Sony's ReRAM plans - what can and can't ReRAM bring to a console? *spawn

[megre]

I'm not sure what you mean by this either. Look up tech like SED TVs and methanol fuel-cell batteries. There are loads of technologies described as just about ready for mainstream that'll revolutionise the world, only to never materialise because of obstacles. Regardless of whatever any company hopes to do, we (and engineers making design decisions) can only go by what actually exists. ReRAM as it is at the moment has not been proven as a viable mainstream storage solution. For the past 20 years, companies have worked on it and the best obtained is small-scale production.

No-one's mass producing it yet, despite companies having been investigating it for a decade, as you say (actually longer). That makes forecasts of cheap, efficient mass production within a year speculative rather than likely. If Sony had been making ReRAM for the past two years and supplying enterprise, and announced they were ramping up production for PS5, that'd be expected to work out well. Sony planning to jump straight in to making tens of millions of chips for a console that they absolutely cannot afford to screw up...that's a risky proposition.

Sony is mass producing STT-MRAM for Avalanche. Sony said they will commercialize ReRam in 2020 going after a well established product from Intel. I don't see how it's a wishful thinking of Sony's part.

Now, will this new product by Sony in 2020 be inside the PS5 though, that's a different question.

Not if they want zero load times. I'd rather have a box that loads any game in 10 seconds than one that loads my current game instantly but takes a minute to copy across a different game..

I assume there will be intelligent precaching. Part of your games resides in the cache enough to quickly launch them. And you don't need to fill the cache to start your game.

NVME SSD are also expensive. And Sony has to buy them from outside source.

A small cache that they make themselves coupled with a low cost qlc nand ssd may not be much more expensive than a top of the line nvme ssd. The 25.6gb/s speed is just an icing on the cake.

So you'll have Sony making $25 chips and putting them in a $500 console while their rivals also making ReRAM will be making $25 chips and selling them at $500 a piece to enterprise users, and Sony's executive and investors will be arguing over why the hell Sony are wasting billions in profits making a console with 'zero load times' when they could have used NAND for 10 second load times and be selling that ReRAM for big bucks.

They can do both. Samsung is using milions of their own OLED screen while also selling them to Apple and other phone makers. I really don't see the point. Supply will adjust according to demand. Sony may even use this for their image sensors. There is both demand within and without Sony.

 

[tunafish]

I'm sorry I can't understand the table. Can you quote the pertinent data?

The chepest SLC in chips is ~3$/GB, the cheapest TLC in chips is ~9 cents per GB. You can go way lower than that in the price for large orders, but the relative cost will stay roughly the same.

Ease of development? No more moving in and out or micromanagement of data every now and then. If the cost is similar to 128gb of slc nand, that's a no-brainer.

A large pool of QLC can have a low read latency. Which is all you need in a console to minimize micromanagement of data movement, and make development easy. Having a pool of SLC alongside would complicate the use of the system, not make it easier.

Writes are either streaming, where having a cache or separate small pool would not help, or rare, and you can choose to not care about the performance of those.

I stand by the idea that there is absolutely no sense in having a write buffer/cache on a console. It provides only things that are strictly not needed in a console environment. I would be willing to bet money on this if you want to.

We can say the same thing in 2015 when Intel announced Optane. I guess the naysayers back then were wrong. We'll see.

Optane was launched to serve an entirely different kind of market.

 

[megre]

The chepest SLC in chips is ~3$/GB, the cheapest TLC in chips is ~9 cents per GB. You can go way lower than that in the price for large orders, but the relative cost will stay roughly the same.

Eh? Optane is just $1.2/gb. We're talking about SSD here, not DIMM.

Optane was launched to serve an entirely different kind of market.

Sony ReRam is going after the same market. Inclusion in the PS5 is still up in the air. But if it's as cheap as SLC Nand then it's a no brainer for me. That 25.6gb/s speed is just a bonus.

Is there no one else here convinced that Sony will use ReRam SSD as cache? Can you help me? Am I alone in this crusade?

 

[Shifty Geezer]

Now, will this new product by Sony in 2020 be inside the PS5 though, that's a different question.

Um, well, no, that's exactly the question we're asking.

They can do both.

Only if they have a surplus of chips to the external market. Let's say Sony can make 1 million chips a month. Let's say the enterprise market is for 1 million chips a month. Every chip Sony stick in PS5 is wasted profits. Now let's say Sony can make 2 million a month and the enterprise market is 1 million a month. Now Sony can make 1 million PS5's as well as serve the more profitable market. It's only if Sony can make surplus to the wider industry that serving their own needs is good business. If, as you say, there's a huge market for ReRAM, then there's good reason to think putting any chips in PS5 is going to deliver the worst ROI versus selling those chips to others.

