Wow, 9W power just for the controller is quite something. Not criticizing them of course, this is not meant for some Laptop SSD, but it's funny to see a regular DataCenter Hard Drive be rated at around 8.5W (UltraStar DC HC 530 SAS) while this is 9W just for the controller, not counting the Flash.
Then again, you don't get to 1M IOPS and 14 GB/s transfer rate for free.
My guess is that consumer mobile versions will be using fewer lanes. PCIe 5.0 x1 will have the same bandwidth as PCIe 3.0 x4, and PCIe 5.0 x2 will have the same as PCIe 4.0 x2.
For consumer uses, I think we might actually soon hit the point of diminishing returns for storage. Just the improvement to random access obtain moving from HDD to SATA SSD already achieved the majority of the returns for consumer uses.
I'm not saying that faster storage won't benefit consumers, but I think that for consumers, advances in storage won't be any more remarkable than advances in CPUs. For someone who edits Word documents and browses the internet and watches movies, the move from one generation of CPUs to another is a nice improvement, but it won't change their life. Now, the move from HDD to SATA SSD *did* utterly transform the way a consumer uses their computer, but I think that we've reached the point of diminishing returns.
Therefore, I suspect that for a lot of consumer uses, new versions of PCIe will translate into fewer lanes (and therefore higher battery life), rather than faster data transfer speeds.
"PCIe 5.0 x1 will have the same bandwidth as PCIe 3.0 x4, and PCIe 5.0 x2 will have the same as PCIe 4.0 x2." I doubt we are ever going to see a single lane SSD, even if it made sense -at first glance- for its target market. It would make far more sense, and would almost certainly be cheaper and more robust (a single PCIe 5.0 lane would be quite weaker, more susceptible to noise and less far-reaching than two PCIe 4.0 lanes all else being equal) to just use two PCIe 4.0 lanes.
Yes, the controller power will only get that high when using the full performance. In such scenarios, the whole drive power will probably be in the 25W range (E1.S with heatsink). Some customers are likely to use these controllers in more power-limited enclosures/systems, where 14GB/s per drive will not be attainable.
All said, that's 9.8w peak + a few watts for the NAND\RAM, so while more than a HDD, it'll idle at much less than a typical data center drive as the only ones that ever really spin down are for cold storage.
Seems like alder lake might just be feasible for those people that want the latest and greatest specs. We already have DDR5 chips, we know have pcie ssd controller.
It would be an odd consumer who'd buy an Alder Lake CPU, slap a PCIe 5.0 4x SSD into the one PCIe 5.0 slot on their system and in the process sacrifice the ability to use anything but the iGPU 😅
If someone's video workflow could be done entirely with Quick Sync, it'd actually be pretty neat -- stuff the whole thing in an ITX case with a handle and go full road warrior.
It's a pretty safe bet - there aren't likely to be consumer PCIe 5.0 devices for at least the first 6 months after launch, if not a full year, and even then they're probably not going to be relevant for 95% of consumer use cases. It'll be interesting to see how quickly GPU makers change across, given that even PCIe 4.0 isn't needed yet.
SSDs usually don't need bleeding edge DRAM. FTL table lookups don't use much bandwidth, and even slow DRAM has much better latency than NAND flash. They also don't need the huge capacity of server DIMMs; just 1GB per 1TB of NAND. Typical datacenter SSDs are still in the low TBs range, though capacities are going up as new interfaces allow for higher per-drive performance.
Also, the on-die ECC in DDR5 and other recent DRAMs is no substitute for the explicit ECC that comes from using a 72-bit bus instead of 64-bit. The latter is a hard requirement for the end-to-end error correction that enterprise SSDs need to provide.
On-die ECC as used in DDR5 is invisible to the host system. It protects data at rest in the DIMM, but does nothing to protect the data as it is transmitted between the DIMM and the processor. So the data that arrived at the DIMM may have been corrupted before the DDR5 chips could even compute and store their ECC data. Enterprise grade products need ECC protection along the entire data path, which means adding more bits to the data bus.
A simple implementation of in-band ECC would require too many extra write cycles to have an acceptable performance.
Complex implementations, like the in-band ECC controllers that are enabled in some embedded SKUs of Intel Tiger Lake U and Elkhart lake use clever caching algorithms to eliminate most of the extra writes.
However, caching is never foolproof, there will always be some applications where most of the writes cannot be cached and have to go to the memory lowering dramatically the performance.
