World's First PCIe 4.0 SSD Demo, Phison E16 Hits 4GB/s

I agree progress is progress, but I was expecting something a bit further from the absolute maximum bandwidth of PCIe 3.0 x4.

https://en.wikipedia.org/wiki/PCI_Express#History_and_revisions

The maximum speed of PCIe 3.0 x4 is 3.94 Gigabytes a second.

The maximum speed of PCIe 4.0 x4 is 7.88 Gigabytes a second.

PCIe 5.0 x4 which according to Wiki is going to be released in Q2 of 2019, doubles this again to 15.75 Gigabytes a second.

I was about to ask a question of what is absolute fastest a cpu could transfer data until I realized that the answer lied with the speed of the ram.

https://en.wikipedia.org/wiki/DDR4_SDRAM#Modules

This should allow us to conclude a few things ... for example.

With a PCIe 4.0 X16 device you should be unable to realize its full potential due to its maximum speed being 31.5 Gigabytes a second, but the maximum speed of a respectable DDR4-3200 ram being 25.6 Gigabytes per second.

This brings up the unexpected and actually kind of funny scenario where you would need to overclock your ram to get the full performance from a PCIe 4.0 x16 SSD adapter.

https://www.tomshardware.co.uk/aplicata-m.2-nvme-ssd-adapter,review-34022-2.html

Comes to mind from about a year ago which actually appears to saturate a PCIe 3.0 x8 at 7.88 gigabytes / second at a queue depth of 4 with 10% or so loss to overhead when reading.

Is this the correct way to think about this?
 

atomicWAR

Glorious
Ambassador


My understanding is running dual channel DDR4 3200 puts you in the 40-50GB/s range depending on you timings. Though I guess you could being going by single channel numbers. My point is that for now we should still have a little wiggle room with PCIe 4.0 vs DDR4 3200 bandwidth but your are correct in the numbers are getting closer. in another few generations an adjustment may need to be made.

https://www.techspot.com/news/62129-ddr3-vs-ddr4-raw-bandwidth-numbers.html
https://ark.intel.com/products/126686/Intel-Core-i7-8700-Processor-12M-Cache-up-to-4_60-GHz i7 8700 listed at 41.6GB/s
 


Ahh thx, that makes perfect sense.

I thought my values were already best case scenario / dual channel.

Well at least this lets us know that in the future not using dual channel can really kill the performance of high end SSDs.

 
Just a few years ago RAM appeared to be this super fast memory whose speed could not be touched by anything.

Even if we double DDR4-3200 from 25.6 gigabytes a second to 51.2 gigabytes a second it still wouldn't be fast enough to support the not released yet PCIe 5.0 x16 specification, 63 gigabytes / second.

Hypothetically they would need to release DDR5 ram (or overclock DDR4 ALOT) before they could fully support the speed of PCIe 5.0 x16
 

mikewinddale

Distinguished
Dec 22, 2016
290
55
18,940
My question is, will we finally get Ryzen motherboard with more than one full-speed M.2? Right now, we get one PCIe 3.0x4 and one PCIe 2.0x4. When Ryzen adds PCIe 4.0, will the second M.2 be upgraded to 3.0?
 

atomicWAR

Glorious
Ambassador


Exactly what i meant about a correction being a couple generations away. As you pointed out PCIe 5.0 is where things hit the fan so to speak. Hopefully DDR5 will be readily available by then.
 


If we look at this logically using the current 2700x it looks like this

https://en.wikichip.org/wiki/amd/ryzen_7/2700x

The 2700X includes 20 PCIe lanes - 16 for a discrete graphics processor and 4 for storage (NVMe or 2 ports SATA Express). or the shorthand way of saying it would be 1x16 + 1x4.

Some Ryzen motherboards .. AM4 have a 2nd M.2 which shares bandwidth with the sata ports and isn't as fast as the main m.2 with the following being an example
https://www.anandtech.com/show/13091/analyzing-b450-for-amd-ryzen-a-quick-look-at-all-the-motherboards/4
Yours also being an example.

