News SMI CEO says no PCIe 6.0 SSDs for PC "until 2030" as Nvidia demands 100M IOPS — Wallace C. Kou on the future of SSDs

Wow, that's quite an interview! Thanks for this. I've missed having interviews with influential tech companies and individuals, lately.

I'm still only part-way through, but will have to return to it, later. Even though the questions are so far quite friendly, I've definitely learned more about SMI and the SSD sector than I knew before.
 
Wow, that's quite an interview! Thanks for this. I've missed having interviews with influential tech companies and individuals, lately.

I'm still only part-way through, but will have to return to it, later. Even though the questions are so far quite friendly, I've definitely learned more about SMI and the SSD sector than I knew before.
It was a good interview. Not everything has to be antagonistic. I think that is what is wrong with our news organizations right now. They are all out to get that "gotcha" moment.
 
  • Like
Reactions: snemarch
Thank you for this interview, it was very interesting and informative about some of the current and upcoming market dynamics. Great bonus that we have SMIs views on that and a lay of the landscape.
 
  • Like
Reactions: bit_user
It was a good interview. Not everything has to be antagonistic. I think that is what is wrong with our news organizations right now. They are all out to get that "gotcha" moment.
I think it's important to be willing to ask tough questions. Sure, you want to be fair, but a journalist should consider themselves a proxy for the readers and therefore try to anticipate what readers will be wondering about or what they'd like to know.

The only time you have to be aggressive is when the interviewee is being evasive. Even then, you just want to make sure they're intentionally not answering your question and make sure it's clear enough to your audience. Once that much has been established, there's no point in tormenting the interviewee, so you just move on.
 
From what I've seen it looks like they want to actually have a product ship with XL-flash mentioned in the article. But Kioxia has talked about XL-flash many times in past with nothing concrete to show for it, so we'll see I guess.
 
Great in depth interview which is something we've been largely missing from the tech space. Interesting to see the strategy that SMI is using and it makes sense why we've largely only seen their more budget controllers in the past on the retail market. With a big focus on high performance and a push for PCIe 6.0 with more advanced nodes I wonder what sort of developments will happen on the consumer side. While I certainly think he's right about consumer PCIe 6.0 not being a thing I wonder what manufacturers will do to try to keep selling new drives as there's a long time between now and then to keep a retail business alive.
 
  • Like
Reactions: snemarch
From what I've seen it looks like they want to actually have a product ship with XL-flash mentioned in the article. But Kioxia has talked about XL-flash many times in past with nothing concrete to show for it, so we'll see I guess.
Here are three shipping examples of SSDs featuring XL-Flash:
 
I wonder what sort of developments will happen on the consumer side. While I certainly think he's right about consumer PCIe 6.0 not being a thing I wonder what manufacturers will do to try to keep selling new drives as there's a long time between now and then to keep a retail business alive.
  • More capacity.
  • More IOPS.
  • Lower QD1 read latency?
  • Faster write speeds.
  • Lower temperatures / better efficiency -> smaller heatsinks and not requiring active coolers.
  • Moar RGB.

There's a lot that matters besides peak sequential reads. Laptops, in particular, sorely need SSDs with better efficiency.

Also, keep in mind that we endured several years of SATA, where it wasn't hard for drives to hit the sequential limit of the SATA interface, yet there was still some meaningful competition on things like IOPS, capacity, price, etc.

Finally, I'm pretty sure the mainstream desktop platforms will see CXL before PCIe 6.0. Perhaps that will allow for lower latency or higher IOPS. It won't be like SATA -> NVMe, but should still be a differentiator - especially for DRAM-less drives.
 
  • Like
Reactions: snemarch
There's a lot that matters besides peak sequential reads. Laptops, in particular, sorely need SSDs with better efficiency.
Well aware of that, but this is something we simply haven't seen to date with NVMe outside of exceptions.

The P31 was the PCIe 3.0 exception, but that was largely driven by superior NAND with a fantastic 4 channel controller.

