SATA Express 3.2 16 Gbps

[Chimel]

Any news at all on this new class of drives, motherboards or chipsets?
I couldn't find any announcement of upcoming or future devices on the web besides the unfinished Asus motherboard prototype from last year. 3.2 was ratified back in 2013, and from past experience, it takes about that long for such new devices to appear on the market, but you'd expect at least some announcements.

Is something fishy? Like the uncanny SATAe contraption of SATA and PCIe being more difficult to implement according to the SATAe 3.2 specs? Some people already complained about the lack of power in the new SATAe connector, unlike the PCIe specs which require to supply power, so I wonder if they are trying to revise the specs. A 8TB SSD recently released, we'll probably need SATA 3.2 soon at this rate of technological advances.

 

[Oleg34]

Holy shit. That's a lot of storage and it's SSD.

 

[Deleted member 126051]

Holy shit. That's a lot of storage and it's SSD.

Now imagine 12 of them in a NAS box running RAIDZ30

Now imagine the taxes you'd pay buying/assembling such a contraption.

Now imagine the disk controller crapping itself when it spends most of the time at capped throughput.

 

[aaronspink]

Any news at all on this new class of drives, motherboards or chipsets?
I couldn't find any announcement of upcoming or future devices on the web besides the unfinished Asus motherboard prototype from last year. 3.2 was ratified back in 2013, and from past experience, it takes about that long for such new devices to appear on the market, but you'd expect at least some announcements.

Is something fishy? Like the uncanny SATAe contraption of SATA and PCIe being more difficult to implement according to the SATAe 3.2 specs? Some people already complained about the lack of power in the new SATAe connector, unlike the PCIe specs which require to supply power, so I wonder if they are trying to revise the specs. A 8TB SSD recently released, we'll probably need SATA 3.2 soon at this rate of technological advances.

AFAICT, SATA Express is basically dead. Its essentially a spec without a point. If you are in the market for that level of performance you are going to go all the way to NVMe 2.5". It actually has drives available and has 2x the headroom, doesn't require any SAS/SATA interface signals, already leverages a robust cabling infrastructure, and is available for basically every platform under the sun.

SATA Express was an attempt to expand sata to double the speed while taking 4x the signals. And it still ties you to all the old SATA issues. NVMe has significantly lower latencies, actually uses the signals for more data, and is more supported and is the standard going forward. There will be multiple form factors for NVMe, from M.2 to straight PCIe to 2.5" using the 12 Gbps Quad SAS connectors.

And so no one has released a SATA Express drive. And it doesn't appear that anyone has plans to. Can't find a single vendor who has plans, any why should they?

 

[R3MF]

^ he's right ^

 

[Armenius]

Any news at all on this new class of drives, motherboards or chipsets?
I couldn't find any announcement of upcoming or future devices on the web besides the unfinished Asus motherboard prototype from last year. 3.2 was ratified back in 2013, and from past experience, it takes about that long for such new devices to appear on the market, but you'd expect at least some announcements.

Is something fishy? Like the uncanny SATAe contraption of SATA and PCIe being more difficult to implement according to the SATAe 3.2 specs? Some people already complained about the lack of power in the new SATAe connector, unlike the PCIe specs which require to supply power, so I wonder if they are trying to revise the specs. A 8TB SSD recently released, we'll probably need SATA 3.2 soon at this rate of technological advances.

Sunrise Point from Intel (Skylake chipset) will supposedly have SATA Express integrated. I believe there are SSDs from Intel available now that use SATA Express.

 

[R3MF]

hmmmm, i do love the smell of a BTX case in the morning!

 

[Chimel]

AFAICT, SATA Express is basically dead.

Thanks for the information, that's what it looked like to me, hence this post.
Even if we won't see SATAe, I still think we'll see more of SATA, maybe with regular (not Express) versions 3.1 and 3.2. PCIe drives such as NVMe seem more aimed to the enterprise market and still years away, even if we are starting to see some. (And again, just like the 840 Pro, great drives from Samsung, and great NVMe hot-swap solutions from Supermicro, these guys rock!)

