this whole news sounds like marketing gibberish ;)

Yeah, not really impressive, especially as there are no 4K Q1T1 numbers. I think Adata is going to have to do some more firmware work on this drive before they release it.

For the 4K Q64T8 and Q32T8 results, my "old" Patriot Viper VPN100 has better read performance and it even beats it in the Q16T8 test. Write speeds are also only slightly faster. It really shows that we need a wider controller/NAND bus to gain any useful performance advantage over current NVMe drives. Alternatively we need a new type of NAND.



Here's my OS drive, an even older WD Black SN700 which also beats it in all of the read tests, by quite some margin.

TheLostSwedeYeah, not really impressive, especially as there are no 4K Q1T1 numbers. I think Adata is going to have to do some more firmware work on this drive before they release it.

For the 4K Q64T8 and Q32T8 results, my "old" Patriot Viper VPN100 has better read performance and it even beats it in the Q16T8 test. Write speeds are also only slightly faster. It really shows that we need a wider controller/NAND bus to gain any useful performance advantage over current NVMe drives. Alternatively we need a new type of NAND.



Here's my OS drive, an even older WD Black SN700 which also beats it in all of the read tests, by quite some margin.

You're 100% correct, and a storage nerd myself, I really dislike how they never put 4k random r/w's as we know that's what determines the user experience of you're going to use it as your OS drive. It reminds me of how WiFi routers are marketed, with a number that's the sum of all speeds on all channels and that is absolutely nothing indicative of the transfer rates a single user/connection will experience.

TheLostSwedeYeah, not really impressive, especially as there are no 4K Q1T1 numbers. I think Adata is going to have to do some more firmware work on this drive before they release it.

For the 4K Q64T8 and Q32T8 results, my "old" Patriot Viper VPN100 has better read performance and it even beats it in the Q16T8 test. Write speeds are also only slightly faster. It really shows that we need a wider controller/NAND bus to gain any useful performance advantage over current NVMe drives. Alternatively we need a new type of NAND.



Here's my OS drive, an even older WD Black SN700 which also beats it in all of the read tests, by quite some margin.

AnarchoPrimitivYou're 100% correct, and a storage nerd myself, I really dislike how they never put 4k random r/w's as we know that's what determines the user experience of you're going to use it as your OS drive. It reminds me of how WiFi routers are marketed, with a number that's the sum of all speeds on all channels and that is absolutely nothing indicative of the transfer rates a single user/connection will experience.

Why are you guys continually surprised that press releases, which are always pure marketing, only ever showcase the best attributes of the product being released?

At least capacity starts at 1TB, which implies 512GB SSDs are one the way out.

Nice lookin heat sink, which will obviously will rule it out for lappys, but as usual, we gotta go diggin somewhere else to find the price :(

bonehead123Nice lookin heat sink, which will obviously will rule it out for lappys, but as usual, we gotta go diggin somewhere else to find the price :(

PCIe 4.0 is not for laptops anyhow, as there are no PCIe 4.0 enabled CPUs or chipsets for laptops, unless you're talking about the luggables with a desktop solution inside them.

TheLostSwedePCIe 4.0 is not for laptops anyhow, as there are no PCIe 4.0 enabled CPUs or chipsets for laptops, unless you're talking about the luggables with a desktop solution inside them.

Tiger Lake laptops are supposedly launching this year and will support PCIe 4.0 NVMe SSDs, ADATA already has them covered: www.anandtech.com/show/16094/adata-xpg-launches-a-pcie-40-x4-nvme-ssd-for-notebooks-gammix-s50-lite

AssimilatorTiger Lake laptops are supposedly launching this year and will support PCIe 4.0 NVMe SSDs, ADATA already has them covered: www.anandtech.com/show/16094/adata-xpg-launches-a-pcie-40-x4-nvme-ssd-for-notebooks-gammix-s50-lite

Doesn't make sense from a power perspective though, as that's not going to be a power friendly SSD.
Something along the lines of this makes more sense for laptops.
www.anandtech.com/show/16012/the-sk-hynix-gold-p31-ssd-review

TheLostSwedeYeah, not really impressive, especially as there are no 4K Q1T1 numbers. I think Adata is going to have to do some more firmware work on this drive before they release it.