Samsung did have the 980 and 990 Pro during PCIe 4.0, but the 980 Pro wasn't very good as a premium product. Everyone else in the high performance space had a single controller (Phison may have put out two controllers over the lifetime, but if I recall it was efficiency not performance which is good if you're using a high end SSD in a laptop and that's about it).

PCIe 5.0 we're seeing some significant efficiency improvements due to better nodes being used, but those are already coming out and they've got another half decade to go. Now it's a matter of which routes they go to improve products and I'm hoping it's not limited to NAND improvements alone due to all of the areas that can be improved.
 
  • Like
Reactions: snemarch
You're right, and I noticed it after I posted and was too lazy to go back and re-edit.
I wish they'd just do a single pass with CDM to see the 4k rnd q1t1 latency results. I know it's not exactly a server benchmark, but still. I can get 1tb of p5800x for as low as $800 for the drive itself. So there are still constraints on pricing.
 
  • Like
Reactions: bit_user
I noticed it after I posted and was too lazy to go back and re-edit.
Would've saved me some time.
; )

I wish they'd just do a single pass with CDM to see the 4k rnd q1t1 latency results. I know it's not exactly a server benchmark, but still.
Agreed. Even though it's an artificial benchmark, it would be nice to have some apples-to-apples way to compare it both with consumer drives and sometimes manufacturers also specify that figure.

I can get 1tb of p5800x for as low as $800 for the drive itself. So there are still constraints on pricing.
Uh, well, they only shipped in 0.4, 0.8, 1.6, and 3.2 TB capacities.

I'm not seeing stock of new, Intel-branded drives in the channel, so I'm guessing those are either used or at least not via authorized resellers? Probably most datacenter operators wouldn't consider using drives from grey market sources, unless they were desperate.

BTW, the lowest ebay price I can find on a new 800 GB drive is $2.4k (from China - probably not including tariffs) and the 1.6 TB capacity drives start at $3k (from China).
 
So, a Q1T1 random 4k read speed of 156.67 MB/s (and I assume they don't mean MiB/s) works out to 38,250 IOPS. That gives a transaction time of 26.1 usec, which isn't exactly the read latency but close enough.

Also, it would be nice to have some confirmation, but I think CrystalMark is measuring file I/O, rather than raw device I/O. Manufacturer specs are definitely the latter.
 
Would've saved me some time.
; )


Agreed. Even though it's an artificial benchmark, it would be nice to have some apples-to-apples way to compare it both with consumer drives and sometimes manufacturers also specify that figure.


Uh, well, they only shipped in 0.4, 0.8, 1.6, and 3.2 TB capacities.

I'm not seeing stock of new, Intel-branded drives in the channel, so I'm guessing those are either used or at least not via authorized resellers? Probably most datacenter operators wouldn't consider using drives from grey market sources, unless they were desperate.

BTW, the lowest ebay price I can find on a new 800 GB drive is $2.4k (from China - probably not including tariffs) and the 1.6 TB capacity drives start at $3k (from China).
Would've saved me some time.
; )


Agreed. Even though it's an artificial benchmark, it would be nice to have some apples-to-apples way to compare it both with consumer drives and sometimes manufacturers also specify that figure.


Uh, well, they only shipped in 0.4, 0.8, 1.6, and 3.2 TB capacities.

I'm not seeing stock of new, Intel-branded drives in the channel, so I'm guessing those are either used or at least not via authorized resellers? Probably most datacenter operators wouldn't consider using drives from grey market sources, unless they were desperate.

BTW, the lowest ebay price I can find on a new 800 GB drive is $2.4k (from China - probably not including tariffs) and the 1.6 TB capacity drives start at $3k (from China).
800gb
https://www.ebay.com/itm/157084426091?_skw=p5800x+800gb $789
not new, but for fraction of the price on a product that is harder to damage, I feel like it's ok.