In my opinion in-band ECC is a bad solution. It is an ugly workaround for the fact that the market for users that value reliability is much smaller that the market for clueless users, so the LPDDR have been not made available also with the larger width required for normal ECC.
In-band ECC is better than no ECC, but much worse than traditional ECC.
> In-band ECC is better than no ECC, but much worse than traditional ECC.
Thanks for sharing your thoughts. I agree, but it does seem attractive for certain embedded applications, especially where the number of DRAM chips might be too low to support conventional ECC.
Typical SSDs maybe, but spaceinvaderone on youtube has a video where he's making an unraid server at his friends business where they're using 100TB SSDs, its absolutely ridiculous. Per his video each one cost roughly $40k
I wouldn't make that comparison though as it's not an apples to apples comparison. This ssd controller is really impressive, even though 9W just for the controller itself is absurd where power consumption is concerned. Keep in mind that the r/w speeds of 16 and 9GB/s respectively are only for short, bursty workloads, not sustained workloads (unless you want the storage chips on the m.2 to burn a hole through the pcb due to the excessive heat produced). There's no ssd that I've seen, either in a liquid cooled m.2 raid0 card or air cooled, that can run at the its peak advertised speeds for more than a few seconds. It won't happen in a pc and it sure won't in super enclosed box like a console. Peak speeds are one thing, sustained speeds are another matter entirely, there's no direct correlation between the two. The former is only attainable for short periods of time in bursty workloads, the latter can be achieved for an indefinite period of time provided the workload calls for it. It's impressive what improvements 5.0 brings over 4.0 in speeds alone though, even though the power consumed is absurd.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
47 Comments
Back to Article
MenhirMike - Thursday, May 27, 2021 - link
Wow, 9W power just for the controller is quite something. Not criticizing them of course, this is not meant for some Laptop SSD, but it's funny to see a regular DataCenter Hard Drive be rated at around 8.5W (UltraStar DC HC 530 SAS) while this is 9W just for the controller, not counting the Flash.Then again, you don't get to 1M IOPS and 14 GB/s transfer rate for free.
Xajel - Thursday, May 27, 2021 - link
I guess when we're closing the consumer mobile version they will already use better process tech to lower the power envelope.mode_13h - Thursday, May 27, 2021 - link
> consumer mobile versionLOL
Mikewind Dale - Thursday, May 27, 2021 - link
My guess is that consumer mobile versions will be using fewer lanes. PCIe 5.0 x1 will have the same bandwidth as PCIe 3.0 x4, and PCIe 5.0 x2 will have the same as PCIe 4.0 x2.For consumer uses, I think we might actually soon hit the point of diminishing returns for storage. Just the improvement to random access obtain moving from HDD to SATA SSD already achieved the majority of the returns for consumer uses.
I'm not saying that faster storage won't benefit consumers, but I think that for consumers, advances in storage won't be any more remarkable than advances in CPUs. For someone who edits Word documents and browses the internet and watches movies, the move from one generation of CPUs to another is a nice improvement, but it won't change their life. Now, the move from HDD to SATA SSD *did* utterly transform the way a consumer uses their computer, but I think that we've reached the point of diminishing returns.
Therefore, I suspect that for a lot of consumer uses, new versions of PCIe will translate into fewer lanes (and therefore higher battery life), rather than faster data transfer speeds.
Mikewind Dale - Thursday, May 27, 2021 - link
"and PCIe 5.0 x2 will have the same as PCIe 4.0 x2"Derp, I meant PCIe 4.0 x4"
Santoval - Saturday, May 29, 2021 - link
"PCIe 5.0 x1 will have the same bandwidth as PCIe 3.0 x4, and PCIe 5.0 x2 will have the same as PCIe 4.0 x2."I doubt we are ever going to see a single lane SSD, even if it made sense -at first glance- for its target market. It would make far more sense, and would almost certainly be cheaper and more robust (a single PCIe 5.0 lane would be quite weaker, more susceptible to noise and less far-reaching than two PCIe 4.0 lanes all else being equal) to just use two PCIe 4.0 lanes.
drmaddogs - Saturday, June 19, 2021 - link
No more Gallium and much more small.. coming over ten years.Daviii - Thursday, May 27, 2021 - link
9W is the max controller power. My guess is that's the peak consumption in those (admittedly rare) scenarios of sustained 14GB/s transfers?Billy Tallis - Thursday, May 27, 2021 - link
Yes, the controller power will only get that high when using the full performance. In such scenarios, the whole drive power will probably be in the 25W range (E1.S with heatsink). Some customers are likely to use these controllers in more power-limited enclosures/systems, where 14GB/s per drive will not be attainable.mode_13h - Thursday, May 27, 2021 - link
> 9W power just for the controller is quite something.If performance increased by enough, then perf/W could still be better than previous gen.