With the jump to PCIe 4.0 and assuming a similar 1x16 + 1x4 layout you would expect them to be able to budget more bandwidth due to all lanes doubling in bandwidth versus PCIe 3.0.

If they choose to they could split the link into 2 PCIe 4.0 x2 which would support the same speed as our current PCIe 3.0 x4

Of course if you wanted tons of PCIe lanes you could look at AMD Threadripper
https://en.wikichip.org/wiki/amd/ryzen_threadripper/2950x
2950x being my favorite with its combination of clockspeed and thread count.

Instead of 20 lanes in total, 16+4, Threadripper has 60 lanes allowing for a much more customizable and diverse set of motherboards.



 

atomicWAR

Glorious
Ambassador


No I don't believe that's how it works. Only the first PCIe slot *might* become PCIe 4.0 and rumor has it AMD has not made their mind up on whether or not to allow this functionality without a X500 series board. A bios update would be required and all AMD has to do is not sign off on the code for PCIe 4.0 on older boards for your hopes to die in regards to PCIe 4.0 on x300/x400 series chipsets..
 




Technically if the motherboard manufacturers wanted to they could create 2 PCIe 4.0 x2 slots that would have the same speed as our current PCIe 3.0 x4 slots.

It wouldn't be full speed at that time, but it would be full speed for our current time.

The current Ryzen 2000 series PCIe configuration is 1x16+1x4
https://en.wikichip.org/wiki/amd/ryzen_7/2700x

And so the question becomes "How many lanes will the Ryzen 3xxx cpus and the new 500-series AMD chipsets support?"

Tom even poses a similar question in an article earlier today;

https://www.tomshardware.com/news/amd-ryzen-pcie-4.0-motherboard,38401.html

"We weren't told the specific lane allocations of the new chipset, but those faster lanes will be useful for numerous types of secondary I/O devices." (secondary I/O devices = SSD/GPU and other things)

If it supports 1x16+1x4 then we are in a similar boat as we are now.

If it supports 1x16+2x4 or higher then you may get your 2 full speed m.2 wish.


 

jpe1701

Honorable
They updated the story today with a reply from AMD and they said they would not block board partners from updating x370 and x470 with pcie 4.0, but it would be up to the manufacturer to make the BIOS updates available.
 

atomicWAR

Glorious
Ambassador


Great news!!!
 

everett mcgee

Distinguished
Sep 15, 2011
2
0
18,510
It will come down to more ram channels probably. I think we need to start looking at double sided mainboards to get past the problems that creep up with timing and EMF interference on long traces between the RAM, CPU, GPU and Drives. I think motherboard will need to look very different than they do today. Think CPU and RAM Pads/sockets(not slots) directly on the opposite side of the GPU and drive socket(not slot). They will have to make a new case standard to accommodate this change too.
 

Ksec

Commendable
Aug 21, 2016
6
1
1,510


Someone already said Dual Channel, then there will be improvement to Memory Speed in DDR5. So it is all good.

Most don't expect you to use PCI-E x16 for SSD, not to mention to push for that bandwidth, you need Controller and NAND to working in perfect harmony, for non Enterprise workload, that likely means multiple 512GB SSD working RAID, i.e you are looking at 4 - 8 x PCI-E 4.0 SSD RAID 0, in minimum 2TB to 4TB capacity, not the usual consumer type of setting really.

First Gen PCI-E 3.0 SSD also barely breaks the 2GB/s barrier, now it could reach 3.5GB/s. So it would take a few more years before reaching 7GB/s+ speed.
 

DOSlover

Reputable
Dec 17, 2015
28
0
4,540
Being a pedant for a moment here, perhaps the reference to QLC not being as 'endurant' as TLC had me slightly confused. Perhaps QLC isn't as durable as TLC?
 