 

[aaronspink]

Thanks for the information, that's what it looked like to me, hence this post.
Even if we won't see SATAe, I still think we'll see more of SATA, maybe with regular (not Express) versions 3.1 and 3.2. PCIe drives such as NVMe seem more aimed to the enterprise market and still years away, even if we are starting to see some. (And again, just like the 840 Pro, great drives from Samsung, and great NVMe hot-swap solutions from Supermicro, these guys rock!)

NVMe drives are basically shipping from every major vendor ATM or will be shortly. Intel has a full line, HGST has a line, Samsung has a line, multiple vendors have M.2 and/or NVMe 2.5" drives coming on market this year.

NVMe drives will be flooding the market this year, esp in the consumer space. There are multiple merchant controller chips that are available and will result in a large number of drives being built.

 

[Happy Hopping]

I want to understand this, M.2 drive is on a circuit board, when will we get actual drive that iis at 16Gb/sec?

And if SATA Express is dead, what's the next generation? is this it?

http://www.extremetech.com/computin...d-performance-up-to-7x-with-new-optane-family

1 other question, is M.2 safe as a root drive? Can Win 7 easily see your M.2 at your PCI e 3.0 slot as drive C? And if you have 1 x M.2 on your PCI E slot and a regular SSD as drive C, would win 7 has any problem knowing which drive is the root drive?

 

[ReCOde]

Who in the HELL needs an 8TB SSD? Lol! I wish!

 

[Happy Hopping]

1 more thing, most people only have as much as 6 PCI E slot. Assuming 1 of the slot is your video card, then you really only have 5. So if you want M.2 speed, then all you can get is 5 drives?

w/ 2.5" SSD, you can JBOD them, so does that mean we can't get 16GB/s or faster on 2.5" size box?

 

[SomeGuy133]

1 more thing, most people only have as much as 6 PCI E slot. Assuming 1 of the slot is your video card, then you really only have 5. So if you want M.2 speed, then all you can get is 5 drives?

w/ 2.5" SSD, you can JBOD them, so does that mean we can't get 16GB/s or faster on 2.5" size box?

you do realize that SATA goes through PCH which is super slow right?

Also what your seeing is the major issue we are now facing...PCIe 3.0 sucks balls. People are running out of bandwidth even on 40 lanes...consumer end has 16 -_-

 

[aaronspink]

you do realize that SATA goes through PCH which is super slow right?

Also what your seeing is the major issue we are now facing...PCIe 3.0 sucks balls. People are running out of bandwidth even on 40 lanes...consumer end has 16 -_-

Actually, we generally have more than enough PCIe bandwidth, whats actually lacking is connectivity, which is primarily an issue because Avago bought PLX and then jacked the prices sky high. AKA no one is running out of bandwidth.

For most use cases, even 16 PCIe lanes is more than enough. Even the top end graphics cards don't really need more than ~4-8 lanes or so. If the graphics card runs out of on board memory, its going to need tens to hundred of GB/s of bandwidth to memory in order to not fall over.

For storage, likewise, except for the enterprise space, 4-8 PCIe are more than enough. That's 4-8 GB/s to/from storage which is generally more than is needed. Even the best NVMe SSDs can really use the full bandwidth of U.2 currently except in bulk sequential transfers which aren't the limiting factors. So 8 U.2 drives connected to a x8 PCIe uplink is unlikely to be a limiting factor short of the enterprise space.

The issue as I said is connectivity. Realistically, it should be possible to make 36-48 port PCIe switches pretty much dirt cheap. Unfortunately the cost of the switches on the market has done nothing but go up in price due to lack of competition (PLX was the only game in town for quite some time with PCIe 3.0 switches). PMC-Sierra has now launched their own line of PCIe 3.0 switches, and hopefully IDT will get off their butts and extend their switch line to 3.0 as well. This should give some competition and drive down the costs.

