How realistic are SSD benchmarks for gamers?

[BSim500]

Does anyone know of any websites that do more realistic gaming-focused testing of SSD's (level load times, stutter elimination, etc)? I know "real usage" itself is going to differ from person to person, but very few SSD "reviews" even remotely reflect my own usage. I have one 512GB SSD (soon to be 1TB) used for OS, programs & games + one secondary 3TB HDD used for data, music, video, photos, downloads, etc. After all stuff is installed, setup and configured during week 1, the bulk of day to day SSD operations from then on are typically 90-95% reads of static data with 5-10% writes being little more than OS data, game saves and web browser cache. I rarely uninstall anything and the total quantity of post-install data written is so low that on a 512GB SSD, the "average block-erase count" is still in single digits after a year. Most "bulk data" is on the HDD. So tests like...

- Copying large videos, music files, etc, back & forward are utterly irrelevant to me since they're all on the much larger 3TB HDD (to maximize precious SSD space for games)

- Synthetics that max out sequential performance do not really reflect the typically lower install speed of games which are installed from a compressed installer (sometimes single-threaded) from a backup / data drive, or for older retail games even an optical disc, where they'll typically be written slower than a pure file copy / synthetic write due to a decompression phase. Likewise, if I'm ripping / burning a CD / DVD / Blu-Ray, it'll go on the HDD as the bottleneck either way is the optical drive speed. The bottleneck for video / audio encoding is the CPU, etc.

- 30mins continuous "consistency tests" involving writing over 150-500GB without any break do not really apply to me or many others who may install one 0.1-50GB game, then there'll be a delay before installing the next one, etc. And that's assuming you put all your games on at once. Same goes for regular people of putting say Office & Photoshop Elements on a laptop, then not much else. "Consistency" tests should really be renamed "saturation" tests.

- HTPC "trace usage" that uses wildly exaggerated demands (eg, accessing 7-8 files simultaneously in a house with 2-3 people). In our house, if we record 2 things, and watch a 3rd whilst downloading a 4th, then it'll peak at 4 (normally is 1-2). But even then, 5,400rpm HDD's in various set-top boxes can handle 3 HD streams at once (record 2 + watch), and every single person we know with a custom built HTPC has a small boot SSD and either a local 2-4TB HDD, or pulls it over a NAS / media server (with a couple of 2-4GB WD Red's) where the bottleneck for multiple streams will be Gigabit network.

- Continuous QD32 "4k" only files are not realistic for anything I use. Most queue depths of mine are single digits, and although the web browser cache is the nearest thing that matches, it gets written at a far slower rate and many files are nearer 128k than 4k. You load a page, a dozen small files get written / read from cache, then there's a break whilst you read the page, etc. This is neither constant nor queued up 32x load, let alone both for 10-30mins solid.

To give one extreme example of the massive "synthetic vs observable reality" disparity, the Samsung 950 PRO is highly impressive in being 3-4x theoretically faster than most SATA bottlenecked drives, and yet is actually losing to the cheap MX200 in multiple games:-
http://techreport.com/review/29221/samsung-950-pro-512gb-ssd-reviewed/4

Now I expected the "real world" gap to be narrower, but a budget MX200 loading levels faster than a +110% more expensive 950 PRO M2 during actual gaming? And some review sites conclude how "the 512GB 950 PRO will be popular amongst gamers" (over a 1TB MX200 / Samsung EVO for less money)...

Is there some site out there that actually attempts to seriously address this stuff in detail beyond token "PC Mark 7 trace" 'realism' tests? Like a clean install of 20 games with a range of "many small files" and "few large files" installs, then measuring performance, level load times, stutter elimination in open-world games that "burst stream" in data, etc? Or benchmarking games install times from actual packed games installers (eg, a large downloaded GOG exe) rather than simple file copies or synthetic produced data? So far the Tech Report is the only one that's actually tried to measure even simple things like level load times.

