News China-made RISC-V PCIe 5.0 SSD controller promises competitive performance — up to 14.2 GB/s without a fan

[Admin]

Yingren (aka InnoGrit) has announced the YRS820, a high-end PCIe 5.0 controller for consumer SSDs based on the RISC-V architecture.

China-made RISC-V PCIe 5.0 SSD controller promises competitive performance — up to 14.2 GB/s without a fan : Read more

 

[Notton]

That IOPS number is insane.
2.0M?
Do we finally have a NAND M.2 drive that beats Optane in IOPS?

 

[thestryker]

That IOPS number is insane.
2.0M?
Do we finally have a NAND M.2 drive that beats Optane in IOPS?

No, and it will never be beaten by consumer NAND drives because these numbers are always high queue depth. High end PCIe 4.0 drives could already beat Optane at high queue depth IOPs since their sequential is higher.

 

[DaveLTX]

That IOPS number is insane.
2.0M?
Do we finally have a NAND M.2 drive that beats Optane in IOPS?

M.2 drives have been maxing out pcie 4.0 in IOPS.
For high queue depth. Optane beats them in low QD only but then again, optane is dead, not much demand for them existed and it was difficult to scale up plus the cost as well. Sad.
But we have proper SLC short bitword SCM now which isn't far and apparently DC operators say optane is largely overkill... probably unless the price was lower.

 

[Amdlova]

Last innogrit I have feel so sluggish big big numbers but the system it's a complete garbage.
I prefer a huge heatsink and a good know brand with prioritary controller (samsung or sk hynix).
Innogrit on paper is amazing but when the system has slowdown or hiccups you see those big number not make any sense at all.

 

[DaveLTX]

Last innogrit I have feel so sluggish big big numbers but the system it's a complete garbage.
I prefer a huge heatsink and a good know brand with prioritary controller (samsung or sk hynix).
Innogrit on paper is amazing but when the system has slowdown or hiccups you see those big number not make any sense at all.

You know your posts are like exactly what you're writing about
Big big numbers but brain empty

 

[bit_user]

That IOPS number is insane.
2.0M?
Do we finally have a NAND M.2 drive that beats Optane in IOPS?

LOL! You're joking, right?

https://www.phoronix.com/news/Linux-14M-IOPS-Per-Core
​

I was so impressed that I bought a P5800X, when they got discontinued. Still haven't figured out what to do with it, though.

 

[bit_user]

High end PCIe 4.0 drives could already beat Optane at high queue depth IOPs since their sequential is higher.

M.2 drives have been maxing out pcie 4.0 in IOPS.
For high queue depth.

Okay, guys. It's "put up or shut up" time. Show me a NAND-based SSD that can sustain 14M IOPS. Any queue depth.

 

[thestryker]

Okay, guys. It's "put up or shut up" time. Show me a NAND-based SSD that can sustain 14M IOPS. Any queue depth.

This is what I'm referring to:

 

[bit_user]

This is what I'm referring to:

First, those are Peak numbers. The P5800X can sustain those rates like all day long, way after those consumer drives' buffers have been exhausted.

Second, those don't actually show higher IOPS, they show higher 4k IOPS. Axboe used 512-byte payloads. If you really want to know the transaction-processing throughput of the hardware, try using a smaller size. Hint: it's not going to look so good for the NAND drives.

Lastly, wrong OS. To really make these drives sing, you need Linux. If you just naively divide his 14M IOPS figure by 8, to get the expected 4k IOPS rate, you get 1.75 M, not 0.949 M. The difference is surely in the OS. Maybe the WD_Black would've scaled up by the same amount, but there are possible reasons why it might not. To truly know, one should test them both on Linux.

 

[thestryker]

First, those are Peak numbers. The P5800X can sustain those rates like all damn day, long after those consumer drives' buffers have been exhausted.

Second, those don't actually show higher IOPS, they show higher 4k IOPS. Axboe used 512-byte payloads. If you really want to know the transaction-processing throughput of the hardware, try using a smaller size. Hint: it's not going to look so good for the NAND drives.

