They make the chips in my PS4 and the discrete video of Macs. They're even the CPU of your precious Xbox One. so yeah, I've seen them lately.
I don't have an Xbox One, AMD is so far behind technologically they might as well be in reverse, and your snide post earned a blocklist spot. Congrats!
According to the keynote (link from my previous post), the first product is a 128 GB chip. They showed a wafer and the chips appeared about the size of a DRAM chip.
On my iMac 5K, the SSD is 128 GB -- so 8 of these chips would be required.
disk0s2: Media Name: APPLE SSD SD0128F Media Size: 120.99 GB (120,988,852,224 bytes) Medium Type: SSD Protocol: PCI Internal: Yes Partition Map Type: GPT (GUID Partition Table) Status: Online S.M.A.R.T. Status: Verified PV UUID: 1694ADBE-6DFC-442B-9D03-43540288FDD2 disk1s2: Media Name: APPLE HDD ST3000DM001 Media Size: 3 TB (2,999,733,223,424 bytes) Medium Type: Rotational Protocol: SATA Internal: Yes Partition Map Type: GPT (GUID Partition Table) Status: Online S.M.A.R.T. Status: Verified PV UUID: C1220402-9B68-40D2-B6EB-90EFDADE293A
One of the uses discussed in the video was [B][I] pattern matching [/I][/B] -- where relative large amounts of data need to be processed at high speed ...
It would seem that things like Siri, facial recognition, gesture recognition, scene recognition, etc. are naturals for a mobile device.
I don't have an Xbox One, AMD is so far behind technologically they might as well be in reverse, and your snide post earned a blocklist spot. Congrats!
We are talking here about speed of 200GB/s ( SSD ~200MB/s x 1000 ) where DDR3 speed ~17GB/s .. pinch me
The speed they referred to was latency not bandwidth:
It's 1000x faster at getting data than NAND (latency in milliseconds/microseconds) but slower than DRAM (latency in nanoseconds). They don't say what the bandwidth capability is but probably also somewhere between Flash at ~2-3GB/s for newer Flash drives and 30-60GB/s for DDR3/DDR4 RAM. The Mac Pro's D700 has 264GB/s of memory bandwidth. NVidia's Pascal is supposed to deliver 1TB/s using 3D stacked memory.
Having 10x the density of DRAM (like SSD density) is good though because they could even put a memory layer inside a CPU or inside a DRAM module. They can keep the operating system inside the 128GB part or really any virtual memory they need. Boot times could almost disappear. It would be like having 128GB of RAM because applications could shuttle data back and forward to it. Apple uses 16x 1GB DDR3 chips inside the laptops. They could cut out maybe 4-8 chips, switch to DDR4 to have 8-12x 2GB = 16-24GB and add 64-128GB of XPoint memory. Apple could store the system recovery partition on there too without partitioning the main storage.
They said this particular memory will be more expensive than NAND. Given that they mentioned they'd be disrupting NAND pricing:
I expect they must be planning to use the same type of memory design to produce something more competitive to existing SSDs. Maybe they were talking about SLC SSD, which is more expensive but they should be able to do something more competitive in consumer SSDs than they have done so far.
XPoint could perhaps make it into mobile devices but they are very price-sensitive. Even $100 extra would be hard to absorb into the price. We'll see what products they deliver in 2016. The eventual goal is to have completely unified memory between GPU, CPU and data storage likely with faster supplemental memory caches.
We are talking here about speed of 200GB/s ( SSD ~200MB/s x 1000 ) where DDR3 speed ~17GB/s .. pinch me
The speed they referred to was latency not bandwidth:
It's 1000x faster at getting data than NAND (latency in milliseconds/microseconds) but slower than DRAM (latency in nanoseconds). They don't say what the bandwidth capability is but probably also somewhere between Flash at ~2-3GB/s for newer Flash drives and 30-60GB/s for DDR3/DDR4 RAM. The Mac Pro's D700 has 264GB/s of memory bandwidth. NVidia's Pascal is supposed to deliver 1TB/s using 3D stacked memory.
Having 10x the density of DRAM (like SSD density) is good though because they could even put a memory layer inside a CPU or inside a DRAM module. They can keep the operating system inside the 128GB part or really any virtual memory they need. Boot times could almost disappear. It would be like having 128GB of RAM because applications could shuttle data back and forward to it. Apple uses 16x 1GB DDR3 chips inside the laptops. They could cut out maybe 4-8 chips, switch to DDR4 to have 8-12x 2GB = 16-24GB and add 64-128GB of XPoint memory. Apple could store the system recovery partition on there too without partitioning the main storage.