Surely I can't be the only one who wants at least one review website to try something a little different beyond the usual : "7 out of 8 pages contain unrealistic workloads for any client use but we included them because..." comprising 90% of the review, only for a single page of half-hearted "real" data tucked in somewhere and treated almost like an aberration (in 1 review out of 5 that can even bother with that) that openly contradicts everything previously written, before ending with some very iffy untested conclusions for "gamers"...

 

[CiPHER]

Most review sites - including AnandTech - use Trace&Replay benchmarks which are unrealistic because they replay all I/O at maximum speed on a fresh/defragmented state and do not properly time the idle times but instead truncate them. This leads to a deviation over realistic performance figures.

Besides, the actual performance is not relevant. What is relevant is the apparent performance -- such as it is perceived by the user. This means many writes disappear in the VFS write-back buffercache and do not get noticed at all by the user. Reads - on the other hand - do get noticed by the user and realistic desktop performance is very much dependent on read performance and not write performance. Almost all benchmarks focus heavily on write performance. Also because SSDs differ mostly in write performance since their read performance is often bottlenecked by the SATA/PCIe interface and not the actual NAND backplane. Internally, simple SATA SSDs have more than 1GB/s of read backplane bandwidth available, but is truncated the interface and capped by the NAND controller.

In reality, you can have some benefit with two SATA/600 SSDs in RAID0 or a PCIe/NVMe capable M.2 SSD like the Samsung 950. But the benefit is not all that big.

You can test it yourself! You can test very easily what the impact would be between your current SSD (or HDD) and the fastest possible SSD that could EVER exist! Sounds cool huh? Well here is how you do that:

1) reboot to clear your RAM filecache; let it settle for a moment so the system is really idle;
2) start your application/game and measure the time it takes to load;
3) close the appliation, wait a moment to let the system settle and then start the application again and measure its time.

The second time you launch the application/game, it will be faster than the first. This is because this time the data is not read from SSD but from RAM memory instead. And it will always be faster than any SSD would be. Because even if your SSD does 2387493287498327498274983274732 gigabytes a second, it will still transfer it to the RAM memory and only then will the system be able to process the data. So the best possible SSD in existence with current computing technology would only be as fast as 99.99999999% as your RAM speed. Your RAM will always be faster.

So let's say the first time you start the game it is 12 seconds and the second time it is 11.4 seconds. That means 0.6 seconds is the maximum benefit between your current SSD and the fastest SSD that could potentially exist. This is not all that much a difference and as such the performance potential is very limited. A single good SATA/600 SSD is already very fast. Some benefit of higher bandwidth with two SSDs in RAID0 or a NVMe M.2 SSD is possible. But still it will be marginal.

So don't get tricked with benchmarks. Many review sites actually do their readers a bad service by implying and/or suggesting that there is a noticeable difference between SSD performance. The difference that exists most likely will fall below the threshold of Placebo and thus is hardly noticeable. Modern SSDs are very fast - newer expensive SSDs are only marginally faster.

My advice: keep your money and spend it on things that DO make a difference!

 

[dave_the_nerd]

Well... the folk wisdom is that loading levels/maps off of an SSD will benefit from the increased read speeds. It's a noticeable improvement playing WoW/SWTOR/Skyrim/Etc. But there's no real increase in FPS.

So adding an SSD to a Pure Gaming Rig is probably the lowest priority in terms of bottlenecks and spending money to increase performance. I don't know of any review sites that have gone to the specific trouble of benchmarking SSDs vs. HDDs for games.

 

[Magic Carpet]

Great explanation CiPHER :thumbsup:

 

[BSim500]

Most review sites - including AnandTech - use Trace&Replay benchmarks which are unrealistic because they replay all I/O at maximum speed on a fresh/defragmented state and do not properly time the idle times but instead truncate them. This leads to a deviation over realistic performance figures.

