Digital Foundry Article Technical Discussion [2023]

Status
Not open for further replies.
Cache Scrubbers for the PC Architecture is worthless because the memory pools are not coherent between CPU and GPU. They would simply be a waste of silicon.
 
in regards to Asus ROG Ally, while it might not have that "console" feel like the Steam Deck, running Windows has a lot of advantages that the Steam Deck doesn't have, specially games' and stores' compatibility.

That doesnt detract from the fact that you can convert it into just a gaming machine, like I did with my desktop computer, running two partitions, one for regular use, and one focused on gaming. If Asus can automate that without user's intervention that's good news.
 
Which is still the latency of the drive itself. None of those things make the access latency of the drive any lower. If it takes X time units in order to access data on the SSD, it will at best take X time units to access data on the SSD. All of those IO subsystems don't change that.

Regards,
SB

Of course it doesn't and I never claimed this. I am talking about system latency from storage to VRAM.

If there is ever a day that there is significant enough bandwidth to render from SSD you would be right. But realties have it that L0 is much faster than L1 which is much faster than L2 whihc is much faster than L3. Which is magnitudes orders faster than GDDR which is magnitudes faster than SSD.
Faster SSD releases footprint pressure on GDDR, but it’s not a replacement for it. Removing latency is a nice trick for when you are trying to cut things as close as possible with using the least amount of footprint as possible. For everyone else, they would just use a little more footprint, or slow things down just enough to make latency.

I never said it can or will replace VRAM. I am strictly talking about latency from storage to main memory. It's also strange for you to describe tech that serves to reduce latency a "nice trick".

It's more accurate to say "the decompression chip is chiefly responsible for reducing CPU requirements while achieving effective bandwidth targets". It doesn't in itself do anything to increase effective bandwidth, the compression scheme is doing that. The decompression unit is just the tool (one of several options) being used to decompress the data stream.

It absolutely increases effective bandwidth because effective bandwidth considers latency into the equation. Games like Demon Souls, Ratchet with specific portal transitions, and Last of Us Pt. 1 would be impossible to run on PS5 at that level of fidelity without the hardware decompression; the CPU would get demolished. Perhaps you meant to say "it doesn't do anything to increase bandwidth" - and my response would be: I never said it did.


Here's the thing though.. and let's see if you'll admit this. Your entire premise of Sony's superiority.. lies in the fact that it depends on developers being able to ignore the other platforms and code the games specifically around the PS5's I/O and memory architecture.. right?

Of course I agree with this. Correct me if I'm wrong but didn't this conversation originate from the context of PS5 games ported to PC?

Well then I hate to break it to you... but if I said the same thing and developers were instead able to code specifically around PC's I/O and memory architecture, then our raw SSD bandwidth, and bigger memory pools.. becomes even more realized and less theoretical than the PS5's. Developers could pre-load and cache far more assets in memory up front if they wanted to, and during gameplay streaming they could pre-fetch much further in advance because they can hold many more assets resident in memory.

Why do you "hate to break it to me" as if my life's happiness is dependent on PS5 architecture reigning supreme? Anyways, I would agree with you that performance would be better, however that defeats the purpose of being more efficient with memory. Regardless of what platform you prefer, the less inactive data stored in main memory, the better. It opens up much more possibilities from a game design perspective. It is wasteful to preload assets.


General observation - It is very strange to see people's excitement about the future possibilities of DirectStorage, while simultaneously downplaying the PS5's i/o architecture.
 
Of course I agree with this. Correct me if I'm wrong but didn't this conversation originate from the context of PS5 games ported to PC?



Why do you "hate to break it to me" as if my life's happiness is dependent on PS5 architecture reigning supreme? Anyways, I would agree with you that performance would be better, however that defeats the purpose of being more efficient with memory. Regardless of what platform you prefer, the less inactive data stored in main memory, the better. It opens up much more possibilities from a game design perspective. It is wasteful to preload assets.


General observation - It is very strange to see people's excitement about the future possibilities of DirectStorage, while simultaneously downplaying the PS5's i/o architecture.
Of course you agree... because you know it's true. You're not speaking like as if it's a matter of code... you're speaking as if it's a matter of pure architectural differences, in which console is infinitely more efficient and of which PC isn't capable of. NOBODY here has said consoles aren't more efficient...

