Printable View
On 7/22/2020, following was reported:
"Sabrent's sizeable PCIe 4.0 SSD will set you back $750
"Sabrent quietly rolled out a 4TB PCIe SSD, making it twice the size of the previously largest PCIe SSD.
"It's also the largest PCIE 4.0 SSD on the market."
As such, 4 @ 4TB = 16TB
$3,000 / 16,000GB = $0.1875 per gigabyte
Also, ASRock's 4x4 design supports multiple AICs
installed in a single motherboard e.g.
2 @ 16TB = 32TB using only 2 x16 PCIe 4.0 expansion slots.
I heard about this recently:
https://www.servethehome.com/microch...memory-future/
Serial interface for RAM, with the flexibility to stick whatever you like at the end of it (DDRX, GDDRX, SSD, HDD, a big stack of uSD cards getting delivered by pigeon, etc). IIRC Micron has a licence to make optane (or at least something very similar to optane, but probably without the branding) - potentially we could see a chip vendor agnostic version of the optane persistent memory, which would be nice
Sure I meant by how RAID setups are for boosting throughput. While very useful on HDDs, and still quite useful on SSDs, the ultra low latency on Optane means the overhead actually serves to reduce performance noticeably(up to 2x in small file random latency) which may somewhat counteract the improvements you get from faster sequential throughput.Quote:
Originally Posted by supremelaw
The M.2 cache drives have crappy sequential throughput(one which I believe is related to power) and in that case it might help to RAID them but its an academic exercise not much useful for people because the high dollar cost per GB, especially for the cache drives. You are better off buying a single 900P/905P series instead because you won't need to sacrifice on latency and you get decent throughput. Still if you are after throughput get a regular SSD instead.
The RAMDisks on DIMMs have an issue too. For the Optane SSDs, you have 10us latencies. With the DIMMs you can get 180-300ns, but that's if you use as regular memory using Memory Mode, or application recompile called App Direct. If you want to use it out of the box as storage, you need to have it work with block storage with 4KB addressing sizes and your latency will shoot up to ~4us. That's true even for regular DRAM not just Optane. There will be value in using the 4KB block storage mode with the DIMMs, because 4us is still pretty low latency but you lose a lot of advantages not skipping the software overhead.
FYI: this showed up in my email INBOX yesterday: at Highpoint's website, see:
"Cutting-edge PCIe Gen 4 x16 NVMe RAID Storage performance"
The model SSD7540 looks interesting: 8 x M.2
A comparison I would like to see is an objective benchmark
of Optane DIMMs and the latest PCIe Gen4 "4x4" add-in cards.
The latter should measure both standard NVMe M.2 SSDs like popular Samsung models,
and the latest Optane M.2 SSDs.
Also, there are 2 major variants of those "4x4" add-in cards:
one variant requires bifurcation in the chipset, and
the other variant does not require bifurcation e.g. Highpoint's designs.
Because PCIe 4.0 is now rolling out and retail products are available,
I'm realistically expecting that Optane devices and non-Optane devices
will each have their own advantages over the other e.g. latency vs capacity.
Also, a wild idea occurred to me recently: since AI is often in the news now,
perhaps AI can be harnessed to implement a self-optimizing storage subsystem?
Many thanks to all the other Commenters here: I do enjoy this topic immensely.