So these "realistic trace tests" consist of gathering highly accurate I/O data over an 8hr workday, and then replaying it back in say "24x Fast Forward" over 20mins? Presumably that then massively skews the ratio of benchmarked I/O activity imposed on the SSD vs the actual real-time I/O load requested by the components (eg, CPU, TRIM / SSD garbage collection, cache, etc)? I suppose that also inflates away other components limitations (eg, streaming 8x 30mb/s 2hr Blu-Ray rips over a LAN back to back over 8hrs = only 2x 30mb/s load at any one time (2hr-2hr-2hr-2hr), but "time compressed" by 24:1 (squeezing 8x movies into 20mins) would in fact require streaming all 8x together each at a 6x faster rate that would exceed Gigabit Ethernet's bandwidth (and is therefore something that no one would do in reality)? Crude example (but not far off Storage Review's gibberish 1175MB/s HTPC "traces" of squeezing 12hrs viewing into 15mins), but that's... utterly absurd if that level of "up to 48x time compression" cheating is the industry standard for recommending SSD purchases. D:

I take it that's also why "game traces" of the 950 PRO vs 850 EVO look like this...
http://www.storagereview.com/images/samsung_ssd_950_pro_m2_nvme_512gb_storagemark2010_gaming.png

...whilst actual real-time, real-life game charts of same drives look like this..
http://techreport.com/r.x/2015_10_21_950_Pro/tombraider.png

...is because the "game trace" in the first chart is "running" 20, 30, 40 times faster than normal (I/O access) purely to deliberately cripple the slower drives (even though the game may run well off a mechanical HDD) and artificially exaggerate the real-time gap?

I guess I just can't believe how large / mature the SSD market has become, and yet all we've got are these incredibly fake tests. It's strangely reminiscent of older Tom's Hardware Guide benchmarking every game at 640x480 to "prove" one CPU is 75% faster than another. Sure you shouldn't defeat the purpose of CPU benchmarks by benchmarking games at a premium 4K gaming intentionally leaving all CPU's the same via a GPU bottleneck, but neither should you wildly over-swing to the other extreme and create fake "usage" scenarios 48:1 times out of whack vs how it would perform as part of a real-time system (which seems to be the norm in SSD land).

Thanks for the reply. Good post and explanation, BTW. :thumbsup: Makes a lot more sense than what most review sites are trying to "demonstrate".

 

[CiPHER]

Well... the folk wisdom is that loading levels/maps off of an SSD will benefit from the increased read speeds. It's a noticeable improvement playing WoW/SWTOR/Skyrim/Etc. But there's no real increase in FPS.

Well some games - including WoW - read textures during gameplay such as when a PVP player enters your world. This can lead to FPS drops that are quite noticeable. This happened to me when i did some gaming in the past with WoW in 'PVP Wintergrasp' area which is a massive PVP event.

Because these games are frequently updated, a lot of fragmentation occurs. On a harddrive this is bad because when reading this information the harddrive has to seek a lot, causing high read latency. This results in frequent small FPS drops or 'spikes' in the game, but only during specific circumstances such as massive PVP events.

It is possible the current state of the game has improved. But in general online MMO games may suffer from this issue. So some games may have an impact on FPS, but a fast SSD will never raise the maximum FPS as you correctly implied.

So adding an SSD to a Pure Gaming Rig is probably the lowest priority in terms of bottlenecks and spending money to increase performance.

I don't game much at all today, but when i do i just want to do it for a short while. So i like the games to be on SSD so they boot quickly, and allow me to resume my work quickly again. So an SSD might in fact be very useful to some gamers. If you are the type that games for multiple hours, i would say a better graphics card and/or CPU is on top of your list. But with SSDs getting bigger, storing the games on the SSD is a very viable option.

I don't know of any review sites that have gone to the specific trouble of benchmarking SSDs vs. HDDs for games.

Even if they do, they will not mimic the realistic usage patterns of gamers. For example, one key aspect is that many games are updated frequently with a 'loader' application that downloads the updates and then modifies the huge data files causing fragmentation.

Benchmarkers usually capture I/O in pristine conditions: a fresh installed operating system and a fresh installed application. This will lead to an unrealistic focus on sequential I/O whereas realistic usage will see more random I/O due to fragmentation instead.

