Kioxia's PCIe 5.0 SSD Hits 14,000 MBps

[Forbidden Wafer]

6 minutes ago, RejZoR said:

Selling Samsung 850 Pro 2TB now would be total waste of money as I'd get peanuts for it yet it's probably better than most crappy cheap NVMe drives sold today and comes in still very respectable 2TB capacity. 

And your ssd will probably outlive the new ones. I know my 8yo Kingston ssd still kicking ass, with smart reporting it's nowhere near the endurance limits 20-30%), while my new nvme drive is at 4% in 8 months

[PeachGr]

Well yeah, there are different but it seems like unifying is a thing. We said the same for ram and VRAM, but Apple runs unified memory, phones offer extended ram etc. I .ran, there are different memory tips, as well as different processors, so we can't compare them 1-1, but they usually follow one another

[cj09beira]

i would love to see latency wars please, instead of bandwidth, thanks.

[porina]

2 hours ago, PeachGr said:

Well yeah, there are different but it seems like unifying is a thing. We said the same for ram and VRAM, but Apple runs unified memory, phones offer extended ram etc. I .ran, there are different memory tips, as well as different processors, so we can't compare them 1-1, but they usually follow one another

DDR ram and VRAM in the form of GDDR is essentially the same thing. Unified makes sense especially in integrated GPU systems, and doesn't make any sense for dGPUs if you care at all about performance. HBM would be different though, it trades bandwidth somewhat with latency, so not quite as direct a replacement for DDR.

 

On phones, perhaps I'm out of date, what does extended ram mean? I've only seen where storage can be expanded by an SD card for example, but neither of those would class as ram. 

 

14 minutes ago, cj09beira said:

i would love to see latency wars please, instead of bandwidth, thanks.

Why not both?

 

In that area I think the first thing that needs to change is the storage medium. Flash doesn't cut it. Optane is far superior for example. Even in consumer class products and systems I was seeing over 3x better latency compared to high end flash SSD, over 5x compared to a more average one. Once the medium has been improved, then the protocol is next to be looked at. NVMe was never designed for ram-like performance, but I don't know what is there to replace it. Maybe some CXL implementation? I haven't looked at that in detail.

[cj09beira]

40 minutes ago, porina said:

DDR ram and VRAM in the form of GDDR is essentially the same thing. Unified makes sense especially in integrated GPU systems, and doesn't make any sense for dGPUs if you care at all about performance. HBM would be different though, it trades bandwidth somewhat with latency, so not quite as direct a replacement for DDR.

 

On phones, perhaps I'm out of date, what does extended ram mean? I've only seen where storage can be expanded by an SD card for example, but neither of those would class as ram. 

 

Why not both?

 

In that area I think the first thing that needs to change is the storage medium. Flash doesn't cut it. Optane is far superior for example. Even in consumer class products and systems I was seeing over 3x better latency compared to high end flash SSD, over 5x compared to a more average one. Once the medium has been improved, then the protocol is next to be looked at. NVMe was never designed for ram-like performance, but I don't know what is there to replace it. Maybe some CXL implementation? I haven't looked at that in detail.

CXL was made to be able to have ram, so one would hope its fast enough, hopefully with cxl coming to the server market next year, optane gets a second wind, and finally sees better prices with better economies of scale.

[leadeater]

49 minutes ago, porina said:

Why not both?

Technically improving latency does both, block size for block size.

 

49 minutes ago, porina said:

In that area I think the first thing that needs to change is the storage medium. Flash doesn't cut it.

Depends, there's a ton of latency improvements that can be made with NAND and Flash, as well as current massive differences in the market now based on workload and market segment target for the product. You can have one SSD doing low to sub ms and another with higher latency than even an HDD on identical workloads. It's just really expensive to latency optimize an SSD if that's what you need to care about.

 

Like most things, pay for what you get. There's a reason $100, $200, $500 and $1000+ 1TB SSD exist and they are only equal in capacity.