Catalina588 - Thursday, May 27, 2021 - link
You can run a stock exchange on a million IOPS. Amazing progression of technology.mode_13h - Saturday, May 29, 2021 - link
According to what source, and from when?Samus - Sunday, May 30, 2021 - link
All said, that's 9.8w peak + a few watts for the NAND\RAM, so while more than a HDD, it'll idle at much less than a typical data center drive as the only ones that ever really spin down are for cold storage.It might average out, energy usage wise.
mode_13h - Sunday, May 30, 2021 - link
First of all, Billy seems the total number will be more like 25W.> It might average out, energy usage wise.
Than HDDs? No. This is not meant to compete with HDDs. Completely weird comparison, on so many levels.
Kurosaki - Thursday, May 27, 2021 - link
Now, we just wait for memory woth the controller. 14GB/s in QD1..Zingam - Thursday, May 27, 2021 - link
Now pair these with OCL or even better HCL Flash and we'll have a true winner!Marlin1975 - Thursday, May 27, 2021 - link
I agree, 9w is a lot for just the controller. Did they say what node these are made on?yeeeeman - Thursday, May 27, 2021 - link
Seems like alder lake might just be feasible for those people that want the latest and greatest specs. We already have DDR5 chips, we know have pcie ssd controller.mode_13h - Thursday, May 27, 2021 - link
> alder lake might just be feasible for those people that want the latest and greatest specs.Huh? This is not for consumer SSDs. And the only way it could be used in one is on a PCIe AIC form factor, definitely not M.2.
Spunjji - Friday, May 28, 2021 - link
It would be an odd consumer who'd buy an Alder Lake CPU, slap a PCIe 5.0 4x SSD into the one PCIe 5.0 slot on their system and in the process sacrifice the ability to use anything but the iGPU 😅lmcd - Monday, May 31, 2021 - link
If someone's video workflow could be done entirely with Quick Sync, it'd actually be pretty neat -- stuff the whole thing in an ITX case with a handle and go full road warrior.nandnandnand - Thursday, May 27, 2021 - link
I predict that most Alder Lake buyers will use DDR4 instead of DDR5 and no PCIe 5.0.Spunjji - Friday, May 28, 2021 - link
It's a pretty safe bet - there aren't likely to be consumer PCIe 5.0 devices for at least the first 6 months after launch, if not a full year, and even then they're probably not going to be relevant for 95% of consumer use cases. It'll be interesting to see how quickly GPU makers change across, given that even PCIe 4.0 isn't needed yet.haukionkannel - Friday, May 28, 2021 - link
First pci 5.0 gpus 10 years from now? At earliest 4 years… unless we get 8 lines highend gpus… that would cut costs somewhat…mode_13h - Saturday, May 29, 2021 - link
> First pci 5.0 gpus 10 years from now? At earliest 4 years…Pretty much.
5080 - Thursday, May 27, 2021 - link
Interesting, PCIe 5.0 but support for cache with DDR4 and not DDR5 with integrated ECC?Billy Tallis - Thursday, May 27, 2021 - link
SSDs usually don't need bleeding edge DRAM. FTL table lookups don't use much bandwidth, and even slow DRAM has much better latency than NAND flash. They also don't need the huge capacity of server DIMMs; just 1GB per 1TB of NAND. Typical datacenter SSDs are still in the low TBs range, though capacities are going up as new interfaces allow for higher per-drive performance.Also, the on-die ECC in DDR5 and other recent DRAMs is no substitute for the explicit ECC that comes from using a 72-bit bus instead of 64-bit. The latter is a hard requirement for the end-to-end error correction that enterprise SSDs need to provide.