Lastly: wrong OS. To really make these drives sing, you need Linux.

At no point did I say NAND is better than Optane just that under specific circumstances they can already pass it.

SCM NAND drives can do the same exact thing as Optane in regards to all day long.

DapuStor Xlenstor2 X2900P 800GB Review The 100 DWPD Next-Gen SLC Optane Alternative

If you are looking for a SLC NAND based alternative with Optane being sunset, the DapuStor Xlenstor2 X2900P is fast with a 100 DWPD rating

www.servethehome.com

Optane is the best storage technology to date, and if Intel had kept developing it there's no NAND which would ever be faster. However because the P5800X is the best we're ever going to get it will be surpassed by PCIe 5.0+ SCM drives in almost every area.

 

[bit_user]

SCM NAND drives can do the same exact thing as Optane in regards to all day long.

DapuStor Xlenstor2 X2900P 800GB Review The 100 DWPD Next-Gen SLC Optane Alternative

If you are looking for a SLC NAND based alternative with Optane being sunset, the DapuStor Xlenstor2 X2900P is fast with a 100 DWPD rating

www.servethehome.com

Yes, though I did cite consumer SSDs in that comment. I know how NAND-based enterprise SSDs' behavior differs. I also have an Intel/Solidigm P5620, which is about 10x the capacity and cost me like 1/3rd as much.

Furthermore, on the subject of sustained random writes, that Dapustor drive still couldn't beat the P5800X.

Also, note the difference in 4k IOPS performance between their testing and Toms'. STH achieved much closer to my extrapolated estimates, though still not what I think Axboe's test would've managed.

because the P5800X is the best we're ever going to get it will be surpassed by PCIe 5.0+ SCM drives in almost every area.

Let me know when it happens. For NVMe drives, you're probably right, but only because NVMe overhead still hides some of Optane's intrinsically low latency & random access.

For NVDIMMs, probably the only thing that can beat it is NAND-backed or battery-backed DRAM. The very highest tier of CXL PMem modules will probably take this approach.

 

[thestryker]

Let me know when it happens. For NVMe drives, you're probably right, but only because NVMe overhead still hides some of Optane's intrinsically low latency & random access.

I'm not sure when we're going to see new SCM in the format Dapustor does as they seem to be the only ones going after the same performance profile as Optane. All of the other SCM I've seen is really workload targeted which makes sense from the vendor standpoint.

For NVDIMMs, probably the only thing that can beat it is NAND-backed or battery-backed DRAM. The very highest tier of CXL PMem modules will probably take this approach.

It'll be interesting to see what storage technologies develop here due to CXL. We're approaching a time where I think backed DRAM makes a lot of sense again due to increasing memory channels. The sheer size of those 256GB MCR DIMMs just reinforces to me that if Intel had stuck it out they'd see a market for that level of density again. Of course I know why they axed the technology and unit, but that doesn't make me any less sad about it.

Furthermore, on the subject of sustained random writes, that Dapustor drive still couldn't beat the P5800X.

Yeah it's going to take PCIe 5.0 to pass Optane here I think. What's really interesting about the Dapustor drives is that the smaller one has higher rated IOPs, but I couldn't find any testing and even then it'd still be behind (STH's test of the 800GB actually matches the rated IOPs almost exactly).

I also have an Intel/Solidigm P5620, which is about 10x the capacity and cost me like 1/3rd as much.

I came really close to getting a couple of those, but I just don't need the sustained write performance and the 3.2TB ones were around 50% more/GB than consumer drives at the time I was buying. I do have a 960GB 905P sitting atop the stack for my next primary build though (really really wish the P5800X had a similar fire sale)!

 

[bit_user]

I came really close to getting a couple of those, but I just don't need the sustained write performance and the 3.2TB ones were around 50% more/GB than consumer drives at the time I was buying. I do have a 960GB 905P sitting atop the stack for my next primary build though (really really wish the P5800X had a similar fire sale)!

Well, it looks like P5800X finally went back up in price. I could flip mine for a small profit, based on current ebay pricing.