[Alder Lake]

15 hours ago, Zodiark1593 said:

The two big obstacles to this use are latency, and endurance. 
 

RAM chips can be written virtually an indefinite amount of times without fail. NAND has a definitive endurance limit. 
 

The two technologies are also written and read in different ways. Nand is written in blocks, and when overwritten, must be done as a block, while RAM can allow for direct access and modification to a given value. 
 

Raw bandwidth is only a part (albeit, not an insignificant one) of the story.

 

For specific applications, it may be possible to treat portions of storage as an extension of read-only RAM, similar to cartridge games, which can greatly cut overall RAM usage. So long as you’re not writing to the area like RAM, you sidestep most of the downsides of NAND storage. 

22 hours ago, Alder Lake said:

Read latency has dropped from 90us to 70us and write latency from 20us to 10us.

well, latency is getting better

[williamcll]

21 hours ago, Lurick said:

Do I need PCIe 5.0 for Microsoft word?!?!?!?!?!

yes because the load times has gotten more worse per generation.

[Alder Lake]

Just now, williamcll said:

yes because the load times has gotten more worse per generation.

isn't that pretty normal for any program (newer version = more features = more data = longer load time)

[Ydfhlx]

22 hours ago, PeachGr said:

I see nvme drives to replace ram in the future on office desktops, or work as ram extensions, same way they do on the recent phones

Always have been working like that. Phone manufacturers only recently found out that they can say they have more RAM than they do. Question is if you will choose to have less ram and not worry of using it all. IMO not - RAM gets faster too (DDR5 is new faster standard after all), so likely there will be a speed difference. And if you can afford PCIe5 SSD, why not have more ram anyways?

[StDragon]

32 minutes ago, williamcll said:

yes because the load times has gotten more worse per generation.

The problem isn't fetching the data (program) from storage. The delay is in processing it once loaded into RAM. Don't blame the HW, blame Microsoft.

[Alder Lake]

22 hours ago, PeachGr said:

I see nvme drives to replace ram in the future on office desktops, or work as ram extensions, same way they do on the recent phones

there is one thing, what if RAM replaces nvme drives instead?

 

 

at the end of this generation (DDR5), DIMMs will have as much as 128GB (I think), so with 4 slots, you can have half a terabyte of RAM in your system, and next GEN that might double (or more) again

 

it's almost like a battle of, "will RAM have enough storage to act as a storage, or will NVME drives have enough speed to act as RAM?"

[Stahlmann]

So the basic TLDR in the SATA vs. NVME already was "you won't spot a difference". Then there was PCIe 4 and it started all over again with the same conclusion. I wonder how PCIe 5 drives will end up  Don't kid yourself, it will still feel the same as SATA SSDs unless you're in a very niche user base that actually needs fast sequential drives.

[porina]

2 hours ago, leadeater said:

Technically improving latency does both, block size for block size.

You could say the same for transfer rate too. If access latency remains constant, a faster transfer rate would mean the transfer part completes sooner.

 

2 hours ago, leadeater said:

Depends, there's a ton of latency improvements that can be made with NAND and Flash

For sure there is a variation in how it is implemented and performs, but fundamentally it is limited by what it is. To go really beyond that means looking at something different.

 

1 hour ago, Alder Lake said:

it's almost like a battle of, "will RAM have enough storage to act as a storage, or will NVME drives have enough speed to act as RAM?"

One big problem is that ram is volatile. You will have to either guarantee power is always available, or still have some other storage anyway. And that's ignoring the cost will always be more than flash. If you want something that has qualities of both flash and ram, that remains Optane. As a generalisation, per capacity it is closer to ram than flash. Performance is also in between the two, but really excels over flash when it comes to access latency and consequently random access. It doesn't use large pages like flash so can be used ram-like for general applications. Endurance might remain a concern. It is much higher than flash, but still not practically unlimited like for ram.

 

In limited cases where those considerations can be controlled, you could make it work. For general uses, we're probably not moving away from the current model for a long time yet.