Benchmarking is quite an art i would say. I have never seen a true good storage benchmark in all my life. Doing it properly requires a good amount of effort and probably writing custom tools. The easier way is to use the tools that are available and using pristine conditions so at least the test is performed equally when repeating the test on other SSDs. Most - probably all - review sites think that is enough.

 

[CiPHER]

So these "realistic trace tests" consist of gathering highly accurate I/O data over an 8hr workday, and then replaying it back in say "24x Fast Forward" over 20mins?

Basically, yes. But remember that they capture the I/O on a pretty 'pristine' system which does not tally with real-life either way. The best way would be to capture the I/O patterns of an actual user system over multiple months time. Obviously that is much more time-consuming.

Presumably that then massively skews the ratio of benchmarked I/O activity imposed on the SSD vs the actual real-time I/O load requested by the components (eg, CPU, TRIM / SSD garbage collection, cache, etc)?

Yes. The defence of most reviewsites - including AnandTech - is that this would not be a big issue because the performance impact would apply equally to all SSDs tested. Such as the argument that bad weather would impact both soccer teams and thus would be 'fair'. But the reality is that one soccer team might be impacted more heavily by bad weather than the other. The same is true for the lack of idle time during trace&replay benchmarks; one SSD might rely on background Garbage Collection more than the other, leading to a higher performance deviation versus realistic circumstances.

Additionally, many benchmarks are performed by writing the entire LBA surface of the SSD - meaning they are written 100% full. The amount of overprovisioning varies considerably between SSDs because some SSDs heavily rely on the fact of there being free space available through TRIM. This also would lead to performance deviation that impacts some SSDs much more than others, something that does not happen in realistic circumstances because almost all SSD users have free space on their SSDs and have TRIM capability.

I take it that's also why "game traces" of the 950 PRO vs 850 EVO look like this...
http://www.storagereview.com/images/samsung_ssd_950_pro_m2_nvme_512gb_storagemark2010_gaming.png

...whilst actual real-time, real-life game charts of same drives look like this..
http://techreport.com/r.x/2015_10_21_950_Pro/tombraider.png

Ha ha ha

That is an awesome comparison! :thumbsup:

I guess I just can't believe how large / mature the SSD market has become, and yet all we've got are these incredibly fake tests.

It has never been different in my years of experience. I remember AnandTech and StorageReview posting a joint statement that RAID0 had no place on the desktop. While such an assertion may or may not have good arguments, they backed up their claims with one of the worst storage benchmarks i have ever witnessed. They did 'everything wrong' by crippling the RAID0 configuration to the point that it could not possibly be worse. Using FakeRAID drivers, using PCI addon card, using migaligned 31.5KiB partitions and using too low stripesize. Ugh! If you are going to test RAID0 viability, do it proper!

To be fair, this was in a time where some of the culprits were not all that well known. But i have seen mistakes like this happen quite frequently. For example that Anand himself was unaware of his Sandforce SSD benchmarks being fake because Sandforce was using compression and deduplication techniques. He was testing zero write performance and not realistic performance. Today this is no more an issue since the capabilities of this specific SSD controller are well known and tools like IOmeter have been patched to use random data instead of static pseudo-random data that Sandforce can defeat by using deduplication.

It's strangely reminiscent of older Tom's Hardware Guide benchmarking every game at 640x480 to "prove" one CPU is 75% faster than another.

Such benchmarks may be viable by proving that the bottlenecks shift from CPU to GPU when increasing resolution and/or detail settings. But i would always start with a generic 'realistic' benchmark and only after that start with using specific benchmarks which must be accurately interpreted to make sense.

Benchmarking has more goals just to give you a realistic indication of real-world performance. Micro-benchmarks for example are used to improve performance of very specific software code. How that translates to actual performance gains considering all factors is not always considered.

Thanks for the reply. Good post and explanation, BTW. :thumbsup: Makes a lot more sense than what most review sites are trying to "demonstrate".