 

[megre]

There's a difference between making and selling ReRAM, and planning it to be in your console. There's a market for high-performance, low(er) cost persistant storage which is what ReRAM manufacture will serve. Sony will be able to serve that market with whatever amounts of ReRAM they can make, and if ramping up is a slow process, that's not a problem. Unless their console is dependent on it, at which point if they have problems ramping up, we have a blue laser diode situation again, only without the 'future standard' gains to be had. So you'll have Sony making $25 chips and putting them in a $500 console while their rivals also making ReRAM will be making $25 chips and selling them at $500 a piece to enterprise users, and Sony's executive and investors will be arguing over why the hell Sony are wasting billions in profits making a console with 'zero load times' when they could have used NAND for 10 second load times and be selling that ReRAM for big bucks.

I agree it's risky. But the thing about ramping up, I don't see it. This thing is verified to scale below 10nm, scalable, stackable.

 

[PSman1700]

Why would sony risk another ps3, for something no one even cares about? From hdd to say nvme is already massive. The spiderman current gen demo didn’t load instantly either, but the increase was still enough to brag about it?

 

[megre]

Um, well, no, that's exactly the question we're asking.

Um no, the quote I was answering to was questioning the possibility of Sony ever releasing ReRam. I said it's Sony is already mass producing STT-MRAM and going after a well-established persistent memory market by Intel. Commercializing ReRam is not a wishful thinking on Sony's part, I don't think.

Now whether it'll be included in the PS5 is another question.

Why would sony risk another ps3, for something no one even cares about? From hdd to say nvme is already massive. The spiderman current gen demo didn’t load instantly either, but the increase was still enough to brag about it?

If this thing indeed cost similar to SLC Nand SSD then that 25.6gb/s is just icing on a cake.

 

[Shifty Geezer]

Um no, the quote I was answering to was questioning the possibility of Sony ever releasing ReRam. I said it's Sony is already mass producing STT-MRAM...

1) That's not ReRAM. 2) How 'mass produced' is Avalanche's STT-MRAM? What sizes are they available in and at what prices? How do you get from 'Sony making STT-MRAM for Avalanche' to 'Sony are/will be mass producing ReRAM'?

 

[megre]

1) That's not ReRAM. 2) How 'mass produced' is Avalanche's STT-MRAM? What sizes are they available in and at what prices? How do you get from 'Sony making STT-MRAM for Avalanche' to 'Sony are/will be mass producing ReRAM'?

So you're indeed questioning whether Sony will really commercialize ReRam in 2020. That could be anybody's guess. Saying they will do it has a 50% chance of happening while saying they won't has also 50%.

But I believe Sony will commercialize ReRam in 2020 just like their engineer (the inventor of memory stick) has spoken.

Now whether it will be included in the PS5 is another question. (Which I happen to also believe they will.)

 

[PSman1700]

Now whether it will be included in the PS5 is another question.

Everything points to then not doing that. What’s the reason you personally like to see reram in the ps5? The spiderman demo, was that on reram?

 

[tunafish]

Is there no one else here convinced that Sony will use ReRam SSD as cache? Can you help me? Am I alone in this crusade?

You are alone because every programmer here can tell you that console SSDs do not need caches.

Caches on storage does not help read speed, or streaming write speed, only random write speed. PCs need the caches because every program on the platform is written to expect relatively fast random writes.

The solution to random writes on console storage is: "just don't do that". There is no actual need to ever do it, no situation where they are indispensable, so you can optimize the whole system by just leaving that out. Consoles are made better by leaving out the write cache. Not only does it save on cost and power, but having a write cache actually means you need to do more work on reads, and thus slows them down.

If you feel like a console should have a write cache, please explain in what sort of use case would it actually be useful or desirable?

Eh? Optane is just $1.2/gb. We're talking about SSD here, not DIMM.

Those are spot prices. As I said, with large, pre-planned orders, prices go down, but the relative price difference between SLC and TLC remains.

 

[Shifty Geezer]

So you're indeed questioning whether Sony will really commercialize ReRam in 2020.

No. I'm questioning whether they can commercialise it effectively in time and cost effectively enough to feature in a cheap console. 'Up and running in 2020' could mean producing a couple million chips a month starting from October, too late to be used in PS5.