They said this particular memory will be more expensive than NAND. Given that they mentioned they'd be disrupting NAND pricing:
I expect they must be planning to use the same type of memory design to produce something more competitive to existing SSDs. Maybe they were talking about SLC SSD, which is more expensive but they should be able to do something more competitive in consumer SSDs than they have done so far.
XPoint could perhaps make it into mobile devices but they are very price-sensitive. Even $100 extra would be hard to absorb into the price. We'll see what products they deliver in 2016. The eventual goal is to have completely unified memory between GPU, CPU and data storage likely with faster supplemental memory caches.
I wonder if XPoint could be used to replace SRAM function -- which would also offset costs -- especially for things like the Apple Watch:
The 4Gb of SRAM is instead of 512 MB of DRAM,
IDK the price of SRAM, but I think it is more expensive than DRAM.
and it would save space. If the overall performance was enough, it could be a good fit for wearables. They aren't playing games or anything like that and the screen resolution is low. They might have a small video memory cache on top but store the bulk of the data in the XPoint memory and this replaces both the 512MB (4Gb) of RAM and the 8GB of NAND.
The suitability of XPoint will all depend on how beneficial it will be for each device. While it seems like a good replacement for a few existing technologies when you look at the advantages e.g higher density and cheaper than RAM, higher durability and performance than NAND, the reason they use separate technology like RAM and NAND is because there is a need for both the highest performance memory for real-time use and cheap memory for bulk storage. XPoint will be more expensive than NAND and slower and less durable than RAM.
If they can get the price down, it can replace NAND no problem. If they can get the performance up, it can replace RAM. If they can't do either of those things, it might end up having limited applications. One suggested use was in the server space where they'd split virtual machines up so they can have lots of cheap memory and share it with storage.
The following site quotes the Micron shareholder presentation saying they will be developing multiple XPoint technologies:
That site also says XPoint was described as 10x faster than NAND over NVMe. So bandwidth must be in the region of 10-25GB/s. If they can manufacture it cheaply the way MLC/TLC NAND is then perhaps they can get it price-competitive with NAND, even if it came with a performance hit. If they optimize the interface to the CPU or integrate the memory into the CPU, perhaps they can improve latency and bandwidth further to replace RAM in some places.
It sounds promising for a brand new technology. If it can be competitive with both RAM and NAND in its first generation manufacturing, it will only get better with scaling while NAND and RAM slow down their improvements.
What if they could offset the additional cost with less battery, smaller size, replace SRAM/DRAM/Flash at acceptable performance.
It could really affect current wearables like the Apple Watch -- and open up whole new possibilities for future wearables.
This is my hope for a second or third version of the Watch.
It appears that are lots of possibilities with AppleWatch 2 and beyond -- it is not insignificant that almost all the Apple/IBM Mobile First offerings include Apple Watch as well as iPad & iPhone!
and it would save space. If the overall performance was enough, it could be a good fit for wearables. They aren't playing games or anything like that and the screen resolution is low. They might have a small video memory cache on top but store the bulk of the data in the XPoint memory and this replaces both the 512MB (4Gb) of RAM and the 8GB of NAND.
The suitability of XPoint will all depend on how beneficial it will be for each device. While it seems like a good replacement for a few existing technologies when you look at the advantages e.g higher density and cheaper than RAM, higher durability and performance than NAND, the reason they use separate technology like RAM and NAND is because there is a need for both the highest performance memory for real-time use and cheap memory for bulk storage. XPoint will be more expensive than NAND and slower and less durable than RAM.
If they can get the price down, it can replace NAND no problem. If they can get the performance up, it can replace RAM. If they can't do either of those things, it might end up having limited applications. One suggested use was in the server space where they'd split virtual machines up so they can have lots of cheap memory and share it with storage.
The following site quotes the Micron shareholder presentation saying they will be developing multiple XPoint technologies:
That site also says XPoint was described as 10x faster than NAND over NVMe. So bandwidth must be in the region of 10-25GB/s. If they can manufacture it cheaply the way MLC/TLC NAND is then perhaps they can get it price-competitive with NAND, even if it came with a performance hit. If they optimize the interface to the CPU or integrate the memory into the CPU, perhaps they can improve latency and bandwidth further to replace RAM in some places.
It sounds promising for a brand new technology. If it can be competitive with both RAM and NAND in its first generation manufacturing, it will only get better with scaling while NAND and RAM slow down their improvements.
Wow wow wow, we are getting too far ahead of ourselves.
First, the speed was referring to latency, not bandwidth. And from Anandtech and other sources, Xpoint latency is likely to be 2x to 10x slower then RAM. Which is why Intel said it wont be a replacement for RAM.
