Intel Optane cache drives don't play with budget Kaby Lake CPUs

[fail_quail]

2. I see no reason for gamers to have SSD for games - i had that but i realised that having my steam library on HDD and os on SSD almost the same - no kidding. HOTS game is loading 15 sec instead of 12 sec.in both cases i usually wait for other players. It is similar for other games, and there is no difference in the runtime of the game.

Depends on the game. It's not just loading times, particuarly games that utilise texture streaming (mostly open world games, GTAV, fallout4, skyrim and so on) benefit immensly from a SSD, greatly reducing texture pop-in and/or stuttering during gameplay.

Most of my games are on a mechanical HDD, but a select few i run from a dedicated 128GB SSD.

[aidanjt]

The biggest advantage of SSD's over HDD's: Not having to wait for the platters to spin up.

[jigger]

Seems Intel have also seen fit to drop support/intentionally stop optane from working with greenlow systems...

[DanceswithUnix]

Only that's not what we see with Optane as it has around the same iOPS as SSD's and NVMe drives, however most SSD's and NVMe drives don't reach their maximum iOPS until they hit maximum QD whereas Optane reaches it's maximum iOPS at much lower QD, they say pictures tell a thousand words so..

Actually I think that graph explains very little

Workloads have an IOPS rating. To fit in with that graph, let's say I am have a VM farm which requires about 40000 IOPS.

The SATA SSD will cope and have a queue depth of around 8.5
The NVMe drive will soak up the same load with a queue depth of 3
The Optane drive will laugh at such a feeble workload, with a queue depth of 1
Spinning rust, well that won't cope. The OS will have a limit to the number of operations it allows on the queue, and leave the queue permanently full.

So hence my observation, you can measure but not generally feel a difference in queue depth between 8 and 3. Between three and one, no chance.

It's not about throughput, it's about accessing lots of small amounts of data really quickly.

Sorry, bad wording on my part. I was thinking about throughput during small random reads, not about the big linear reads that consumer SSDs give as ratings. I have spent the last few years in enterprise storage, large linear reads don't really happen there so the thought just doesn't occur to me! Storage is usually sold as so many petabytes at so many IOPS, and the game is keeping up with all those little reads from hundreds of clients!

[Corky34]

Actually I think that graph explains very little

That's because you're not reading it right, look turn it around...no, not that way, like this...that's it, now fold it like this....you see now it looks like she's smiling.

Workloads have an IOPS rating. To fit in with that graph, let's say I am have a VM farm which requires about 40000 IOPS.

The SATA SSD will cope and have a queue depth of around 8.5
The NVMe drive will soak up the same load with a queue depth of 3
The Optane drive will laugh at such a feeble workload, with a queue depth of 1
Spinning rust, well that won't cope. The OS will have a limit to the number of operations it allows on the queue, and leave the queue permanently full.

Yes workloads issue certain iOPS but i think you maybe mistaking where these queues exist and how they work, the OS doesn't limit its requests and it's not aware of what the storage controller is doing, it send a I/O operation then waits for the reply just like all it does with all I/O operations.

The storage controller then services those requests as quickly as it can and if more I/O operations come in than it can responded to they start backing up, the queues are there simply to allow the controller to organise those I/O operations efficiently during times that the I/O operations outstrip the drives ability to services them.

Using your example the SATA SSD would have a latency of 8.5ns, the NVMe drive 3ns, Optane 1ns, and the HDD would have 32ns (I'm just using those numbers as an example, they're probably very different than that)

So hence my observation, you can measure but not generally feel a difference in queue depth between 8 and 3. Between three and one, no chance.

If we feel it is rather subjective that's why we measure things, some people started to notice stuttering in games because of nanosecond delays between frames while others didn't, it wasn't until we started measuring frame delay that it was discovered these delays were a problem, so while you or i may not 'feel' the difference others will and I'd suggest that even if you don't think you or I can't 'feel' the difference that we do.

[CAT-THE-FIFTH]

2. I see no reason for gamers to have SSD for games - i had that but i realised that having my steam library on HDD and os on SSD almost the same - no kidding. HOTS game is loading 15 sec instead of 12 sec.in both cases i usually wait for other players. It is similar for other games, and there is no difference in the runtime of the game.

Some games it isn't that important but with others it is. Fallout 4 is noticeably quicker loading from an SSD,and with mods even more so. My Sandisk 480GB had the common "suddenly going kaput" syndrome and I reinstalled my modded Fallout 4 game to a 7200RPM HDD. Not only were load times much longer,at some points the game literally stalls until the texture and assets are loaded and I confirmed it looking at the performance manager - the HDD was at 100% utilisation.

[DanceswithUnix]

Yes workloads issue certain iOPS but i think you maybe mistaking where these queues exist and how they work,

Well I hope I'm not too far out, I have an interview coming up to write firmware for SSD drives so I hope I can get the basic block IO stack right

I just twigged, the queue depth in that graph goes up to 32 and that is the NCQ maximum command depth isn't it. Silly me, should have spotted that, so that is what they are measuring.

I still contend that in domestic and most business use you won't feel any difference.

Put it this way, 1000 IOPS is lot for a real world task, yet from the graph you referenced the SATA SSD can manage 20000 with a queue of 1. So it is within the capabilities of the cheapest SSD to keep up with strenuous real world tasks without so much as pipelining the next command. How much overkill do you need?

[aidanjt]

How much overkill do you need?

There's no kill quite like overkill!

I agree, though, NVMe is ridiculously overhyped in tech press. It's definitely got its applications, but consumer computing certainly isn't it.

[Corky34]

How much overkill do you need?

You're asking that on a tech enthusiast site like Hexus where some people spend thousands on hardware that gives them an extra 1-2% performance.

Maybe we should all still be using DDR as DDR2, 3, 4, 5 are just overkill.

[aidanjt]

Maybe we should all still be using DDR as DDR2, 3, 4, 5 are just overkill.

Bit of an apples to oranges, there. DDR4 comes in higher densities and lower prices for the same capacity. Optane and NVMe SSDs are significantly more expensive for the same capacity over SATA SSD's without any real world benefit. Being able to fit more games on an SSD is much more important than them starting fewer of them 0.01s sooner.

[CAT-THE-FIFTH]

Bit of an apples to oranges, there. DDR4 comes in higher densities and lower prices for the same capacity. Optane and NVMe SSDs are significantly more expensive for the same capacity over SATA SSD's without any real world benefit. Being able to fit more games on an SSD is much more important than them starting fewer of them 0.01s sooner.

I think I would also rather have more reliability too.

[Corky34]

Bit of an apples to oranges, there. DDR4 comes in higher densities and lower prices for the same capacity. Optane and NVMe SSDs are significantly more expensive for the same capacity over SATA SSD's without any real world benefit. Being able to fit more games on an SSD is much more important than them starting fewer of them 0.01s sooner.

No one was comparing DDR to anything, it was used as an example because DanceswithUnix asked how much overkill do you need, and that IMO is like saying why bother with buying a faster car, a bigger TV, faster RAM, or anything that's better than what you currently have.

In other words if we take the how much overkill do you need to its logical conclusion we'd still be using HDD because for most people, most of the time, an SSD is overkill.

Also like i said the "real world" benefits are subjective that's why we measure things, it's also not about games starting 0.01s sooner, it's about the drive being able to service more I/O operations per second, if one CPU had an IPC of 100 and another 1000 would you buy the slower one?