And I said that because it really does seem like your only purpose on this forum is to play up the PS5's supposed strengths and play down the PCs.

To your statement of having less inactive data in main memory being better... well I disagree. In fact I think the more memory you can fill up from the start, the more possibilities there are from a game design perspective. It's only wasteful to pre-load assets if you don't have the space to... and yet every single developer would preload more assets... if the consoles actually had the memory to do it. RAM to VRAM latency on PC is far less than SSD > RAM latency on PS5.. so the more upfront loading you can do the better.. reducing streaming requirements while in game. However, on the PC.. they could actually do both.. since they have the memory capacity to spare, and the speed to stream in tons of data as you play.

The fast SSD bandwidth of the PS5.. is a compromise made by Sony because they can't afford to throw 24 or 32GB of RAM in the console.. So while Sony made careful decisions to optimize the I/O capabilities to maximize the potential of the console with just 16GB of RAM.. that doesn't mean that it's wasteful to utilize more if you have it. The lower the amount of RAM, the faster you can fill it.. so in Sony's case it made more sense to keep the lower 16GB of RAM and improve the I/O of the console to ensure that it can be fed as quickly as possible.

That's efficient.. but it doesn't mean it's the best. People aren't downplaying the PS5's I/O... you only feel that way because you have been overhyping it... which you're also only doing because you feel like the PS5 has something that nothing else has. It's an amazingly impressive kit for $500... but PC is capable of more.
 
I never said it can or will replace VRAM. I am strictly talking about latency from storage to main memory. It's also strange for you to describe tech that serves to reduce latency a "nice trick".
If we’re referring to the cache scrubbers? Or do you mean just latency on nvme in general?

Cache scrubbers are a nice trick. A nice to have, but not one that developers can’t live without.
 
Of course you agree... because you know it's true. You're not speaking like as if it's a matter of code... you're speaking as if it's a matter of pure architectural differences, in which console is infinitely more efficient and of which PC isn't capable of. NOBODY here has said consoles aren't more efficient...

And I said that because it really does seem like your only purpose on this forum is to play up the PS5's supposed strengths and play down the PCs.

There's nothing I can really say here. I can't help how you feel, even if those feelings are invalid.

To your statement of having less inactive data in main memory being better... well I disagree. In fact I think the more memory you can fill up from the start, the more possibilities there are from a game design perspective. It's only wasteful to pre-load assets if you don't have the space to.

You're wrong and I would be mildly surprised if many others here agree with you on this. I don't care if you have 1,000gb of VRAM, preloading would still be a wasteful practice if you had a capable memory subsystem and engine to ensure data can be sent to VRAM when it's needed. You talk as if memory, specifically VRAM, is infinite. As if decently spec'd GPUs haven't been incapacitated by sizeable increase in VRAM requirements. At the core, the idea behind PS5 i/o is to ensure the GPU is the bottleneck and not the ability to move data around as needed. You need not look further than the 3070 to determine whether this was a smart design choice.

Yea, it will be amusing watching people keep pointing to the next thing that they believe can't be done on PC... and then it happens..

Well by 2025 I would hope so.

The fast SSD bandwidth of the PS5.. is a compromise made by Sony because they can't afford to throw 24 or 32GB of RAM in the console.. So while Sony made careful decisions to optimize the I/O capabilities to maximize the potential of the console with just 16GB of RAM.. that doesn't mean that it's wasteful to utilize more if you have it. The lower the amount of RAM, the faster you can fill it.. so in Sony's case it made more sense to keep the lower 16GB of RAM and improve the I/O of the console to ensure that it can be fed as quickly as possible.

Oh I see now, you're talking about system RAM. Well since you've already written me off as some sort of die-hard console fan, perhaps you will accept the messaging if it's coming from the Godfather of Digital Foundry, for which this thread is dedicated to:

 
There's nothing I can really say here. I can't help how you feel, even if those feelings are invalid.
Indeed there's nothing you can say. So instead of babbling nonsense, just don't say anything at all.