Aug 1, 2018
11
0
10
What we need in reality is higher low QD performance. I'd take a lower bandwidth, lower latency SSD over a higher bandwidth, higher latency one. Optane is the perfect example. It doesn't even saturate PCIE 3.0 x4, but it sure as h* is way faster than every other consumer NVMe SSD:s out there.
 

As far as most real-world scenarios go, existing drives are already significantly limited by the rest of the system. Unless you're just doing simple tasks like copying files, even the current NVMe drives need to wait for the CPU, RAM and other components. Take load times in games for example. The system isn't simply loading data off the drive, but also processing that data. Compressed art assets are getting decompressed, procedural data is getting generated, an online server might be getting communicated with, and everything's getting pieced together and prepared for the game to start. Only a portion of that time involves reading data off the drive. So, an NVMe drive ends up not loading much faster than even a SATA SSD that's limited to around 500 MB/s, and the differences will tend to be completely indistinguishable, maybe reducing load times by a second or so in some games. Even if you were loading the data from a RAM disk, the load times of most games wouldn't improve much. So, unless one is doing specialized tasks that involve lots of file reads and writes and little else, even today's PCIe 3.0 SSDs are arguably overkill.
 

supremelaw

Distinguished
Dec 24, 2006
73
0
18,630
Forgive me if this point was already mentioned: I have noticed a few articles which describe an engineering limit to the overall length of a PCIe 4.0 trace. This limit prevents all PCIe slots from oscillating at 16 GHz; only the slot closest to the CPU socket will be able to oscillate at 16 GHz. Maybe future R&D will eliminate this limitation? Onwards and upwards!
 

atomicWAR

Glorious
Ambassador


Anytime you increase the bit count on flash the durability drops proportionately. So yeah QLC is going to have less endurance the TLC just as TLC has less endurance then MLC and MLC has less then endurance then SLC. It's the way it works. Sadly I feel like many QLC SSDs just don't have enough over partitioning to counter act it's low endurance. I worry the market is going to be flooded with crappy QLC SSDs and most folks won't know better. I do like the drop in cost though. It just remains to been seen if it is worth it long term or if the SSD manufactures will need to up their over partitioning game to make up the deficit.
 


I understand a hypothetical PCIe 5.0 x16 SSD wouldn't typically be used for loading World of Warcraft.
(Although if you had 100+ VDI setup and Wow just happened to be in your image and a lightning fast SSD is what was called for ...)

The average consumer, prosumer and even most professionals wouldn't need that kind of speed.

By high end SSDs I meant a hypothetical PCIe 5.0 x16 SSD that could theoretically push 60+ gigabytes/second which almost by definition and price would be for professional and or corporate usage.

The point I was trying to make was that for the absolute most expensive worst case scenario your DDR4 ram would be slower than the PCIe 5.0 x16 specification, a fact that I find slightly amusing.

From what I've read, at no other point in history has the transfer rate of a storage device come so close to exceeding the transfer speed of the current generation ram.

Even going back to SDRAM100 which had a transfer speed of 800 Megabytes a second the hard drive standard back then I believe was was IDE or EIDE/PATA which topped out at Ultra ATA/167 or 167 megabytes a second. (I don't believe SDRAM had dual channel mode ... it was before my time lol ... in any case 334 megabytes a second is still slower than 800 megabytes a second by a large margin.)

Most people won't be affected nor care that this could eventually happen, but it just goes to show how quickly storage speed has caught up to the speed of ram.




 

USAFRet

Titan
Moderator


hahahaha...
A year from now, people here WILL be asking about exactly that.

We already have people trying to shoehorn NVMe or even an Optane 905p into a 10 year old Z68 system.
 

atomicWAR

Glorious
Ambassador


No one would ever do that....<coughs> I mean I might have made a custom UEFI/Bios so I can maybe run a Samsung 960 EVO 500GB as my boot drive on a X79 based rig, which should have of been supported out of box.<cough cough>....Yeah I don't know who would do some crazy stuff like that