Numbers I've heard for current pricing on the PLX/Avago 32 port switches is >$100 which is pretty horrible considering you used to be able to buy the 48 port switches for <$100.

 

[SomeGuy133]

what are you talking about? A SLI of 980 is capped (a little) by lack of bandwidth in PCIe 3.0 and then you want to throw in PCIe SSD? There is not enough bandwidth for that period. Quad SLI is also bandwidth limited with current gen GPUs It will be horrible with next gen GPUs being twice as face.

PCIe 8x is too slow for current gen so how are you going to run top end cards at 4x if 8x is already showing issues? Connectivity isn't the issue ATM. Bandwidth is.

In the test below a 980, which is much slower then 980TI and won't even touch next gen cards is already showing bandwidth issues at 8x and is having detrimental effects at 4x. Throw a 980TI into this exquation it is even worse. Next gen? HA. So how are you going to run a 980 TI or next gen that is 2x faster in a 4x slot? Remember Skylake had only 16 lanes...not 20 like we hopes so you are going to need to run 1 GPU at 8x and another at 4x and the SSD at 4x. That is a terrible set up especially counting the next gen GPUs coming this summer.

I saved you the trouble and gooded the article. I thought there was a newer one using the 980 TI or Titan X but I can't find it but the 980 isn't even that fast and at 4x it is showing major bottle necks in some of the newest games.

https://www.techpowerup.com/reviews/NVIDIA/GTX_980_PCI-Express_Scaling/19.html

So even high end systems with 40 lanes can run into bandwidth issues too depending on how much crap they want to run on PCIe.

 

[aaronspink]

what are you talking about? A SLI of 980 is capped (a little) by lack of bandwidth in PCIe 3.0 and then you want to throw in PCIe SSD? There is not enough bandwidth for that period. Quad SLI is also bandwidth limited with current gen GPUs It will be horrible with next gen GPUs being twice as face.

PCIe 8x is too slow for current gen so how are you going to run top end cards at 4x if 8x is already showing issues? Connectivity isn't the issue ATM. Bandwidth is.

In the test below a 980, which is much slower then 980TI and won't even touch next gen cards is already showing bandwidth issues at 8x and is having detrimental effects at 4x. Throw a 980TI into this exquation it is even worse. Next gen? HA. So how are you going to run a 980 TI or next gen that is 2x faster in a 4x slot? Remember Skylake had only 16 lanes...not 20 like we hopes so you are going to need to run 1 GPU at 8x and another at 4x and the SSD at 4x. That is a terrible set up especially counting the next gen GPUs coming this summer.

I saved you the trouble and gooded the article. I thought there was a newer one using the 980 TI or Titan X but I can't find it but the 980 isn't even that fast and at 4x it is showing major bottle necks in some of the newest games.

https://www.techpowerup.com/reviews/NVIDIA/GTX_980_PCI-Express_Scaling/19.html

So even high end systems with 40 lanes can run into bandwidth issues too depending on how much crap they want to run on PCIe.

Yeah so that article you wanted to point out: PCIe x4 3.0 ranges from 94% to 99% of the performance of PCIe x16 3.0. Doesn't really seem to be a bottleneck to me.

And quad SLI? WTF cares. Seriously. Its a setup like .000000001% of the market will ever have, and those people are idiots.

And as I said, the issue is a connectivity issue,. not a bandwidth issue. For high end dual GPU MBs, we should be using PCIe switches but they've priced themselves out of the market due to a lack of competition and greed.

 

[SomeGuy133]

12% do some math. 31/35fps....