Yes. Keep in mind though that no really good storage benchmark has ever been performed to my knowledge. Also not by myself. I have investigated in the past what it takes to do it Properly® but i quickly realised it would take a shitload of time and effort and without any pay that is not very motivating.

Yet, somehow i expected more from the 'big guys' like AnandTech and StorageReview and TheTechReport. They may have the funds to develop their own tools and do unsurpassed benchmarks that take all the right factors into account. Thus far, i have not been very impressed. All those benchmarks they do seem to accentuate the minuscule differences between SSDs and imply that buying a faster SSD actually makes a difference. The reality, however, is that such hopes are in vain. SSDs perform almost equally but do differentiate themselves in terms of price and reliability.

 

[Deders]

Well... the folk wisdom is that loading levels/maps off of an SSD will benefit from the increased read speeds. It's a noticeable improvement playing WoW/SWTOR/Skyrim/Etc. But there's no real increase in FPS.

So adding an SSD to a Pure Gaming Rig is probably the lowest priority in terms of bottlenecks and spending money to increase performance. I don't know of any review sites that have gone to the specific trouble of benchmarking SSDs vs. HDDs for games.

I found Arkham City, Watchdogs and Arkham Knight ran better on an SSD. The Arkham games ran on a modified Unreal 3 engine which was made with small levels in mind, so a lot of effort was put in to stitch lots of smaller maps together seamlessly, in a way that became quite demanding for disk IO and resulted in stutters on HDD's.

The SSD has also been a godsend for level loading times in games like Dragon Age: Inquisition which as unbearable for me a HDD.

Another factor to consider is Hard Page Faults which will vary from system to system. If Windows can't find the data it needs in memory, this causes a hard page fault and Windows will have to look elsewhere. Usually the page file, but sometimes the original location of the file. Using a HDD will mean that games are more likely to stutter more often as Windows will pause everything until the data is found.

Another reason that disabling the Page file with an SSD is counter productive.

 

[dave_the_nerd]

I found Arkham City, Watchdogs and Arkham Knight ran better on an SSD. The Arkham games ran on a modified Unreal 3 engine which was made with small levels in mind, so a lot of effort was put in to stitch lots of smaller maps together seamlessly, in a way that became quite demanding for disk IO and resulted in stutters on HDD's.

The SSD has also been a godsend for level loading times in games like Dragon Age: Inquisition which as unbearable for me a HDD.

Another factor to consider is Hard Page Faults which will vary from system to system. If Windows can't find the data it needs in memory, this causes a hard page fault and Windows will have to look elsewhere. Usually the page file, but sometimes the original location of the file. Using a HDD will mean that games are more likely to stutter more often as Windows will pause everything until the data is found.

Another reason that disabling the Page file with an SSD is counter productive.

If you're running out of ram and having to hit disk, you need more ram. SSDs just make a horrible situation slightly less agonizing.

 

[ShintaiDK]

The big change was HD to SSD, with speed changes in the area of 100-1000x where it mattered most (seektime).

SSD to SSD is quite minimal. Even something like a cheap SSD vs a 3DXpoint NVME it may be 10-20x. Sequel read/write is of little benefit. Most user I/O is random.

I also think one of the issues behind is that OS, games etc essentially is still developed around the lowest denominator. And that is HD speed/seektime.

 

[Deders]

If you're running out of ram and having to hit disk, you need more ram. SSDs just make a horrible situation slightly less agonizing.

Don't want to derail the thread, but this is on the assumption that the pagefile is only used for excess memory usage. You are still likely to get page faults without a pagefile, and if they can't be resolved from an SSD then it will take longer to get the data from an HDD and Windows will pause until it is resolved.

 

[BFG10K]

Does anyone know of any websites that do more realistic gaming-focused testing of SSD's (level load times, stutter elimination, etc)?

I did a round of load time tests a while ago:

The reality is even with the cheapest SATA SSD you're no longer I/O bound in typical desktop situations except in edge cases such as copying files all day. Heck, even the VelociRaptor can hold its own against SSDs for level load times.