This is still using old B27A or B27B flash - 96L Micron. We won't see E18 with Micron's new RG flash until 2021. B27A/B27B is 512Gb/die with four planes, in many respects just an updated B16A/B17A design to allow for higher sequentials. No real 4K improvements with B27A (which the leaked E18 from Ramseyer utilized, I asked), B27B has appeared in one ADATA product but even he says it might be 2021 for that (and later in 2021 for B47x rollout). If you calculate out the seq. writes for 96-way interleaving - which can be achieved on older controllers, but they were limited to 800 MT/s per channel - you see they can only meet the specs at 2TB. tl;dr you're right, and these early E18s are really only for sequentials and really only at 2TB.

AnarchoPrimitivYou're 100% correct, and a storage nerd myself, I really dislike how they never put 4k random r/w's as we know that's what determines the user experience of you're going to use it as your OS drive. It reminds me of how WiFi routers are marketed, with a number that's the sum of all speeds on all channels and that is absolutely nothing indicative of the transfer rates a single user/connection will experience.

It's unfortunately more complicated than that. I'm lucky to now be on the forefront of flash analysis but it's moving rapidly to the point it will be very difficult for reviewers moving forward. For example, we have SLC caching algorithms that not only rely on behavioral (workload) profiles but can shuffle data into SLC for reads and utilize AI/ML for ECC per zones (e.g. static vs. dynamic SLC, with the latter shifting through native flash). We have adaptive read schemes (Samsung's ARC), tiles, SBL, etc. to improve tR for 4K requests. Samsung even uses 8KB subplanes on their new flash - see the 980 Pro. They also use a new TurboWrite. Gauging "user experience" is going to get very difficult with these new drives, relying on 4K random no longer cuts it as you have all sorts of tricks, as another example Kioxia has an adaptive scheme for programming BiCS5 that adjusts for detected tR needs. New QLC from Kioxia and Intel have parallel/independent plane reads which can mask QLC's poor tR (in addition to them "hiding" OS files in SLC cache for reads), etc.

AssimilatorWhy are you guys continually surprised that press releases, which are always pure marketing, only ever showcase the best attributes of the product being released?

At least capacity starts at 1TB, which implies 512GB SSDs are one the way out.

They're not - it's just that E18s have to use older, 96L flash for now which requires more flash to saturate the channels. This is why the Samsung 980 Pro is using 256Gb dies at 250GB-1TB (likely 512Gb at 2TB) despite launching as 512Gb, SK hynix is using 512Gb dies in the Gold P31 despite launching that flash at 1Tb (likely to be used in the 2TB Platinum P31), etc. At 500GB you have only one 64GiB die per channel (technically 256Gb dies are possible as on lower capacity E16s, though) and 96L flash is too slow. Upcoming flash will be much faster per die, from 82 MB/s (TLC) with 6th Gen V-NAND to 132 MB/s with BiCS5, so we'll see 500GB SKUs again soon.

MaxxThis is still using old B27A or B27B flash - 96L Micron. We won't see E18 with Micron's new RG flash until 2021. B27A/B27B is 512Gb/die with four planes, in many respects just an updated B16A/B17A design to allow for higher sequentials. No real 4K improvements with B27A (which the leaked E18 from Ramseyer utilized, I asked), B27B has appeared in one ADATA product but even he says it might be 2021 for that (and later in 2021 for B47x rollout). If you calculate out the seq. writes for 96-way interleaving - which can be achieved on older controllers, but they were limited to 800 MT/s per channel - you see they can only meet the specs at 2TB. tl;dr you're right, and these early E18s are really only for sequentials and really only at 2TB.

I wasn't really expecting huge improvements when it comes to 4K, but at the same time, I wasn't expecting my old WD Black to be faster in the read tests, even less so by quite a margin.

MaxxWe have adaptive read schemes (Samsung's ARC), tiles, SBL, etc. to improve tR for 4K requests. Samsung even uses 8KB subplanes on their new flash - see the 980 Pro. They also use a new TurboWrite. Gauging "user experience" is going to get very difficult with these new drives, relying on 4K random no longer cuts it as you have all sorts of tricks, as another example Kioxia has an adaptive scheme for programming BiCS5 that adjusts for detected tR needs. New QLC from Kioxia and Intel have parallel/independent plane reads which can mask QLC's poor tR (in addition to them "hiding" OS files in SLC cache for reads), etc.