 

[DaveLTX]

I'm not sure when we're going to see new SCM in the format Dapustor does as they seem to be the only ones going after the same performance profile as Optane. All of the other SCM I've seen is really workload targeted which makes sense from the vendor standpoint.

It'll be interesting to see what storage technologies develop here due to CXL. We're approaching a time where I think backed DRAM makes a lot of sense again due to increasing memory channels. The sheer size of those 256GB MCR DIMMs just reinforces to me that if Intel had stuck it out they'd see a market for that level of density again. Of course I know why they axed the technology and unit, but that doesn't make me any less sad about it.

Yeah it's going to take PCIe 5.0 to pass Optane here I think. What's really interesting about the Dapustor drives is that the smaller one has higher rated IOPs, but I couldn't find any testing and even then it'd still be behind (STH's test of the 800GB actually matches the rated IOPs almost exactly).

I came really close to getting a couple of those, but I just don't need the sustained write performance and the 3.2TB ones were around 50% more/GB than consumer drives at the time I was buying. I do have a 960GB 905P sitting atop the stack for my next primary build though (really really wish the P5800X had a similar fire sale)!

My P4800X 375gb on a per GB cost... 8x more than my 3.2TB CD6. Ouch.
And that was the cheapest per GB optane I could get.
Anyway, the DC developers are saying Optane is just plain overkill. Kioxia XL flash SCM and dapustor Kioxia XL flash is already very suitable

 

[thestryker]

My P4800X 375gb on a per GB cost... 8x more than my 3.2TB CD6. Ouch.

Yeah it wasn't until the first gen fire sales came that I got a 960GB 905p for ~$300 (and of course the 1.5TB went on sale for ~$400 when I did but I opted to save the money) which is a bit more than my 2x 2TB P44 Pro drives cost around the same time.

And that was the cheapest per GB optane I could get.
Anyway, the DC developers are saying Optane is just plain overkill. Kioxia XL flash SCM and dapustor Kioxia XL flash is already very suitable

Yeah 3D XPoint absolutely needed another generation minimum, but maybe two before the cost factor would have made sense. This is of course assuming that it was even possible which we'll likely never know. I just know the technology had linear scaling as layers increased so if it was feasible to do so it would eventually make sense for enterprise/workstation/HEDT.

Just the same I would be interested in a workstation/HEDT drive with XL Flash assuming they could make the pricing work.

 

[DaveLTX]

Okay, guys. It's "put up or shut up" time. Show me a NAND-based SSD that can sustain 14M IOPS. Any queue depth.

At small small sizes that makes no sense to anyone
There's on paper and technically then there's reality

Yeah it wasn't until the first gen fire sales came that I got a 960GB 905p for ~$300 (and of course the 1.5TB went on sale for ~$400 when I did but I opted to save the money) which is a bit more than my 2x 2TB P44 Pro drives cost around the same time.

Yeah 3D XPoint absolutely needed another generation minimum, but maybe two before the cost factor would have made sense. This is of course assuming that it was even possible which we'll likely never know. I just know the technology had linear scaling as layers increased so if it was feasible to do so it would eventually make sense for enterprise/workstation/HEDT.

Just the same I would be interested in a workstation/HEDT drive with XL Flash assuming they could make the pricing work.

Apparently they quickly found out that optane layer scaling isnt quite linear or possible. Sad. It is already layered but not enough to grow substantially more
XL flash is exactly what Kioxia sells as FL6 and the flash driving the dapustor iirc

 

[thestryker]

Apparently they quickly found out that optane layer scaling isnt quite linear or possible. Sad. It is already layered but not enough to grow substantially more

This is still just a rumor as far as I'm aware as it would have cost Intel billions more to continue manufacture due to Micron stopping. While it's entirely possible scaling wasn't going to happen it's equally believable that blowing billions more for what amounts to as a small guaranteed market simply wasn't feasible with all of the other issues facing the company.

 

[bit_user]

At small small sizes that makes no sense to anyone

That's not true! Database reads are frequently well under 4k!

There's on paper and technically then there's reality

This tells me you didn't even follow the link, because these are actual speeds Axboe achieved on real hardware.