You're bringing a lot of faith into this discussion which doesn't fit well with the data-driven debate. We can show you loads of examples of technologies with roadmaps that either didn't release or released late or struggled to keep costs down, but you seem to be basing your idea on everything working perfectly. Step back from your hopes and dreams for a second and look at the rational argument. Is it more likely that:

1) PS5's designers will base their design on the availability of a part that does not yet exist in the hopes it'll all work without hitch and cost effectively and a few months before PS5 is ready, the plant is up and running producing ReRAM chips en masse sufficient to both serve the lucrative enterprise market and to satisfy PS5 production.

or

2) PS5's designers will play it safe, use existing NAND tech that'll already provide a generational advance in storage performance, is known to be cost effective, is known to be readily available on time and won't limit PS5 production rate, and the memory division will be free to serve the memory market as and when they are up and running and able to.

If we hear ReRAM is being mass produced economically six+ months before PS5 is supposed to release, then we can consider it as a viable option. Until then, the weight of logic is against it.

 

[megre]

Those are spot prices. As I said, with large, pre-planned orders, prices go down, but the relative price difference between SLC and TLC remains.

No. That $1.2/gb of optane is retail price with retail mark-up and intel profit.

No. I'm questioning whether they can commercialise it effectively in time and cost effectively enough to feature in a cheap console. 'Up and running in 2020' could mean producing a couple million chips a month starting from October, too late to be used in PS5.

I cannot give you a data to prove that Sony can commercialize it in time and cost effectively enough to feature in a console. All I know is Sony said they will commercialize in 2020 and the professional analysis of its pricing approximating that of SLC Nand. Is that enough to prove that Sony will use ReRam in PS5? I'm not saying it's enough of a proof. But I'm saying it's enough to warrant speculation. And in my book, the probability is high.

You are alone because every programmer here can tell you that console SSDs do not need caches.

Really? That's fascinating.

 

[iroboto]

I agree it's risky. But the thing about ramping up, I don't see it. This thing is verified to scale below 10nm, scalable, stackable.

You’ve done good research. But a business person would look at the next best technology that is guaranteed to work and see the risk factor drop a great deal of many points.

Sony doesn’t need ReRAM to beat MS next gen; right now it’s Sony’s generation to lose. if Sony chooses ReRAM and screws the pooch, that could be a disaster. I don’t think the executive team would greenlight a high risk situation when you are winning. the payoff has to be BIG. Really big. Much bigger than PlayStation. Much bigger than loads of things.

Really? That's fascinating.

Honestly; nvme is fairly massive. Standard Sata 3 is like supports 1 thread or something. Very few I/O requests in parallel.
nvme supports 64K of them IIRC in parallel.

That type of parallel I/O threads happening in action can max out the bandwidth, in a way that can totally speed up loading. As long as it’s multithreaded.

 

[megre]

I don’t think the executive team would greenlight a high risk situation when you are winning. the payoff has to be BIG. Really big. Much bigger than PlayStation. Much bigger than loads of things.

Fair enough. How about the importance of first to market and commercialize? If Sony saturates the market with low priced ReRam in the future, it'll be hard to new comers to enter. See how Samsung has almost monopolized small screen oled market for years. It's only now that a Chinese company is giving Samsung a competition.

Honestly; nvme is fairly massive. Standard Sata 3 is like supports 1 thread or something. Very few I/O requests in parallel.
nvme supports 64K of them IIRC in parallel.

That type of parallel I/O threads happening in action can max out the bandwidth, in a way that can totally speed up loading. As long as it’s multithreaded.

I'm not a developer. I'll take your word for it. But now I wonder why Sony even attempted to include an ssd with raw bandwidth higher than any ssd on pc. Cerny said that when pc ssd tops at 3gb/s, so at least what cerny has is more than that. Is that the optimum bandwidth for SSD? Is going more than 4gb/s an overkill? Why even shoot for 4gb/s, why not go for an even cheaper say 2gb/s?

So going for 25.6gb/s with fast latency with possibly amd ssg configuration does not give benefit. But going from 2gb/s to 4gb/s does? It doesn't make sense from a business point of view to choose an expensive ultra-fast ssd when a small bandwidth will suffice? What's 5-10 seconds more of loading time, is that even a selling point?

 

[Shifty Geezer]

I cannot give you a data to prove that Sony can commercialize it in time and cost effectively enough to feature in a console. All I know is Sony said they will commercialize in 2020 and the professional analysis of its pricing approximating that of SLC Nand. Is that enough to prove that Sony will use ReRam in PS5? I'm not saying it's enough of a proof. But I'm saying it's enough to warrant speculation. And in my book, the probability is high.

Which is fine, but without more to discuss, the discussion needs to end with at 'agree to disagree'. Here, we'll have posit and then arguments and then probabilities as we decide what is and isn't likely. There's no more speculation to be had here, or arguments in favour of a ReRAM design, until there's data that can change the argument.

Fair enough. How about the importance of first to market and commercialize? If Sony saturates the market with low priced ReRam in the future, it'll be hard to new comers to enter.