CPU is extremely sensitive to Latency, That is why we now have L1 - L3 Cache, and L4 if you count Intel eDRAM. As a matter of fact the Latency improvement over the years for DRAM does not match the speed of CPU. That is why you see little to no performance improvement in DDR3 to DDR4. Since the latency is similar and they trade improvement for Bandwidth. DDR4 or LPDDR4 mainly focuses on low power and higher density, both are important to the target market, Server and Mobile Phones / Tablet.
Theoretically, Xpoint could server as RAM once we have huge L4 / eDRAM. But on mobile SoC the graphics will still need a huge amount of bandwidth which is not available from Xpoint. So you would still need LPDDR4 / Wide I/O 2 for Graphics Operation.
For SSD, Xpoint does not require a sophisticated controller like NAND do, with all sort of clever tricks and GC just to make it faster. This will save the power requirement from SSD, but i am not sure if this will justify the cost.
What if they could offset the additional cost with less battery, smaller size, replace SRAM/DRAM/Flash at acceptable performance.
It could really affect current wearables like the Apple Watch -- and open up whole new possibilities for future wearables.
Not everything has an equal value, in all use cases, Today we are trading off storage reliability for size and battery life. As long as you keep updating your hardware before the memory wears out then it is not an issue, but would you trust all you personal important information on a device which at some point in the near future will stop working and you can not recover that information. This is why I still store all my infromation on a HDD, when ever it is on a flash based product. This is one of the reason I like Apple since they make it easy to back up your i-dives unlike Andriod, Google backs up but not everything but only into the cloud where they can look around your information.
What if they could offset the additional cost with less battery, smaller size, replace SRAM/DRAM/Flash at acceptable performance.
It could really affect current wearables like the Apple Watch -- and open up whole new possibilities for future wearables.
Not everything has an equal value, in all use cases, Today we are trading off storage reliability for size and battery life. As long as you keep updating your hardware before the memory wears out then it is not an issue, but would you trust all you personal important information on a device which at some point in the near future will stop working and you can not recover that information. This is why I still store all my infromation on a HDD, when ever it is on a flash based product. This is one of the reason I like Apple since they make it easy to back up your i-dives unlike Andriod, Google backs up but not everything but only into the cloud where they can look around your information.
Above Italics mine - I just want add that Apple bought an Israeli company several years ago that had expertise in writing controllers for the flash storage. Now and then I'll hear a bit of rumor about the company, but for some reason Apple has never made an announcement about how this company's controllers allow the flash memory to last longer then others. If this is part of Apple products, and I think it is, somewhere down the line it should start showing itself as all the other competitor's flash start to die and older Apple iDevices keep on keeping on...
Above Italics mine - I just want add that Apple bought an Israeli company several years ago that had expertise in writing controllers for the flash storage. Now and then I'll hear a bit of rumor about the company, but for some reason Apple has never made an announcement about how this company's controllers allow the flash memory to last longer then others. If this is part of Apple products, and I think it is, somewhere down the line it should start showing itself as all the other competitor's flash start to die and older Apple iDevices keep on keeping on...
Everyone is looking for the secert sauce which will will extend falsh memory's life. The problem is they shink the lithography faster than they can come up with algorithms to make it last longer. The long and short of it, most of what they are trying to do is wear level the data storage cells. They continue to move data around so one cell does not get erase more than any other. If it does, then they map around it so data is not writen to it. There has been some advances in how to write and refeash the data so not to wear it out too quickly but it is not significant enough to make much of a difference. Flash is really good for write once and read many application like a music player. But when you writing and erasing data all the time like most modern OS and web browsers due you can wipe out flash in a few years. Apple has been using the technology for the company from Israel for a long time, since they inbed the flash in the SOC and use their own controller verse using the controller from the flash suppliers. This allows them to switch between flash suppliers with out having to rewrite softwere to work with each suppliers controller.
Comments
I don't have an Xbox One, AMD is so far behind technologically they might as well be in reverse, and your snide post earned a blocklist spot. Congrats!
On my iMac 5K, the SSD is 128 GB -- so 8 of these chips would be required.
disk0s2:
Media Name: APPLE SSD SD0128F Media
Size: 120.99 GB (120,988,852,224 bytes)
Medium Type: SSD
Protocol: PCI
Internal: Yes
Partition Map Type: GPT (GUID Partition Table)
Status: Online
S.M.A.R.T. Status: Verified
PV UUID: 1694ADBE-6DFC-442B-9D03-43540288FDD2
disk1s2:
Media Name: APPLE HDD ST3000DM001 Media
Size: 3 TB (2,999,733,223,424 bytes)
Medium Type: Rotational
Protocol: SATA
Internal: Yes
Partition Map Type: GPT (GUID Partition Table)
Status: Online
S.M.A.R.T. Status: Verified
PV UUID: C1220402-9B68-40D2-B6EB-90EFDADE293A
One of the uses discussed in the video was [B][I] pattern matching [/I][/B] -- where relative large amounts of data need to be processed at high speed ...
