Question - SSDs - Why are Ultra fast SSD no faster ?

[Andy14]

Hi guys / gals.

As I was spec'ing my new PC, one element to get a lot of attention was the the storage. My initial impulse was to buy the Samsung 960 Pro, it's the spiritual successor to the 840 Pro in my current PC and seems to offer class leading performance.

Then SCAN advised me to instead spec a "normal" SSD and either save the money or double the capacity. This lead to me to doing a fair bit of research into whether I'd appreciate any difference in speed for "normal" PC usage. Boot up, loading software, etc. Bottom line is "no". There are even videos out there proving that loading modern games on a cheap SSD is no slower than a 960 Pro in side-by-side comparisons.

So my question is not "whether it's worth it for me" but Why is it the case that the 960 Pro is no faster ? Theoretically the 960 Pro is 5-7x faster in practically every measure, but this doesn't show in the real world at all (forget database work / 4k video production - which I rarely do). What is the bottleneck that prevents the 960 Pro from reaching it's full potential here, or is it a case of the numbers being very deceiving ?

Cheers.

[shaithis]

because you are normally looking at sequential read and write speeds when looking at specs......it isn't very often that you read or write sequentially.

Random reads and writes are more down to latency, which won't be very different between models.

[Andy14]

because you are normally looking at sequential read and write speeds when looking at specs......it isn't very often that you read or write sequentially.

Random reads and writes are more down to latency, which won't be very different between models.

I'm not convinced. According to the numbers even Random Read / Writes are massively different.

960 Pro

RANDOM READ (4KB, QD32)
512 GB: Up to 330,000 IOPS (Thread 4)
1024 GB: Up to 440,000 IOPS (Thread 4)
2048 GB: Up to 440,000 IOPS (Thread 4)

RANDOM WRITE (4KB, QD32)
512 GB: Up to 330,000 IOPS (Thread 4)
1024 GB: Up to 360,000 IOPS (Thread 4)
2048 GB: Up to 360,000 IOPS (Thread 4)

850 Evo

RANDOM READ (4KB, QD32)

500 GB: Up to 98,000 IOPS
1 TB: Up to 98,000 IOPS
2 TB: Up to 98,000 IOPS

RANDOM WRITE (4KB, QD32)

500 GB: Up to 90,000 IOPS
1 TB: Up to 90,000 IOPS
2 TB: Up to 90,000 IOPS

[kalniel]

But if your computer is only calling for 5,000 IOPS then it doesn't matter if drive A can perform 300,000 IOPS to drive B's 90,000.

IE, you don't have a usage pattern for needing a faster drive.

[Andy14]

But if your computer is only calling for 5,000 IOPS then it doesn't matter if drive A can perform 300,000 IOPS to drive B's 90,000.

IE, you don't have a usage pattern for needing a faster drive.

That's still not an explanation. If we take the example of a game loading, it's trying to pull as much data as possible, as fast as possible. The loading times from HDD to Hybrid to SSD clearly show huge gains.

But there's little / no difference in speed between the fastest and slowest SSD's.

[kalniel]

That's still not an explanation. If we take the example of a game loading, it's trying to pull as much data as possible, as fast as possible. The loading times from HDD to Hybrid to SSD clearly show huge gains.

But there's little / no difference in speed between the fastest and slowest SSD's.

Game loading doesn't try and pull as much data as possible, as fast as possible. The loading times get better going from a HDD because they're trying to do more IOPS than a HDD can provide, but once you've satisfied that condition you don't get anything faster by going to a faster SSD.

Made up example: Game loading required pulling some data, decompressing it, loading it into memory, compiling some shaders. Depending on the speed of the GPU, CPU, RAM, and the bandwidth between all of those (not to mention any OS limitations), say the maximum rate the rest of the system can process is 10,000 IOS. Say our imaginary HDD has a limit of 1,000 IOPS, and our two SSDs have a limit of 90,000 and 300,000 IOPS.

Our HDD is going to be limiting the loading speed, so going to an SSD will have an impact on loading times. But the difference between the SSDs makes no change to the loading times, because now the other processes are fed as fast as they can manage and caching irons out the miniscule initial latency gain from the faster SSD.

