 905P review: https://www.tomshardware.com/reviews/intel-optane-ssd-905p,5600-2.html

The controller used on the 905P seems far better at extracting multi-queue performance than the one on the 900P. The performance on random read and write QD4 are whopping 40% higher than the 900P! That seems to help it additionally in load times over the 900P(of course its small). Some other reviews bear this out.

 

[cbn]

Tomshardware 905P review: https://www.tomshardware.com/reviews/intel-optane-ssd-905p,5600-2.html

The controller used on the 905P seems far better at extracting multi-queue performance than the one on the 900P. The performance on random read and write QD1 are whopping 40% higher than the 900P! That seems to help it additionally in load times over the 900P(of course its small). Some other reviews bear this out.

The Sequential Read is also higher (by about 80 MHz) at QD1 and maybe 200 MHz at QD2 and greater.

P.S. IntelUser2000 QD1 (for random read and write) was a typo right? Random performance boost doesn't happen till a bit after QD2 and maxes out once reaching QD4:

(Load times affected more by Sequential Read improvement or improvement in 4K QD2+?)

 

[IntelUser2000]

You are right. I changed it. I meant QD4.

I don't think load times are as simple as one metric. Also, benchmarks may not reflect everything that happens in a complex application such as a game. The 905P is probably more balanced in that regard.

 

[Brahmzy]

Yeah, for 90% workstation workloads the 900p and 905p are identical.

I’m excited about a performance-equivalent M.2 version, with hopefully a price drop from my U2/M2 280GBers.

 

[nosirrahx]

If that memory scheme works for server, then a dual channel consumer system (with four DIMM slots) could have 2 DRAM DIMMs and one Optane DIMM while maintaining dual channel at the full speed of the DRAM? (This via software (or OS) handling the Optane DIMM differently than it would if it saw a regular DRAM DIMM in that same 3rd slot).

BIOS and chipset limitations will likely be in the way of consumer boards having this functionality. It would be nice but Intel would need to bet on consumers being willing to buy expensive boards and even more expensive Optane DIMMS.

 

[IntelUser2000]

I’m excited about a performance-equivalent M.2 version, with hopefully a price drop from my U2/M2 280GBers.

I think the gains may be partially because its firmware and revision to the controller are more consumer oriented. Don't count on the price decrease. Performance has been going up, but so is the price. Their NVMe chief made their strategy clear: Optane for performance, and 3D NAND for capacity.

It would be nice but Intel would need to bet on consumers being willing to buy expensive boards and even more expensive Optane DIMMS.

Based on them putting the Xeon Platinum into a single socket board and calling it a consumer chip, I wouldn't be surprised if Cascade Lake-X supports a cheaper version of the Optane DIMMs. I think charging nearly $2/GB for the 800P is ridiculous, but even $3/GB for a 128GB Optane DIMM would make sense.

P.S. My guess for the development system is 2x512GB

 

[cbn]

(Thinking maybe 2 years ahead for client, but hopefully sooner)

If dual channel DDR4 is good enough for a 95W 6C/12T Coffee Lake processor then wouldn't DDR interface Optane (even with it being slower) be good enough for a low power x86 processor?

Maybe 2GB DRAM (single channel, soldered-on) +16GB DDR interface Optane replacing 8GB DRAM on certain low wattage mobile systems?

At what point does Bottleneck occur in compute tasks? (Thinking CPU and iGPU tasks)

(I am optimistic because I found in the past that even 2GB dual channel DDR2 800 did not hold me up in Linux Mint (outside of a internet browser) when using Xeon X3323 (equivalent to Core 2 Quad Q9300) and GT 630 GDDR3 2GB. Internet Browsing is another story because it uses lots of RAM while not needing much compute.)

 

[cbn]

Thinking about 2GB DRAM + 16GB DDR interface Optane vs. 8GB DRAM a bit more.....I am interested in how well this could mesh with Intel's Lower Display Technology:

https://liliputing.com/2018/06/intels-low-power-display-tech-could-extend-battery-life-by-hours.html

If everything works as expected, users wouldn’t notice any difference. For example, you might want your screen to refresh 60 times a second when watching videos or playing games. But a lower rate is fine when you’re looking at static images. So your device’s GPU would recognize the activity, send a message to the display, make an adjustment, and… everything would look normal to the user.

So maybe 2GB DRAM + 16GB DDR interface Optane + Intel's Lower Display Technology could be a better than normal match when the PC is used mostly used for browsing? (This if 2GB DRAM + 16GB DDR Optane uses less power than 8GB DRAM)

 

[IntelUser2000]

Maybe 2GB DRAM (single channel, soldered-on) +16GB DDR interface Optane replacing 8GB DRAM on certain low wattage mobile systems?

Do you have a certain usage scenario in mind, or are you looking for a general low end system? NV memory like Optane is first and foremost aimed at performance. Businesses(like Intel) will be selling like so. What you are describing, Optane being used in low end systems are likely a decade away, when its proliferated into most of the high end markets.

It's interesting to think, but its an extremely niche way of taking advantage of a new technology. It's like if Apple's first iPhone wasn't a $600 one with all the features, but a $30 one meant for developing markets. Now, its a different story because billions of Smartphones are out there.