Plus this is with a 980 not a 980 TI and next gen is going to also be 3.0 so you will have cards 3 times faster then a 980. So yes there is a serious bottleneck

16 lanes for a dual GPU plus SSD? 8x 4x 4x. That will be shit. Plus the high end SSDs use 8x so the SSD will also be bottlenecked....so yes there is a serious bottleneck issue especially if someone needs single threads like myself. I want a skylake build for single thread performance but if I do that I will have serious PCIe bottleneck issue if I want dual next gen GPUs for 4k and VR with XPoint. Please tell me how 2 next gen GPUs and XPoint won't bottleneck with only 16 lanes of 3.0

Please explain the magical math that shows 3.0 has plenty of bandwidth.....I am waiting.


EDIT: If intel had skylake medium core CPUs out when regular ones were out I wouldn't have an issue but the medium core CPUs are always a gen behind....why for the love of god do medium cores always lag.

I would have built 2 6 core skylake xeons if they were out but instead I am stuck with a 1650v3 and a 6700k (waiting for parts to build it)


On a side note, I am very wary about these switches because of latency. I heard they add a significant amount of latency but I have not heard or see any reports showing one or the other. I would not touch those switches if they added any latency because that will murder XPoint and I be damned to nerf that.

I want XPoint in RAM but supposedly those are not coming out first :/ and require special chipset support, which is odd considering ULTRADIMMS don't have that issue. I would actually buy that if it was available but sadly it is still blocked from sale. (or at least last I heard)

 

[aaronspink]

12% do some math. 31/35fps....

Plus this is with a 980 not a 980 TI and next gen is going to also be 3.0 so you will have cards 3 times faster then a 980. So yes there is a serious bottleneck

No its not cause you are going to have a card with 50-200% more memory on it. Basically the only time the PCIe bus gets used is when the VRAM is filled or to send cmd lists (and cmd lists aren't taking any real bandwidth). In fact, I would bet that a 980TI would be LESS dependent on pcie performance right now than a 980.

And no my percents were correct. You want to focus on one game on a dead end engine which even the architect of the engine has admitted is a dead end and specifically stresses bus bandwidth due to poor architectural choices. Its not really representative at all.

16 lanes for a dual GPU plus SSD? 8x 4x 4x. That will be shit. Plus the high end SSDs use 8x so the SSD will also be bottlenecked....so yes there is a serious bottleneck issue especially if someone needs single threads like myself. I want a skylake build for single thread performance but if I do that I will have serious PCIe bottleneck issue if I want dual next gen GPUs for 4k and VR with XPoint. Please tell me how 2 next gen GPUs and XPoint won't bottleneck with only 16 lanes of 3.0

8-12 lanes into a 24-32 port switch supporting 2 PCIe x8 slots. 4-8 lanes into a 20-24 port switch supporting 4-5 U.2/M.2 slots. More than enough bandwidth for either graphics or storage.

And no the high end SSDs don't use x8, they use x4. U.2 is the high end. Its what everyone is designing towards and has product in the market for. And XPoint? LOL, I don't think you even understand the market it is going into. No one from either product side is giving any though to -S dies and Xpoint. Besides, the only way you could potentially saturate x4 or x8 via NVMe storage is with moderate to deep queue depth sequential accesses, which isn't really useful. For real work, none of the drive announced or coming are pushing more than 1GB/s aka 1 PCIe lane.

As far as nextgen GPUs go, get a board with an Avago/PLX or PMC-Sierra switch chip. Done.

EDIT: If intel had skylake medium core CPUs out when regular ones were out I wouldn't have an issue but the medium core CPUs are always a gen behind....why for the love of god do medium cores always lag.

Cause its a market that basically no one cares about. Not ever been enough market demand for it. That it exists at all is as basically a bin dumping ground for server parts. And server parts require significantly more work and validation. One possible way it comes out faster is if they switch to sharing with Xeon-D instead of Xeon-EX.

On a side note, I am very wary about these switches because of latency. I heard they add a significant amount of latency but I have not heard or see any reports showing one or the other. I would not touch those switches if they added any latency because that will murder XPoint and I be damned to nerf that.

Lol, what adds latency is having to go to PCIe period. The latency added by any of the PCIe switches on the market is in the noise. And hell even with NVMe, the software stack is still the limiting factor in performance.