[PeachGr]

Ram sticks to change maybe, but as ram, by definition you can't store anything. They are all deleted after shutting down the PC

 

[Alder Lake]

Just now, PeachGr said:

Ram sticks to change maybe, but as ram, by definition you can't store anything. They are all deleted after shutting down the PC

 

I mean, that could be changed, but I agree that at the current state, RAM has absolutely no chance to replace storage for extremely obvious reasons like that

[PeachGr]

1 hour ago, Ydfhlx said:

Always have been working like that. Phone manufacturers only recently found out that they can say they have more RAM than they do. Question is if you will choose to have less ram and not worry of using it all. IMO not - RAM gets faster too (DDR5 is new faster standard after all), so likely there will be a speed difference. And if you can afford PCIe5 SSD, why not have more ram anyways?

I have found my self choosing between those two in rendering solutions. If you want your assets to load faster, you use ram, if you don't have enough ram, you let them load from disk. So I said office use, because for gaming is always "more faster, more better"

[PeachGr]

3 hours ago, porina said:

n phones, perhaps I'm out of date, what does extended ram mean?

As you said, they are on a short of integrated GPU, and it came recently on some android phones, like Samsung and Xiaomi ones

[flibberdipper]

17 hours ago, Forbidden Wafer said:

And your ssd will probably outlive the new ones. I know my 8yo Kingston ssd still kicking ass, with smart reporting it's nowhere near the endurance limits 20-30%), while my new nvme drive is at 4% in 8 months

That could change with a CrystalDiskInfo update, though (assuming you use it). For a long time my 840 EVO was sitting at 100% health, then a CDI update came along and knocked it down to 79%.

[Alder Lake]

6 minutes ago, PeachGr said:

Ram sticks to change maybe, but as ram, by definition you can't store anything. They are all deleted after shutting down the PC

 

watch DDR6 have a solution for this, something like "semi-random-access-memory" SRAM (unless that's already taken), where it would have 128gb per stick, and would be somewhere between the speeds of RAM and NVME, and stay even after a complete shutdown, that way things are still snappy on startup

[Forbidden Wafer]

6 minutes ago, flibberdipper said:

That could change with a CrystalDiskInfo update, though (assuming you use it). For a long time my 840 EVO was sitting at 100% health, then a CDI update came along and knocked it down to 79%.

Mine have been below 100% for a long while, so I hope that won't be the case.

[porina]

5 minutes ago, Alder Lake said:

watch DDR6 have a solution for this, something like "semi-random-access-memory" SRAM (unless that's already taken), where it would have 128gb per stick, and would be somewhere between the speeds of RAM and NVME, and stay even after a complete shutdown, that way things are still snappy on startup

SRAM is taken: Static RAM. However it is still volatile. It differs from DRAM in that it doesn't need to be constantly refreshed in use. It is higher performance, at higher cost. I think it is used in things like the higher performance CPU caches.

 

DRAM is DRAM so that part of it isn't going to change. It will take a fundamentally different technology to get around it.

[Alder Lake]

5 minutes ago, porina said:

SRAM is taken: Static RAM. However it is still volatile. It differs from DRAM in that it doesn't need to be constantly refreshed in use. It is higher performance, at higher cost. I think it is used in things like the higher performance CPU caches.

 

DRAM is DRAM so that part of it isn't going to change. It will take a fundamentally different technology to get around it.

thought it would be

 

like I said, we'll just have to see what happens in the end

[Zodiark1593]

2 hours ago, Alder Lake said:

well, latency is getting better

Keep in mind, RAM is measured on the nanoseconds scale, and often single-digit numbers at that. There’s a 1000x difference from 10 ns to 10 μs. 

[Alder Lake]

2 minutes ago, Zodiark1593 said:

Keep in mind, RAM is measured on the nanoseconds scale, and often single-digit numbers at that. There’s a 1000x difference from 10 ns to 10 μs. 

yeah, but the point was how it's getting better

 

 

I don't see either of them replacing each other for another 30 years, maybe longer, but it's not inconceivable that one will be replaced at some point