According to my mate at Adata/XPG, the AI controllers aren't all they're cracked up to be. I guess it's one of those things that will improve over time, but right now it would seem that some of what they do doesn't actually improve the performance.

TheLostSwedeDoesn't make sense from a power perspective though, as that's not going to be a power friendly SSD.
Something along the lines of this makes more sense for laptops.
www.anandtech.com/show/16012/the-sk-hynix-gold-p31-ssd-review

Except that ADATA has supposedly designed the S50 Lite to be power-efficient... I agree it's unlikely to be anywhere near a PCIe 3.0 NVMe drive, and in particular the Gold P31, but there are people who will happily take the hit in battery just for the PCIe 4.0 bragging rights (because let's be honest - people who actually need PCIe 4.0 storage speeds on desktop are a handful - and on mobile an even smaller subset).

MaxxWe won't see E18 with Micron's new RG flash until 2021.

Is that a Micron or Phison issue? Since Micron announced RG NAND nearly a year ago, if there aren't at least samples shipping to partners by now, one would assume there's something wrong with their fab process...

MaxxNew QLC from Kioxia and Intel have parallel/independent plane reads which can mask QLC's poor tR...

Isn't that a good thing, though? Should help with QLC adoption and higher capacities for storage, which will hasten HDDs going the way of the dinosaur.

Maxx... relying on 4K random no longer cuts it...

This was always going to happen eventually, and while it will make benchmarking more difficult, I've always been leery of using a single metric as the be-all end-all determinator of whether an SSD is "good" or "bad" - it's as problematic as marketing using sequential R/W for the same, IMO. However, I don't know enough about SSDs to suggest a better/fairer way to generally rank them going forward - would be really interested to hear your thoughts on this.

AssimilatorThis was always going to happen eventually, and while it will make benchmarking more difficult, I've always been leery of using a single metric as the be-all end-all determinator of whether an SSD is "good" or "bad" - it's as problematic as marketing using sequential R/W for the same, IMO. However, I don't know enough about SSDs to suggest a better/fairer way to generally rank them going forward - would be really interested to hear your thoughts on this.

I guess it comes down to what you want/need from an SSD, as so much else these days. For most people, there's little to no extra benefit to be had from the generation of drives that are launching right now. I think we can just keeping hoping for a wide range of tests and then look at what matters for our use cases.
It'll be interesting to see what DirectStorage will bring for games for example. That will obviously bring with it a new set of potential things to test and compare.
I see the 4K numbers (low que depths) as a simple filter between drives I would consider and drives I won't consider, since the sequential speed numbers are simply not interesting any more.

TheLostSwedeI wasn't really expecting huge improvements when it comes to 4K, but at the same time, I wasn't expecting my old WD Black to be faster in the read tests, even less so by quite a margin.


According to my mate at Adata/XPG, the AI controllers aren't all they're cracked up to be. I guess it's one of those things that will improve over time, but right now it would seem that some of what they do doesn't actually improve the performance.

I don't consider 4K beyond around QD/T4 or maybe 8 to be relevant for consumers, but in any case - yeah, it's highly dependent on the flash. Some controllers do eek out more like the SMI designs (thanks to help from Intel) but even SK hynix didn't do bad with their new Cepheus controller. However, there's just a ton of tricks they're using on the flash now to improve tR, which I try to document over at Reddit for those that want the technical details. However we still only have SK hynix's "128L" flash on the market although I've seen 980 Pro results already...in any case, we need to wait for Micron's new TLC and BiCS5.

I haven't worked with ADATA but I have written a whitepaper on SSD ML recently (which is under NDA) - I largely agree it's fluff and primarily for marketing, however it would be a big mistake to disregard the fact that the vast majority of gains in flash and NAND performance have come from increasingly complex algorithms. AI/ML in particular benefits more on the enterprise end, especially with cTLC as a budget option, however it's already used and will be used on consumer drives (e.g. Samsung's flash) to improve tR (latency) but more importantly keep the drive performing better, longer. You have far less degradation of performance after writes and wear due to the improved read retry cycles, and it also enables you to do better programming (e.g. 8-16 vs. 4-8-16) including with interruptions (i.e. doing a read mid-program sequence). So while that seems minor to the consumer, it means you can get more out of the flash.