KarlKastor - Thursday, May 27, 2021 - link
Where is the difference of ECC we know from actual DIMM from that on-die ECC of DDR5?Billy Tallis - Thursday, May 27, 2021 - link
On-die ECC as used in DDR5 is invisible to the host system. It protects data at rest in the DIMM, but does nothing to protect the data as it is transmitted between the DIMM and the processor. So the data that arrived at the DIMM may have been corrupted before the DDR5 chips could even compute and store their ECC data. Enterprise grade products need ECC protection along the entire data path, which means adding more bits to the data bus.mode_13h - Thursday, May 27, 2021 - link
What's your take on in-band ECC? Do you think it'll displace OOB ECC, for less performance-intensive applications?AdrianBc - Thursday, May 27, 2021 - link
A simple implementation of in-band ECC would require too many extra write cycles to have an acceptable performance.Complex implementations, like the in-band ECC controllers that are enabled in some embedded SKUs of Intel Tiger Lake U and Elkhart lake use clever caching algorithms to eliminate most of the extra writes.
However, caching is never foolproof, there will always be some applications where most of the writes cannot be cached and have to go to the memory lowering dramatically the performance.
In my opinion in-band ECC is a bad solution. It is an ugly workaround for the fact that the market for users that value reliability is much smaller that the market for clueless users, so the LPDDR have been not made available also with the larger width required for normal ECC.
In-band ECC is better than no ECC, but much worse than traditional ECC.
mode_13h - Saturday, May 29, 2021 - link
> In-band ECC is better than no ECC, but much worse than traditional ECC.Thanks for sharing your thoughts. I agree, but it does seem attractive for certain embedded applications, especially where the number of DRAM chips might be too low to support conventional ECC.
theno1patrick - Friday, July 2, 2021 - link
If you want some reliability on your SSD/NVME, check out the new plot ripper by sabrent. insane PBW!saratoga4 - Thursday, May 27, 2021 - link
It is first and foremost a cost savings measure to enable vendors to sell DRAM cells that would otherwise have an error rate too high to be usable.KarlKastor - Thursday, May 27, 2021 - link
Makes sense. Thank you for the explanation.theno1patrick - Friday, July 2, 2021 - link
Typical SSDs maybe, but spaceinvaderone on youtube has a video where he's making an unraid server at his friends business where they're using 100TB SSDs, its absolutely ridiculous. Per his video each one cost roughly $40kSamus - Sunday, May 30, 2021 - link
I'm surprised it's even DDR4. Most SSDs were still DDR3LP until a year or two ago.mode_13h - Sunday, May 30, 2021 - link
Look at the data rates! And it's only using a 64-bit datapath to DRAM, so DDR3L was out of the question.DigitalFreak - Thursday, May 27, 2021 - link
14GB/sec... So for consumer SSDs with QLC to hit that number you'd need 64 channels and 16TB of flash. LOLKarlKastor - Thursday, May 27, 2021 - link
Why that? That's the read speed. Is not much lower with QLC.MrCommunistGen - Thursday, May 27, 2021 - link
Referencing the original press release, it looks like the Bravera SC5 are currently sampling to select customers.MrCommunistGen - Thursday, May 27, 2021 - link
Ugh... removing foot from mouth... it's literally the last line in the article.ballsystemlord - Thursday, May 27, 2021 - link
Don't kick yourself. It happens.Kenshiro70 - Thursday, May 27, 2021 - link
Playstation and XBox: Our storage is a generation faster!Marvell: Hold my beer.
MrTechie - Friday, May 28, 2021 - link
I wouldn't make that comparison though as it's not an apples to apples comparison. This ssd controller is really impressive, even though 9W just for the controller itself is absurd where power consumption is concerned. Keep in mind that the r/w speeds of 16 and 9GB/s respectively are only for short, bursty workloads, not sustained workloads (unless you want the storage chips on the m.2 to burn a hole through the pcb due to the excessive heat produced). There's no ssd that I've seen, either in a liquid cooled m.2 raid0 card or air cooled, that can run at the its peak advertised speeds for more than a few seconds. It won't happen in a pc and it sure won't in super enclosed box like a console. Peak speeds are one thing, sustained speeds are another matter entirely, there's no direct correlation between the two. The former is only attainable for short periods of time in bursty workloads, the latter can be achieved for an indefinite period of time provided the workload calls for it. It's impressive what improvements 5.0 brings over 4.0 in speeds alone though, even though the power consumed is absurd.Kaique Gerais - Wednesday, June 2, 2021 - link
From this power consumption this might be the same TSMC 12FFC process they used on there PCIE 4.0 Controllers and not a 7nm Process.doylecc - Saturday, July 10, 2021 - link
Pricing?