[Andy14]

Cheers. The above explanation helps.

So basically the CPU / general memory bandwidth becomes the bottleneck as decryption needs to take place before data can be used. That makes sense and also means that the above loading time example will be faster on my new pc than the current one. SSD speed won't change, but everything else will be substantially upgraded.

[kalniel]

So basically the CPU / general memory bandwidth becomes the bottleneck as decryption needs to take place before data can be used. That makes sense and also means that the above loading time example will be faster on my new pc than the current one. SSD speed won't change, but everything else will be substantially upgraded.

Yes. Note there will always be bottlenecks - eventually you hit things like wait times deliberately written in the code/OS to prevent problems with clashes/locks etc.

[Dashers]

The question you're asking is, why isn't your professional-class device showing benefits for a domestic work-load.

One of the more expensive reasons for the "pro" products is the chips they use that can handle more writes, combined with more hidden/spare capacity. They're designed for environments where there is a lot of writes going on. Loading up a relatively small game map doesn't put any load on a SSD.

If you want to increase the speed on your domestic system, you're far better off putting cheap SSDs into a RAID-0 stripe.

[kalniel]

If you want to increase the speed on your domestic system, you're far better off putting cheap SSDs into a RAID-0 stripe.

But again that doesn't do anything for situations where a single SSD isn't particularly limiting, which in the given example of game loading, they generally aren't. If one isn't limiting, moving to two won't help

[DanceswithUnix]

But again that doesn't do anything for situations where a single SSD isn't particularly limiting, which in the given example of game loading, they generally aren't. If one isn't limiting, moving to two won't help

More than that, I would expect having one large SSD to be better than two SSDs of half the size as all the free space in on the one drive giving the controller more flexibility in moving sectors around.

There is another way of looking at this. On rotating media a disc operation consists of:

1/ Move head, make sure it is over the correct track. (really slow)
2/ Wait for sector you want to pass under the head. (quite slow)
3/ Read and return data. (fast)

On an SSD, you pretty much just go to step 3. At that point, just how fast the fast bit is starts getting academic. Now it is faster, you can measure it, but you can't *feel* the difference as human reaction time can't tell the difference between 1 millisecond and 5 milliseconds so a factor of five can get lost on us carbon units. To some people with the money to spare, just knowing they have the fastest is important. To most of us on a budget, it really doesn't matter and the money is usually better spent elsewhere like the graphics card where you can feel the difference (or more SSD capacity, as upgrading the SSD is annoying and expensive).

[Dashers]

If one isn't limiting, moving to two won't help

You can never have too much speed!

[EndlessWaves]

I'm not convinced. According to the numbers even Random Read / Writes are massively different.

960 Pro

RANDOM READ (4KB, QD32)
512 GB: Up to 330,000 IOPS (Thread 4)
1024 GB: Up to 440,000 IOPS (Thread 4)
2048 GB: Up to 440,000 IOPS (Thread 4)

RANDOM WRITE (4KB, QD32)
512 GB: Up to 330,000 IOPS (Thread 4)
1024 GB: Up to 360,000 IOPS (Thread 4)
2048 GB: Up to 360,000 IOPS (Thread 4)

850 Evo

RANDOM READ (4KB, QD32)

500 GB: Up to 98,000 IOPS
1 TB: Up to 98,000 IOPS
2 TB: Up to 98,000 IOPS

RANDOM WRITE (4KB, QD32)

500 GB: Up to 90,000 IOPS
1 TB: Up to 90,000 IOPS
2 TB: Up to 90,000 IOPS

You want to be looking at the queue depths of 1, 2 and 4 for home use, few programs are written to issue 32 simultaneous disk accesses.

Having said that, they're still faster so the bottlenecks are somewhere else. I haven't seen anyone do the tests to determine which components are being the bottlenecks though

[Andy14]

I haven't seen anyone do the tests to determine which components are being the bottlenecks though

Indeed, would love to see someone like Linus or Chris (Battle(non)sense) break this down. It's quite interesting.