It would seem that things like Siri, facial recognition, gesture recognition, scene recognition, etc. are naturals for a mobile device.
You are offended by the strangest things...
The speed they referred to was latency not bandwidth:
It's 1000x faster at getting data than NAND (latency in milliseconds/microseconds) but slower than DRAM (latency in nanoseconds). They don't say what the bandwidth capability is but probably also somewhere between Flash at ~2-3GB/s for newer Flash drives and 30-60GB/s for DDR3/DDR4 RAM. The Mac Pro's D700 has 264GB/s of memory bandwidth. NVidia's Pascal is supposed to deliver 1TB/s using 3D stacked memory.
Having 10x the density of DRAM (like SSD density) is good though because they could even put a memory layer inside a CPU or inside a DRAM module. They can keep the operating system inside the 128GB part or really any virtual memory they need. Boot times could almost disappear. It would be like having 128GB of RAM because applications could shuttle data back and forward to it. Apple uses 16x 1GB DDR3 chips inside the laptops. They could cut out maybe 4-8 chips, switch to DDR4 to have 8-12x 2GB = 16-24GB and add 64-128GB of XPoint memory. Apple could store the system recovery partition on there too without partitioning the main storage.
They said this particular memory will be more expensive than NAND. Given that they mentioned they'd be disrupting NAND pricing:
http://www.thessdreview.com/daily-news/latest-buzz/intels-rob-crooke-speaks-3d-memory-10tb-ssds-2014-intel-investor-meeting/
I expect they must be planning to use the same type of memory design to produce something more competitive to existing SSDs. Maybe they were talking about SLC SSD, which is more expensive but they should be able to do something more competitive in consumer SSDs than they have done so far.
XPoint could perhaps make it into mobile devices but they are very price-sensitive. Even $100 extra would be hard to absorb into the price. We'll see what products they deliver in 2016. The eventual goal is to have completely unified memory between GPU, CPU and data storage likely with faster supplemental memory caches.
Impressive tech!
I wonder if XPoint could be used to replace SRAM function -- which would also offset costs -- especially for things like the Apple Watch:
The 4Gb of SRAM is instead of 512 MB of DRAM,
IDK the price of SRAM, but I think it is more expensive than DRAM.
It's amazing what they can do when their CPU business is gone to shit...
Margins on memory is not that great, so its not like they're replacing what they're losing. More like cutting the bleeding.
It would retain the benefit of not needing to be refreshed like DRAM:
Difference between static RAM and dynamic RAM
and it would save space. If the overall performance was enough, it could be a good fit for wearables. They aren't playing games or anything like that and the screen resolution is low. They might have a small video memory cache on top but store the bulk of the data in the XPoint memory and this replaces both the 512MB (4Gb) of RAM and the 8GB of NAND.
The suitability of XPoint will all depend on how beneficial it will be for each device. While it seems like a good replacement for a few existing technologies when you look at the advantages e.g higher density and cheaper than RAM, higher durability and performance than NAND, the reason they use separate technology like RAM and NAND is because there is a need for both the highest performance memory for real-time use and cheap memory for bulk storage. XPoint will be more expensive than NAND and slower and less durable than RAM.
If they can get the price down, it can replace NAND no problem. If they can get the performance up, it can replace RAM. If they can't do either of those things, it might end up having limited applications. One suggested use was in the server space where they'd split virtual machines up so they can have lots of cheap memory and share it with storage.
The following site quotes the Micron shareholder presentation saying they will be developing multiple XPoint technologies:
http://www.theregister.co.uk/2015/07/29/having_a_looks_at_imtfs_crosspoint/?page=3
That site also says XPoint was described as 10x faster than NAND over NVMe. So bandwidth must be in the region of 10-25GB/s. If they can manufacture it cheaply the way MLC/TLC NAND is then perhaps they can get it price-competitive with NAND, even if it came with a performance hit. If they optimize the interface to the CPU or integrate the memory into the CPU, perhaps they can improve latency and bandwidth further to replace RAM in some places.
It sounds promising for a brand new technology. If it can be competitive with both RAM and NAND in its first generation manufacturing, it will only get better with scaling while NAND and RAM slow down their improvements.