And if you think you are going to be able to get/afford Xpoint anytime soon.... Its probably best to think of XPoint as cheaper memory and not more expensive storage. AKA, expect NAND:XP cost ratios in the range of 1:20 to 1:40. So um, good luck with that. The first application is as memory extender and it will be priced for that market (think along the lines of 4K+ per TB at a minimum and likely higher). Plus its highly unlikely that you will have any workload that will take actual advantage of the performance of XPoint.

I want XPoint in RAM but supposedly those are not coming out first :/ and require special chipset support, which is odd considering ULTRADIMMS don't have that issue. I would actually buy that if it was available but sadly it is still blocked from sale. (or at least last I heard)

XPoint is basically only coming to the enterprise space and XPoint dimms will require e5/e7 14nm xeons.

 

[SomeGuy133]

No its not cause you are going to have a card with 50-200% more memory on it. Basically the only time the PCIe bus gets used is when the VRAM is filled or to send cmd lists (and cmd lists aren't taking any real bandwidth). In fact, I would bet that a 980TI would be LESS dependent on pcie performance right now than a 980.

And no my percents were correct. You want to focus on one game on a dead end engine which even the architect of the engine has admitted is a dead end and specifically stresses bus bandwidth due to poor architectural choices. Its not really representative at all.



8-12 lanes into a 24-32 port switch supporting 2 PCIe x8 slots. 4-8 lanes into a 20-24 port switch supporting 4-5 U.2/M.2 slots. More than enough bandwidth for either graphics or storage.

And no the high end SSDs don't use x8, they use x4. U.2 is the high end. Its what everyone is designing towards and has product in the market for. And XPoint? LOL, I don't think you even understand the market it is going into. No one from either product side is giving any though to -S dies and Xpoint. Besides, the only way you could potentially saturate x4 or x8 via NVMe storage is with moderate to deep queue depth sequential accesses, which isn't really useful. For real work, none of the drive announced or coming are pushing more than 1GB/s aka 1 PCIe lane.

As far as nextgen GPUs go, get a board with an Avago/PLX or PMC-Sierra switch chip. Done.




Cause its a market that basically no one cares about. Not ever been enough market demand for it. That it exists at all is as basically a bin dumping ground for server parts. And server parts require significantly more work and validation. One possible way it comes out faster is if they switch to sharing with Xeon-D instead of Xeon-EX.



Lol, what adds latency is having to go to PCIe period. The latency added by any of the PCIe switches on the market is in the noise. And hell even with NVMe, the software stack is still the limiting factor in performance.

And if you think you are going to be able to get/afford Xpoint anytime soon.... Its probably best to think of XPoint as cheaper memory and not more expensive storage. AKA, expect NAND:XP cost ratios in the range of 1:20 to 1:40. So um, good luck with that. The first application is as memory extender and it will be priced for that market (think along the lines of 4K+ per TB at a minimum and likely higher). Plus its highly unlikely that you will have any workload that will take actual advantage of the performance of XPoint.



XPoint is basically only coming to the enterprise space and XPoint dimms will require e5/e7 14nm xeons.

Dude they already stated it to be inbetween RAM and NAND so that is 2-3 bucks a GB and that means this guy is getting a 512GB the day it is released!

and this is a guy who also buys Xeons.. I got a 1650v3 next to me right now...again sucks the damn crap is always a gen behind :/ (if it wasn't a gen behind I would have bought the 8 core for a grand.

and again show me a data analysis of those chipsets not adding latency. I will be damned if I lower my 70k single IOPs or whatever it was.

Also XPoint DIMMs are server only. The PCIe cards are open game and I am getting a 512GB day it is released. DIMMS I would too but that isn't til 2017 :/ 256GB DIMM and 512GB PCIe with 2 TB standard SSD for data.....mmmmmmm

 

[aaronspink]

Dude they already stated it to be inbetween RAM and NAND so that is 2-3 bucks a GB and that means this guy is getting a 512GB the day it is released!