The usual rebuttal I get is "you used a slow Intel SSD", to which my reply is "no SSD is faster than a RAM disk", which is the last column in my tests. I pat myself on the back for being forward-thinking enough to perform RAM disk tests.

It's always funny to see the NVMe crowd claiming their devices instill some kind of massive performance gain over SATA SSDs. But then this is the same crowd that claims they can "feel" the difference between 1T and 2T RAM timings.

I liken them as the audiophiles of the storage world. Just don't tell them a coat hanger is as good as a Monster Cable. :awe:

 

[Sabrewings]

I found Arkham City, Watchdogs and Arkham Knight ran better on an SSD. The Arkham games ran on a modified Unreal 3 engine which was made with small levels in mind, so a lot of effort was put in to stitch lots of smaller maps together seamlessly, in a way that became quite demanding for disk IO and resulted in stutters on HDD's.

The SSD has also been a godsend for level loading times in games like Dragon Age: Inquisition which as unbearable for me a HDD.

Some games I play load all of their assets (up to 4GB at times) into RAM when they launch. These games benefit hugely from running off an SSD during startup. It's the difference from a few minutes down to 20 seconds or less. A few examples right now are KSP and Rogue System.

 

[Madpacket]

My first SSD was an Intel 80GB MLC drive (still works) and have owned well over 20 SSD's since then of various brands with different controllers. I currently run a 128K formatted 960GB stripe (dual 480GB Toshiba SSD's) for my latest gaming rig.


There's been virtually no perceivable difference in game loading times out of all of these drives and setups. SATA 2 still isn't a bottleneck for the most part.

The only real world benefit is when I'm installing a large number of Steam or Origin games. These applications like to precache or allocate disk space to make room for the game installs. I've noticed a fairly significant decrease in install times with my Raid stripe over single drive SSDs. Other areas that help is when verifying the game integrity of huge installs (multiple CRC) but beyond these cases there's not much improvement.

Mechanical hard drives I only use for large file storage and video playback. Loading games like BF4 with large sets of small asset data off a mechanical HDD is excruciating compared to an SSD.

What it boils down to is marketing and companies trying to come up with reasons to sell faster drives. These new NVMe drives are not worth the asking price for a gaming desktop. They have their place in the server / VM world though.

New SSD's are kind of like new Intel processor's. Not very exciting...

 

[Deders]

When I had my 830 on Sata 2, I would notice that disk reads would on very rare occasion peak at the 260MB/s limit when loading games, but it would only be for a second or so, resulting in possibly an extra second in loading time. That was with my old i5-750@3800MHz.

I then upgraded to my current system with Sata 3, was more interested in the CPU improvement to look too much into disk speeds, but if the disk isn't the bottleneck, then the processing of that data is. For instance with Dragon age: Inquisition, with a HDD the CPU activity would be low as it is waiting for the disk. With an SSD it was much higher.

Can't remember if it was 100% on one core or all 4 but when you consider that my current Skylake in some circumstances benchmarks twice as fast as my old i5 did, it can deal with the data at much a faster rate. Obviously there is also work done on the GPU to build a 3D level which will also bottleneck the loading time.

What I have found with monitoring my 950 pro disk speeds when loading a few games is that it can breach the 550MB/s Sata 3 limit. I've seen 633 and once just over 750, again only lasting for one 'reading' that updates every second. On my old system that would have taken 3x as long to load that (albeit small) segment.

Whilst running the GTAV benchmark, I noticed that at one point it was reading over 100MB/s (not continuously) whist flying low, iirc before it flew under the bridge.

It's much harder to quantify reads that can't make full use of the raid nature of SSD's, (4k random reads for instance), when just looking at MB/s. Currently my 950 is capable of one-and-a-half times as many 4k reads as my old i5/830 combo (50MB/s vs 20MB/s) That has to count for something somewhere.

 

[VirtualLarry]

There's been virtually no perceivable difference in game loading times out of all of these drives and setups. SATA 2 still isn't a bottleneck for the most part.