What if they could offset the additional cost with less battery, smaller size, replace SRAM/DRAM/Flash at acceptable performance.
It could really affect current wearables like the Apple Watch -- and open up whole new possibilities for future wearables.
This is my hope for a second or third version of the Watch.
It appears that are lots of possibilities with AppleWatch 2 and beyond -- it is not insignificant that almost all the Apple/IBM Mobile First offerings include Apple Watch as well as iPad & iPhone!
Thx for your knowledgeable perspective!
Wow wow wow, we are getting too far ahead of ourselves.
First, the speed was referring to latency, not bandwidth. And from Anandtech and other sources, Xpoint latency is likely to be 2x to 10x slower then RAM. Which is why Intel said it wont be a replacement for RAM.
CPU is extremely sensitive to Latency, That is why we now have L1 - L3 Cache, and L4 if you count Intel eDRAM. As a matter of fact the Latency improvement over the years for DRAM does not match the speed of CPU. That is why you see little to no performance improvement in DDR3 to DDR4. Since the latency is similar and they trade improvement for Bandwidth. DDR4 or LPDDR4 mainly focuses on low power and higher density, both are important to the target market, Server and Mobile Phones / Tablet.
Theoretically, Xpoint could server as RAM once we have huge L4 / eDRAM. But on mobile SoC the graphics will still need a huge amount of bandwidth which is not available from Xpoint. So you would still need LPDDR4 / Wide I/O 2 for Graphics Operation.
For SSD, Xpoint does not require a sophisticated controller like NAND do, with all sort of clever tricks and GC just to make it faster. This will save the power requirement from SSD, but i am not sure if this will justify the cost.
What if they could offset the additional cost with less battery, smaller size, replace SRAM/DRAM/Flash at acceptable performance.
It could really affect current wearables like the Apple Watch -- and open up whole new possibilities for future wearables.
Not everything has an equal value, in all use cases, Today we are trading off storage reliability for size and battery life. As long as you keep updating your hardware before the memory wears out then it is not an issue, but would you trust all you personal important information on a device which at some point in the near future will stop working and you can not recover that information. This is why I still store all my infromation on a HDD, when ever it is on a flash based product. This is one of the reason I like Apple since they make it easy to back up your i-dives unlike Andriod, Google backs up but not everything but only into the cloud where they can look around your information.
What if they could offset the additional cost with less battery, smaller size, replace SRAM/DRAM/Flash at acceptable performance.
It could really affect current wearables like the Apple Watch -- and open up whole new possibilities for future wearables.
Not everything has an equal value, in all use cases, Today we are trading off storage reliability for size and battery life. As long as you keep updating your hardware before the memory wears out then it is not an issue, but would you trust all you personal important information on a device which at some point in the near future will stop working and you can not recover that information. This is why I still store all my infromation on a HDD, when ever it is on a flash based product. This is one of the reason I like Apple since they make it easy to back up your i-dives unlike Andriod, Google backs up but not everything but only into the cloud where they can look around your information.
Above Italics mine - I just want add that Apple bought an Israeli company several years ago that had expertise in writing controllers for the flash storage. Now and then I'll hear a bit of rumor about the company, but for some reason Apple has never made an announcement about how this company's controllers allow the flash memory to last longer then others. If this is part of Apple products, and I think it is, somewhere down the line it should start showing itself as all the other competitor's flash start to die and older Apple iDevices keep on keeping on...
Above Italics mine - I just want add that Apple bought an Israeli company several years ago that had expertise in writing controllers for the flash storage. Now and then I'll hear a bit of rumor about the company, but for some reason Apple has never made an announcement about how this company's controllers allow the flash memory to last longer then others. If this is part of Apple products, and I think it is, somewhere down the line it should start showing itself as all the other competitor's flash start to die and older Apple iDevices keep on keeping on...
Everyone is looking for the secert sauce which will will extend falsh memory's life. The problem is they shink the lithography faster than they can come up with algorithms to make it last longer. The long and short of it, most of what they are trying to do is wear level the data storage cells. They continue to move data around so one cell does not get erase more than any other. If it does, then they map around it so data is not writen to it. There has been some advances in how to write and refeash the data so not to wear it out too quickly but it is not significant enough to make much of a difference. Flash is really good for write once and read many application like a music player. But when you writing and erasing data all the time like most modern OS and web browsers due you can wipe out flash in a few years. Apple has been using the technology for the company from Israel for a long time, since they inbed the flash in the SOC and use their own controller verse using the controller from the flash suppliers. This allows them to switch between flash suppliers with out having to rewrite softwere to work with each suppliers controller.