I think you are using the wrong pricing here. You should be thinking of the RAM pricing being things like 64/128GB DIMM pricing and the NAND pricing being things like the HGST SN100 pricing. AKA enterprise level pricing, not consumer level pricing.

and this is a guy who also buys Xeons.. I got a 1650v3 next to me right now...again sucks the damn crap is always a gen behind :/ (if it wasn't a gen behind I would have bought the 8 core for a grand.

If you are buying 1650v3s you aren't really buying Xeons, you are buying renamed HEDT products.

and again show me a data analysis of those chipsets not adding latency. I will be damned if I lower my 70k single IOPs or whatever it was.

You apparently have a whole lot more faith in the software stack than anyone who's actually in the industry. The switch chips are adding ns/10s of ns to paths that are in the 10s of microseconds range. AKA noise.

Also XPoint DIMMs are server only. The PCIe cards are open game and I am getting a 512GB day it is released. DIMMS I would too but that isn't til 2017 :/ 256GB DIMM and 512GB PCIe with 2 TB standard SSD for data.....mmmmmmm

You are assuming they will bother with anything as low as 512GB. The markets they are aiming products at are the markets where they already have 3-6TB of DRAM but need more. Where they've already switched over to high capacity NVMe, etc. Unlikely that they play around in the 512GB range except for maybe the baseline DIMMS.

 

[SomeGuy133]

I think you are using the wrong pricing here. You should be thinking of the RAM pricing being things like 64/128GB DIMM pricing and the NAND pricing being things like the HGST SN100 pricing. AKA enterprise level pricing, not consumer level pricing.



If you are buying 1650v3s you aren't really buying Xeons, you are buying renamed HEDT products.



You apparently have a whole lot more faith in the software stack than anyone who's actually in the industry. The switch chips are adding ns/10s of ns to paths that are in the 10s of microseconds range. AKA noise.



You are assuming they will bother with anything as low as 512GB. The markets they are aiming products at are the markets where they already have 3-6TB of DRAM but need more. Where they've already switched over to high capacity NVMe, etc. Unlikely that they play around in the 512GB range except for maybe the baseline DIMMS.

Hardly those HEDT are just feature stripped Xeons. Not sure how you go from lower quality product to higher lol

Also I would buy that 10/12 core unlocked but I obviously don't have 1800 bucks Other people on this form have those...lucky's.

They claimed consumer end being released 2016 so go back and read their press releases -_- Seriously, you are making this up as you go. Plus micron and Intel are both releasing this next year and each have difference markets so it is possible that micron had consumer end more in mind...or did you not know they both were in on this together and releasing independent products?

yea and 10ns can add up. Not sure how the math works out because not sure if the latency is static addition or only in switching but if that is a static 10ns that can murder IOs. Look at the difference in IOs between DIMMs and PCIe drives. The latency has a major affect. Again look at the documents intel released.

 

[aaronspink]

Hardly those HEDT are just feature stripped Xeons. Not sure how you go from lower quality product to higher lol

No, they are skus that are designed for HEDT that are resold as Xeons to fill an old product gap. But what do I know, I just worked on them.

Also I would buy that 10/12 core unlocked but I obviously don't have 1800 bucks Other people on this form have those...lucky's.

If they are using them for DT or WS then they are pretty much just throwing money away.

They claimed consumer end being released 2016 so go back and read their press releases -_- Seriously, you are making this up as you go. Plus micron and Intel are both releasing this next year and each have difference markets so it is possible that micron had consumer end more in mind...or did you not know they both were in on this together and releasing independent products?

yes consumer products priced higher than a P3700... And Micron also plays in the enterprise space as well or do you think P320h, P420m, S410m, and S6x0DCs are aimed at the consumer space? Christ, I knew this was coming years ago. Do you really think Intel/Micron are going to throw away money in the first year of a new technology when they can easily sell every chip they get into the enterprise space and make significantly more money?

yea and 10ns can add up. Not sure how the math works out because not sure if the latency is static addition or only in switching but if that is a static 10ns that can murder IOs. Look at the difference in IOs between DIMMs and PCIe drives. The latency has a major affect. Again look at the documents intel released.