Currently my 950 is capable of one-and-a-half times as many 4k reads as my old i5/830 combo (50MB/s vs 20MB/s) That has to count for something somewhere.

Both good points. I do think that the 4K Random Read (QD=1) numbers are probably the most real-world related benchmark scores. Most SATA6G SSDs, regardless of controller or NAND type, generally peak at something around 20-25MB/sec there. So a doubling of that number, would seem like it might have some sort of real-world noticeable effect.

 

[JimmiG]

Well... the folk wisdom is that loading levels/maps off of an SSD will benefit from the increased read speeds. It's a noticeable improvement playing WoW/SWTOR/Skyrim/Etc. But there's no real increase in FPS.

So adding an SSD to a Pure Gaming Rig is probably the lowest priority in terms of bottlenecks and spending money to increase performance. I don't know of any review sites that have gone to the specific trouble of benchmarking SSDs vs. HDDs for games.

It depends on the game. Many games pre-load the entire level into RAM. Also the loading is quite CPU-intensive as it's usually also de-compressing the data and compositing the level at the same time. The loading is often very serial in nature (optimized for HDD's with high seek times), so the HDD might not be the bottleneck at all. The data is often packed into a few large multi-gigabyte files, and HDD's are quite fast at sequential reads. That's why in some games you might see the loading time go from something like 35 seconds to 33 seconds with an SSD. The exception is, as you say, open-world games since they continually stream data from the drive. So you might experience fewer hickups when entering a new area in such a game.

 

[Magic Carpet]

The usual rebuttal I get is "you used a slow Intel SSD", to which my reply is "no SSD is faster than a RAM disk", which is the last column in my tests. I pat myself on the back for being forward-thinking enough to perform RAM disk tests.

Now that modern games can easily outdo the storage available by typical ram disks these sort of benches would represent quite a challenge, if you were to re-do it

 

[StrangerGuy]

I did a round of load time tests a while ago:

The reality is even with the cheapest SATA SSD you're no longer I/O bound in typical desktop situations except in edge cases such as copying files all day. Heck, even the VelociRaptor can hold its own against SSDs for level load times.

The usual rebuttal I get is "you used a slow Intel SSD", to which my reply is "no SSD is faster than a RAM disk", which is the last column in my tests. I pat myself on the back for being forward-thinking enough to perform RAM disk tests.

It's always funny to see the NVMe crowd claiming their devices instill some kind of massive performance gain over SATA SSDs. But then this is the same crowd that claims they can "feel" the difference between 1T and 2T RAM timings.

I liken them as the audiophiles of the storage world. Just don't tell them a coat hanger is as good as a Monster Cable. :awe:

I think those as "Hi, I'm superior because I can spend more money than you and I also have a massive case of buyer's remorse that I'm not going to admit."

It's much harder to quantify reads that can't make full use of the raid nature of SSD's, (4k random reads for instance), when just looking at MB/s. Currently my 950 is capable of one-and-a-half times as many 4k reads as my old i5/830 combo (50MB/s vs 20MB/s) That has to count for something somewhere.

My iPhone 6S has 5x the ST performance of my lame Snapdragon 400 equipped Xiaomi phone yet the latter still feels just as smooth in general use. Just because there's a big difference in raw specs doesn't mean it actually does anything in the real world.

 

[Deders]

My iPhone 6S has 5x the ST performance of my lame Snapdragon 400 equipped Xiaomi phone yet the latter still feels just as smooth in general use. Just because there's a big difference in raw specs doesn't mean it actually does anything in the real world.

I presume they run very different operating systems and the iPhone is capable of doing a lot more?

I get your point, bigger numbers don't always make for better performance if the software can't make use of it, but your analagy isn't quite apple to apples.

