Using Optane as a (hopefully cheaper and more dense) virtual memory backing store is a possibility if the processor itself did the paging from optane to dram (like TLBs are handled now). At 12 GB/s a 4K page should load in less than a microsecond. Could have some interesting performance impacts on things like "memory resident" databases.
Obviously, this would be a big change to the processor, so not something that is going to happen soon...
Enthusiasts need capacity, at the very least 100-200GB at 2$ or better per GB. 16-32GB can be sold on marketing to uneducated users. It's an OEM play.
Games are very large and so are serious apps and the input/output files. Hybrids are hard to benchmark in consumer and show their value even when they offer good value. With 16-32GB capacity, it would compete against DRAM not NAND as more DRAM would offer better value for enthusiasts or proffessionals in desktop. DRAM prices have somewhat doubled vs a year ago and likely to keep rising this year, that could help Xpoint a bit- a lot in server. On the other hand, midrange PCIe 3.0 x4 SSDs, retailing at bellow 50 cents per GB, are rather compelling. PCIe 4.0 should help Xpoint and ,soon enough, WD's 3D ReRAM as it would make it easier to show the difference vs NAND.
The one odd part is that Intel is about to be kicked out of the enthusiast market in desktop. AMD seems about ready to offer twice the cores at same price point and if Intel ties to keep Optane Intel only, they would severely limit its potential.
Sounds like you're advocating Chromebooks. I agree that they are an approach that makes a lot of sense in many environments. I also agree that they don't need a high-capacity HDD and Optane - a small SSD would do just fine.
Perhaps, in time, the PC business will split into two distinct markets: cloud-based and self-sufficeint. Let's check back in a decade!
Although I do believe that some of the more sophisticated PC buyers that you mention (gamers, photographers, etc.) will like what Optane m.2 XPoint SSDs will do for their system, I am not convinced that this will drive enough volume to make your first two factors play out. Time will tell.
I agree that a block interface helps to hide a lot of sins. That's why Samsung's been shipping all of its 3D NAND in SSDs.
Thanks for the info on limited m.2 sockets in notebooks. I didn't know that.
I don't think, though, that you need a "faster than DRAM" layer. That's what the processor caches are for. Since the speed difference between flash and XPoint is said to be 1,000:1 then XPoint will be a good layer to add between DRAM and an SSD. Note that Intel is positioning it as a layer between an HDD and DRAM to bridge the same gap that is addressed by an SSD.
In the near term we (at Objective Analysis) see 3D XPoint being used as memory, rathe than storage. There isn't yet off-the-shelf software that can use its persistence. Over time this will change, but in the near term XPoint should serve as a fast memory layer. This should appeal to server users, but less so to PC users.
I am a lot more positive on the technology than you appear to be.
As long as 3D XPoint memory can sell for less than DRAM and can perform faster than NAND flash then it's a reasonable technology to use as a new layer in the memory/storage hierarchy.
My question in this article was whether the average PC buyer would feel comfortable spending money for this feature. I don't expect for that to happen. I do expect to see adoption in the server, though.
I think there needs to be a distinction whether Optane will try to compete in the storage or memory industry. Both are quite different, but I will address issues for both.
As a storage device, it will be competing with 3DVNAND which currently has a very strong hold on the market. Even without knowledge of the pricetag, the announced capacities of 16 and 32 gb leaves a lot to be desired, as M.2 and SATA Flash SSDs have already broken the 1TB barrier at not unreasonable prices. Optane does boast some performance advantages such as byte level addressing, and if Optane uses some variation of PCRAM technology, its write/erase speed can be estimated to be in the 100 ns range which is significantly faster than Flash's 1 ~ 2 us erase speed. Whether those advantages will have significant measurable performance increases to make up for its small capacity to be accepted in the mainstream market however, is extremely doubtful in my opinion.
As a memory device, it does boast non volatility, but PCRAM's write speed of 100 ns is unsuitable as a memory application, whereas DRAM can have write speeds of less than 10 ns. In addition, I haven't seen evidence that Optane has infinite write/erase endurance, which is a requirement for a memory application.
While there could be surprises awaiting us, the way it looks now and based on my knowledge in this field, Optane will stand in the shadows of Flash and DRAM, making it a niche product at best.
True, but for fast boot up on a client PC, how much memory do you need? 4 GB? I rarely actually reboot my laptop, so I "restart" instantly from battery backed DRAM, which costs about $160 for 32GB. I think to benefit clients, we need to focus on "faster than DRAM", even if it costs a little more, and make DRAM the "slower" layer.
On my client PC, when I first switched from a rotating hard drive to a first gen SSD at just 300 MB/second, the difference was dramatic. But when I later upgraded to a 1.5GB/sec SSD, the incremental difference was just barely noticeable. I only notice faster speeds when transferring large files, but then a cache system doesn't work - you need the entire hard drive to be that fast.
While Optane (and QuantX) isn't a great fit for the PC market before the spigot turns on to 9, it may turn into one.
To be clear, it will be hard for 3DXP to ever surplant NAND ever as cost/die size always matters. But consider the timing: a 2 TB 3DXP SSD this year would be as foolish a product as a 2 TB NAND SSD would have been in 1985 (a year after NAND first shipped).
Factor one: From a memory maker's perspective, manufactured wafers must be shipped. This is a solid truth in solid state. There are few technical barriers in the PC market from adoption other than price/performance. The market for block controlled 3DXP will find an economic balance as the early devices want to be sold.
Factor two: Nothing helps improve the manufacturing process more than the data of manufacturing. 3DXP (or any memory) needs to be manufactured to improve manufacturing to result in an eventual, volume-supported, high-value memory solution. What to do with the early wafer results? See Factor one.
Factor three: The PC market and block usage (vs. memory usage) is more tolerant of 'good' early memory that doesn't meet all specifications as the block controller can do well-known 'magic' to correct many issues. Until the manufacturing process is dialed in to enable the high value memory market with volume, the PC market is a great win-win outlet. Gamers, video processors, pro photographers and other storage centric pro's will see a benefit from an 10x speed up at a cost less than DRAM for 'small' (e.g., 16, 32 GB) accelerators.
There are multiple protocol specifications for storage cache/acceleration already developed for NAND/HDD that can be taken advantage of for this application (or a heck of a lot of learning already accomplished). NVMe has some too,
One interesting physical barrier is how to shoe-horn a M.2 3DXP-based NVMe device into lap-top/2-in-1 designs (or XBox 2?). The one M.2 slot is often busy with the main M.2 NAND-based NVMe SSD. Perhaps the newer controllers enabling NVMe SSDs in BGA form (e.g., Phison E8T) will tend to free the one slot. Two slots isn't impossible, but this choice is possibly one that Intel is mulling over as we discuss this. (I'd expect the drop in peak performance from a four-lane PCIe NVME M.2 device moving to a two-lane PCIe NVMe BGA device would be more than offset by a four PCIe lane Optane/QuantX M.2)
Balancing process/memory/SCM/Flash will happen over time, My educated guess is that Optane/QuantX will start as a performance plugin that will price itself into a supply-absorbing market.
Perhaps the issue is that in a corporate environment, today's HDDs are way bigger than needed. It is hard to find a desktop with anything smaller than a 500-1000 GB HDD. But a typical install might be only 50-100 GB, since user files are on a file server. If speed is important, it is probably cheaper to go with 128 GB of SDD than HDD + Optane. For a home PC, fast startup from some flash is probably more cost effective. Even for home PCs the trend is for more cloud storage of user files, which makes all-SDD look attractive, especially considering that SDD instead of HDD and a slower processor might be faster for the same price.