No it isn't going to add up. You are comparing things that have latencies of nS to things that have latencies of uS and mS. Christ the storage controller for Optane drives alone is going to add at least 100-200ns. And you'll have a couple hundred ns just getting out of the PCIe complex on chip. So baseline latencies are going to be at least in the range of .5us and likely closer to 1-2us. Sure it won't be in the 10s of uS like flash based drives, but the latency through a PCIe switch isn't really going to be noticeable.

 

[SomeGuy133]

No, they are skus that are designed for HEDT that are resold as Xeons to fill an old product gap. But what do I know, I just worked on them.



If they are using them for DT or WS then they are pretty much just throwing money away.



yes consumer products priced higher than a P3700... And Micron also plays in the enterprise space as well or do you think P320h, P420m, S410m, and S6x0DCs are aimed at the consumer space? Christ, I knew this was coming years ago. Do you really think Intel/Micron are going to throw away money in the first year of a new technology when they can easily sell every chip they get into the enterprise space and make significantly more money?



No it isn't going to add up. You are comparing things that have latencies of nS to things that have latencies of uS and mS. Christ the storage controller for Optane drives alone is going to add at least 100-200ns. And you'll have a couple hundred ns just getting out of the PCIe complex on chip. So baseline latencies are going to be at least in the range of .5us and likely closer to 1-2us. Sure it won't be in the 10s of uS like flash based drives, but the latency through a PCIe switch isn't really going to be noticeable.

you have no idea what peoples uses are and are completely talking out of your ass. Some people do Photoshop with 4k screens and Photo shop with viewing 100 pixel for pixel images on screen requires insane amount of CPU

Others use these for rendering and video so please just stop. You are embarrassing yourself.

dude you really have no idea what your talking about these things latency are in the ns 100s. If and I mean if because i don't know if you even know anything your talking aobut but if there is a static 10ns added time on each transation that can murder IOs. Just stop talking. You can see the difference when you look at standard PCIe NAND vs ULTRADIMMs there is a huge difference in performance and that is simply because of the interface.....just stfu already. You have no idea what your saying.

I am going off of their press releases and their documentation of each technology. You are clearly making this up as you go roflcopter


since ignorance is bliss here go look at ULTRADIMM engineer sample review. This is standard NAND...nothing special. It is just in the DIMM and look how ridiculously faster it is because of the interface.

http://www.tweaktown.com/reviews/68...-ddr3-400gb-ssd-enterprise-review/index7.html

So yes 10 more ns per transaction can murder IOPs. If these thing are freakin cool image how much better XPoint will be when it naturally has 1000 times lower latency. There is a tangible difference going from SATA to PCIe to DIMM so please stop fooling yourself.

 

[aaronspink]

y
dude you really have no idea what your talking about these things latency are in the ns 100s. If and I mean if because i don't know if you even know anything your talking aobut but if there is a static 10ns added time on each transation that can murder IOs. Just stop talking. You can see the difference when you look at standard PCIe NAND vs ULTRADIMMs there is a huge difference in performance and that is simply because of the interface.....just stfu already. You have no idea what your saying.

No, you have no idea what you are talking about. Sandisk UltraDIMMs have a read latency of 150uS and a write latency of 5uS. In fact, UltraDIMMs have a read latency roughly 2x that of Sandisk FusionIO SX350s. The main advantage of the UltraDIMMs is their write latency or would be if you could still buy them and the weren't embroiled in a patent lawsuit and a trade secret lawsuit.