 

[BSim500]

I liken them as the audiophiles of the storage world. Just don't tell them a coat hanger is as good as a Monster Cable. :awe:

LOL! Ironically I used to work part-time in an audiophile shop for a while (I eventually left out of honesty). I showed one guy a cheap "deaf peasant" £3 interconnect cable then plugged it in (he couldn't see the back of the unit) and watched him sneer. Then I held up a premium "golden ear" £99.99 "Monster" cable and... plugged the same £3 cable back in again. "Oooh that's MUCH better" he squealed. Then there was another guy who wanted a 192khz DVD-Audio system yet couldn't hear a brick-wall low-pass filter at 14khz. And my personal favorite - cable risers (which stop the bits falling out the cable and making a mess on your carpet)... And I know exactly which "telephone wire & coathanger" test you're referring to. A lot of hurt ego's on that day...

Thanks all for your replies. I went with a 1TB MX200 in the end that was over £100 cheaper than a Samsung Pro 850 (yet still MLC tech and same real-world game load times). :thumbsup: I'm still very wary of TLC. From what I read, data retention was 40 weeks on Samsung's 21nm TLC (840) and down to 8 weeks on 19nm TLC (840 EVO) due to voltage overlap / degradation causing slowdowns. In contrast, a read speed test on my Crucial MX100 (16nm MLC) showed a solid +500-540MB/s on 16 month old data. Samsung's 40nm V-NAND 850 EVO may have solved it, but I still hope SSD manufacturers don't push for a race to the bottom and go all TLC on non-premium drives in future (like the BX100 to BX200 move).

Edit: BFG10K, I'm not sure if it's just me but the chart you posted earlier is just a thumbnail? Do you still have the full size chart? I can't say I've really looked at RAMDisk testing. Thanks.

 

[VirtualLarry]

but I still hope SSD manufacturers don't push for a race to the bottom and go all TLC on non-premium drives in future (like the BX100 to BX200 move).

The takeaway I got from AT's web site article / review on the BX200, was that TLC (non V-NAND, at least) was CRAP. Slow write times, bad access times, basically not much better than a HDD. Sure, cheap, but not worth it.

 

[flexy]

Well... the folk wisdom is that loading levels/maps off of an SSD will benefit from the increased read speeds. It's a noticeable improvement playing WoW/SWTOR/Skyrim/Etc. But there's no real increase in FPS.

So adding an SSD to a Pure Gaming Rig is probably the lowest priority in terms of bottlenecks and spending money to increase performance. I don't know of any review sites that have gone to the specific trouble of benchmarking SSDs vs. HDDs for games.

A lot of games DO access data frequently (textures) etc. and also the waiting times for loading levels can be significantly lower with SSD. (BF4 maps etc.)

No one halfway technically savvy would expect a direct "FPS increase" from an SSD although it can help to eliminate stutter..and thus indeed have an (indirect) effect on FPS.

A good example here was Diablo 3 where (as with many games) there is constantly something loaded...in extreme cases, say if you run from a mediocre HD there is indeed a short hitch/pause before an animation (when the animation is loaded/pre-loaded). This I consider SIGNIFICANT. Imagine a small hitch/stutter many/most times before an animation. I remember sometimes this was an entire second...a literal mini-freeze of the game due to some HD access.

(Of course this also depends on how well a game does pre-loading, caching etc.)

This, the faster level load times, less waiting etc.. and less hitching I do actually consider the most "real life benefit" and I personally would rate an SSD way up on the list for an overall smoother experience.

 

[UaVaj]

If you're running out of ram and having to hit disk, you need more ram. SSDs just make a horrible situation slightly less agonizing.

exception: does not help with stupid games like BF4 which refuse to use physical memory (when there is ample) and want to revert back to page file on ssd.



back to the original ssd question. once you get to a certain level of ssd performance (~300MB/s). any higher performance become diminishing.

weakest performance link is the 4k read. even a 950 pro is still marginal. that is where performance needs to be focus.

 

[BFG10K]

Edit: BFG10K, I'm not sure if it's just me but the chart you posted earlier is just a thumbnail? Do you still have the full size chart? I can't say I've really looked at RAMDisk testing. Thanks.

Definitely not a thumbnail, it's 433 x 932 pixels and very easy to see.