They'd also wipe out their potential profits. A new tech will be sold at a premium in reduced amounts, until natural ramping up due to competition. To don't dump your stuck onto the market cheap just to own it when you can control supply to keep price and profits up.

I'Is going more than 4gb/s an overkill? Why even shoot for 4gb/s, why not go for an even cheaper say 2gb/s?

It depends on the cost/benefit ratio. If twice the speed is 10% more expensive, it may well be worth it. If 10% more speed is twice the price, it definitely isn't.

So going for 25.6gb/s with fast latency with possibly amd ssg configuration does not give benefit.

Again, it's cost/benefit. Of course 25.6 GB/s is better, but if it's also vastly more expensive, it's a poor choice for a console. It's very important to note that load times isn't just transfer form storage, but game processing. The CPU has to process workloads to generate the world form the assets loaded. I pointed out earlier that straight transfer of content would be very fast, but as others rightly argued, that doesn't mean load times will be fastest transfer possible. PS5's solution isn't just storage hardware but IO stack and keeping everything as lean as possible. 4GB/s is plenty of data to process and load, capable of filling RAM in a few seconds. The bottleneck is more likely to be processing speeds working through that data.

 

[PSman1700]

But now I wonder why Sony even attempted to include an ssd with raw bandwidth higher than any ssd on pc.

He said at the time it was faster then anything pc. This could be the result of PCIe3.0 vs 4.0. They are careful with their wording but still can be perceived as the fastest thing ever.

 

[megre]

Which is fine, but without more to discuss, the discussion needs to end with at 'agree to disagree'. Here, we'll have posit and then arguments and then probabilities as we decide what is and isn't likely. There's no more speculation to be had here, or arguments in favour of a ReRAM design, until there's data that can change the argument.

I guess the discussion has to end with pessimism which is fair because that is the status quo. ReRam won't be in the PS5 until proven otherwise. I had the itch to counter the pessimism with optimism maybe because I want ReRam to happen. Ok I'll let you have the last word with your pessimism then.

A new tech will be sold at a premium in reduced amounts, until natural ramping up due to competition. To don't dump your stuck onto the market cheap just to own it when you can control supply to keep price and profits up.

Sorry I can't help myself. Optane is already competition in the enterprise market. PS5 will give ReRam the momentum early as opposed to relying solely on the enterprise market.

Of course 25.6 GB/s is better, but if it's also vastly more expensive, it's a poor choice for a console.

At the price similar to SLC Nand (hypothetically, I'm not saying the analysis is accurate), is it high on the cost/benefit ratio in your book?

He said at the time it was faster then anything pc. This could be the result of PCIe3.0 vs 4.0. They are careful with their wording but still can be perceived as the fastest thing ever.

That's why I limited myself to 4gb/s because I'm expecting this argument. Anyway, it's a cost/benefit ratio like Shifty already said.

 

[Shifty Geezer]

I guess the discussion has to end with pessimism which is fair because that is the status quo. ReRam won't be in the PS5 until proven otherwise. I had the itch to counter the pessimism with optimism maybe because I want ReRam to happen. Ok I'll let you have the last word with your pessimism then

It's not pessimism as there's no good or bad outcome. This is a logic discussion based on logic, without regard for the outcome. Like going to a horse race and trying to determine which horse will win based on form and conditions but without placing a bet, so you don't care whether the horse wins or loses other than to see if your reasoning was correct or not.

There is nothing to be gained from this discussion from convincing oneself some awesome tech is going to feature only for it not to. Likewise, there's nothing to be gained from expecting a tech won't feature only for it to do so. The joy of this discussion is in the discussion itself and working through the logic of the possibilities, like a puzzle.

PS5 will give ReRam the momentum early as opposed to relying solely on the enterprise market

Momentum to do what? Why would it need additional momentum beyond being cheaper than Optane?

At the price similar to SLC Nand (hypothetically, I'm not saying the analysis is true), is it high on the cost/benefit ratio in your book?

If you can get something 6x the speed for the same price, sure. Not sure what good picking hypotheticals is for grounded speculation.

 

[megre]

If you can get something 6x the speed for the same price, sure. Not sure what good picking hypotheticals is for grounded speculation.

I mean cost/benefit ratio compared to top of the line 1tb nvme ssd. If ReRam is about similar in price to SLC Nand SSD, does it have an advantage in cost/benefit ratio if used as a cache say 128GB with a 1tb low-cost ssd against a high bandwidth 1tb nvme ssd?

Momentum to do what? Why would it need additional momentum beyond being cheaper than Optane?

I'm speculating here. I'm guessing Sony would want to eat into the nand chips market share little by little. They have a roadmap to stack and scale and give nand flash a run for its money.