As far as XPoint, the chips themselves are estimated to have latencies in the 100s of nS, but the raw chips won't do you that much good. They need a controller and that controller is going to have latency and that latency is almost assuredly going to be in the 100s of nS range realistically (they are basically uC engines running firmware, generally using low to mid end ARM processors). It takes 100s of nS generally just to get through the PCIe root complex and into a PCIe device, so add that on. So you are looking at in the absolute best case close to or above 1000nS for a PCIe based XPoint device. They'll still be significantly faster than a NAND based device because those generally run in the 40-60uS range in actual use. Hell just the NVMe software stack alone is ~2800nS! So you are looking at latency in the range of 4-5uS or 4000-5000nS.

So, I would suggest that it is in fact, you, that has no idea what they are saying.

I am going off of their press releases and their documentation of each technology. You are clearly making this up as you go roflcopter

Then you should probably stay away from things like press releases and documentation because you have a demonstrated lack of ability to understand what you are reading and what it actually means.

since ignorance is bliss here go look at ULTRADIMM engineer sample review. This is standard NAND...nothing special. It is just in the DIMM and look how ridiculously faster it is because of the interface.

http://www.tweaktown.com/reviews/68...-ddr3-400gb-ssd-enterprise-review/index7.html

Call me when you can buy an UltraDIMM or an XPointDIMM and have a server it will work in... I'm expecting initial prices for XPoint DIMMs in the 4-8k+ range and support in 2P+ Xeons only.

So yes 10 more ns per transaction can murder IOPs. If these thing are freakin cool image how much better XPoint will be when it naturally has 1000 times lower latency. There is a tangible difference going from SATA to PCIe to DIMM so please stop fooling yourself.

None of the technologies you are comparing even have a latency below 5000nS. Even Fusion IO's SX350 (which AFAIK is the lowest latency write device available on market) or HGST's SN100 series (which is the lowest latency read device on the market) have latency in the 15000nS to 20000nS range. UltraDIMM has latency of 150000nS for reads and 5000nS for writes. And the only reason UltraDIMMs get that low on writes is that they remove the whole entire I/O complex.

So unless they are going to do something different than NVMe (unlikely), no, the latency of a PCIe switch chip is basically immaterial.

If you actually want to learn, this might be a decent research paper to start off with: https://www.usenix.org/system/files/conference/fast14/fast14-paper_vucinic.pdf This is basically on overview of work that HGST was doing to get to the absolute lowest read latency possible for PCM memory (~130ns device read latency) by basically going to a super eager pulling protocol instead of the doorbell protocol used by NVMe. You should probably pay attention to Figure 4 and Table 5. But the whole paper is pretty good and goes through some of the various bottlenecks in NVMe wrt latency and low latency storage devices.

 

[aaronspink]

And now we have published latency for an Optane NVMe drive of 9uS. I fully rest my case.

http://www.theplatform.net/2015/10/28/intel-shows-off-3d-xpoint-memory-performance/

 

[SomeGuy133]

And now we have published latency for an Optane NVMe drive of 9uS. I fully rest my case.

Give me that 250ns please!

http://www.theplatform.net/2015/10/28/intel-shows-off-3d-xpoint-memory-performance/

awesome thanks! You proved my point. Their press release was completely accurate and DIMMs are going to be freaking awesome

Tjhey said less than 10 us and they got 9 us in that test so spot on. That leads you to believe that the 250ns for DIMMs will be realistic!

So according to that the DIMMS will have 36 times lower latency Tangible

I fully rest my case.

oh and dido

 

[aaronspink]

awesome thanks! You proved my point. Their press release was completely accurate and DIMMs are going to be freaking awesome

Tjhey said less than 10 us and they got 9 us in that test so spot on. That leads you to believe that the 250ns for DIMMs will be realistic!

So according to that the DIMMS will have 36 times lower latency Tangible

Yes but I expect DIMMs to be PRICEY. Probably looking at 1-2TB dimms at the start priced in the 2-4k+ range. Basically 8-16k+ for a set. Just don't see anyone outside of the enterprise space paying the money for that. Esp when you'll likely need a 4-8k processor to go along with it.