TheLostSwedeAccording to my mate at Adata/XPG, the AI controllers aren't all they're cracked up to be. I guess it's one of those things that will improve over time, but right now it would seem that some of what they do doesn't actually improve the performance.

Please can we stop using "AI" in scenarios where it doesn't mean anything - at best you could say these controllers are using ML, more honestly you can say that they are using MRU algorithms (which have been around for decades but haven't been labelled as AI or ML). The whole AI/ML thing is 99% marketing, I wish it would die.

TheLostSwedeIt'll be interesting to see what DirectStorage will bring for games for example.

DirectStorage is still a long way away for PC at least... Microsoft will only ship a preview to game devs next year. I don't understand why they didn't design it as part and parcel of DirectX 12 so that it could be available immediately. My guess is that marketing wanted a halo feature that they can sell more of their crappy consoles with...

Maxx... mention of AI/ML (example) - you're right, it's more of a "buzz" thing. I suppose I should have been more clear and just said they are utilizing adaptive algorithms to improve the flash. Although, this certainly involves feedback and learning at least in enterprise. It's unfortunately not something that's going to go away, what with Tensor cores in RTX cards for example.

Oh, I'm well aware that ML is actually a thing in big data companies - but in the components in your desktop or laptop, not a chance. At best you'll have a component that uses an algorithm that has been optimised via ML, but that's done by the vendor - in the case of SSDs, it will mostly be baked-in to the firmware.

NVIDIA tried client ML with Tensor cores in the original RTX with DLSS, and it worked so well that they fell back to using algorithms they've pretrained themselves for DLSS 2.0... ML for individual consumer workloads still has a long way to go before it's actually a thing.

AssimilatorIs that a Micron or Phison issue? Since Micron announced RG NAND nearly a year ago, if there aren't at least samples shipping to partners by now, one would assume there's something wrong with their fab process...

Isn't that a good thing, though? Should help with QLC adoption and higher capacities for storage, which will hasten HDDs going the way of the dinosaur.

This was always going to happen eventually, and while it will make benchmarking more difficult, I've always been leery of using a single metric as the be-all end-all determinator of whether an SSD is "good" or "bad" - it's as problematic as marketing using sequential R/W for the same, IMO. However, I don't know enough about SSDs to suggest a better/fairer way to generally rank them going forward - would be really interested to hear your thoughts on this.

Phison supposedly has at least improved SLC algorithms (for their full-drive SLC caching), beyond that I can't comment although you do need new firmware for new flash. In any case, yes, it is a Micron issue, they've stay latched on to FG with Intel for example with the 96L QLC, they have to retool for something based on TCAT. I mean these time tables are almost always optimistic anyway and then you have the pandemic...

I'm all for better QLC performance, I more meant that it can be deceptively good for example in reviews or specification listings. One could argue that if the consumer doesn't ever "see" it then it's not relevant - which is a valid point, but underlines that NAND is a very murky business. See the recent Level1Techs stream with Malventano to see what I mean.

I can assure you this is something discussed at great length by reviewers since I talk to almost all of the prominent ones at this point. More generally the ability to benchmark has improved the last two years and information has been much better-shared. That being said, staying cutting-edge requires a lot of research and keeping up with the technical side (through conferences and such, livestreams more commonly now, but also reading academic journals and patents). Applying that to "real world" testing is more difficult as I find the engineers to be reticent to reveal proprietary information - for example, I had to track down patents for Crucial's P5 SLC caching scheme as they wouldn't tell me anything. And I had to study their flash geometry (tiles and circuitry) to determine the difference between B27A and B27B since they refused to say. I'm getting off track though - I suspect that, at least for now, 4K isn't really the bottleneck for most things (although software could and will improve), so a better measure is consistency, however most drives this generation will be focused on efficiency and sequentials.

AssimilatorPlease can we stop using "AI" in scenarios where it doesn't mean anything - at best you could say these controllers are using ML, more honestly you can say that they are using MRU algorithms (which have been around for decades but haven't been labelled as AI or ML). The whole AI/ML thing is 99% marketing, I wish it would die.

DirectStorage is still a long way away for PC at least... Microsoft will only ship a preview to game devs next year. I don't understand why they didn't design it as part and parcel of DirectX 12 so that it could be available immediately. My guess is that marketing wanted a halo feature that they can sell more of their crappy consoles with...

I think he was replying to my mention of AI/ML (example, example 2) - you're right, it's more of a "buzz" thing. I suppose I should have been more clear and just said they are utilizing adaptive algorithms to improve the flash. Although, this certainly involves feedback and learning at least in enterprise. It's unfortunately not something that's going to go away, what with Tensor cores in RTX cards for example.

I agree that it will be a while for DirectStorage (and thus related technology like RTX IO) but either way, it's clear the PS5's storage claim to fame isn't ultimately that impressive.

AssimilatorPlease can we stop using "AI" in scenarios where it doesn't mean anything - at best you could say these controllers are using ML, more honestly you can say that they are using MRU algorithms (which have been around for decades but haven't been labelled as AI or ML). The whole AI/ML thing is 99% marketing, I wish it would die.

DirectStorage is still a long way away for PC at least... Microsoft will only ship a preview to game devs next year. I don't understand why they didn't design it as part and parcel of DirectX 12 so that it could be available immediately. My guess is that marketing wanted a halo feature that they can sell more of their crappy consoles with...

No need to get riled up, was only continuing the terminology used by others. Obviously it's not AI, as it's not as if the controller can think.

As for DirectStorage, does it matter if it's now or next year, or later? I was simply saying it'll be one more thing where SSDs will have an impact and it'll one more thing that can hopefully be benchmarked to show us another difference between drives.

TheLostSwedeNo need to get riled up, was only continuing the terminology used by others. Obviously it's not AI, as it's not as if the controller can think.

Sorry about that, I replied as much above but it's currently hidden since I edited it to add some sources to show what I was talking about. Nevertheless the point was that there is adaptive technology at play as we have increasingly capable controllers and AI/ML isn't going away either - even if it's still mainly marketing at this point.

'For added ease of use and convenience, the S70 is backward compatible with PCIe 3.0.'

What kind of monkey installs a PCIe 4.0 M2 SSD on a 3.0 compatible mobo ?

TheLostSwedeI think Adata is going to have to do some more firmware work on this drive before they release it.

FYI I just got word that new E18 firmware is coming out that improves 4K with the older flash, so seems it will make an okay gap product until the new flash arrives.

There's a thing which people are easily overlooking and it's pretty much almost straight from sci-fi: this SSD throughput rivals the RAM speed of computers released just ten years ago.

I guess it can hold that speed for 10 seconds and then it throttles down to a crawl due to overheating ^^

PCIe SSDs still have a long way to go, pushing speeds to the limit isn't the problem.

MaxxFYI I just got word that new E18 firmware is coming out that improves 4K with the older flash, so seems it will make an okay gap product until the new flash arrives.

Yeah, my mate mentioned they're waiting for a new firmware. Hopefully that'll make it more competitive with older drives across the board.

TheDeeGeeI guess it can hold that speed for 10 seconds and then it throttles down to a crawl due to overheating ^^

PCIe SSDs still have a long way to go, pushing speeds to the limit isn't the problem.

I'm sure it can do it for a lot longer than that, especially considering how big that heatsink is.

birdieThere's a thing which people are easily overlooking and it's pretty much almost straight from sci-fi: this SSD throughput rivals the RAM speed of computers released just ten years ago.

That's more an artifact of how slowly (relatively) that DDR speeds have increased over the same time period, thanks to a stagnant design. DDR5 will be much faster than its predecessors.

Still waiting intel to release PCI-E 4 gen so we can buy these faster nvme...What are you waiting for Intel, AMD and Nvidia will eat your lunch ...and dinner too.

AssimilatorWhy are you guys continually surprised that press releases, which are always pure marketing, only ever showcase the best attributes of the product being released?

At least capacity starts at 1TB, which implies 512GB SSDs are one the way out.

Did my writing connotate surprise? I do not think so.... My intent was purely to commiserate with a like minded individual