You're wrong and I would be mildly surprised if many others here agree with you on this. I don't care if you have 1,000gb of VRAM, preloading would still be a wasteful practice if you had a capable memory subsystem and engine to ensure data can be sent to VRAM when it's needed. You talk as if memory, specifically VRAM, is infinite. As if decently spec'd GPUs haven't been incapacitated by sizeable increase in VRAM requirements. At the core, the idea behind PS5 i/o is to ensure the GPU is the bottleneck and not the ability to move data around as needed. You need not look further than the 3070 to determine whether this was a smart design choice.
No I'm not wrong. If you have the memory... you can do more interesting things if you fill it....

I'm not speaking as if VRAM is infinite... I'm simply speaking to the fact that there's more of it on PC, and developers can utilize it. Decent spec'd PCs "incapacitated" by increase in VRAM requirements has everything to do with games being designed around consoles... VRAM that games "allocate" isn't the same thing as VRAM that games require. You should learn that.. people mix that up all the time. Like say TLOU P1... By your own measure, that game PROVES that GPUs are VRAM limited... except we have games like A Plague Tale which demonstrates either comparable or better texture quality, which runs well on lower spec'd machines in a fraction of the memory... I've noticed that a lot of these games in which people are complaining about VRAM requirements are typically Sony games ported from console without the architecture of the PC in mind when they were made.

Another thing... you are not informing anyone *anything* on this forum about the PS5 I/O system that they didn't already know.

Well by 2025 I would hope so.
It could happen right this minute... or even years ago we could have had games doing what PS5 will do later this year. It's not our issue that Sony ports their games later.

Oh I see now, you're talking about system RAM. Well since you've already written me off as some sort of die-hard console fan, perhaps you will accept the messaging if it's coming from the Godfather of Digital Foundry, for which this thread is dedicated to:

Consoles only have one type of RAM...........................
 
Last edited:
Indeed there's nothing you can say. So instead of babbling nonsense, just don't say anything at all.

Such bitterness over video games. Pathetic.

No I'm not wrong. If you have the memory... you fill it....

Yes, with active data. You said "pre-load", meaning inactive - don't be dense.

Another thing... you are not informing anyone *anything* on this forum about the PS5 I/O system that they didn't already know.

Your lack of basic knowledge and understanding suggests otherwise.

It could happen right this minute... or even years ago we could have had games doing what PS5 will do later this year. It's not our issue that Sony ports their games later.

You have now entered delusional territory. Sad.

Consoles only have one type of RAM...........................

Instead of admitting you were wrong to think more RAM would solve the issue for PC, you throw out a meaningless comment that has nothing to do with the topic we're discussing. Because surely you can't be foolish enough to state Sony "compromised" by passing on 32gb of unified GDDR memory. Please tell me you're not that clueless.
 
-Such bitterness over video games. Pathetic.
-Yes, with active data. You said "pre-load", meaning inactive - don't be dense.
-Your lack of basic knowledge and understanding suggests otherwise
-You have now entered delusional territory. Sad.
-Instead of admitting you were wrong to think more RAM would solve the issue for PC, you throw out a meaningless comment that has nothing to do with the topic we're discussing. Because surely you can't be foolish enough to state Sony "compromised" by passing on 32gb of unified GDDR memory. Please tell me you're not that clueless.
-More Nonsense
-With the data you need... plus data you will need soon.... You can do MORE things by having the data you immediately need... AND the data that you will need very soon... already resident in memory (System RAM for PC) instead of trying to pull it off disk right when you need it
-No, it's just you and your lack of comprehension
-You're the one who's delusional to think that PCs haven't had this capability for years now..
-In the quote you posted I'm talking about console memory... you literally bolded it... Don't blame me that you completely misunderstood what was said and posted a video attempting to explain something completely different. As for that DF question... having more system RAM isn't a replacement for VRAM... having more system RAM can instead reduce the bottleneck of bringing data off storage which is even slower.

So nobody here is saying that on PC more RAM replaces VRAM... so stop outright lying about what I'm saying.

And yes Sony 100% passed on 32GB of RAM because it was too cost prohibitive for them.... If they COULD have increased their memory capacity over the PS4 by 4x.. they would have. They instead compromised by devising a way to maximize the effectiveness of the 16GB of RAM they would have, by increasing the bandwidth of the SSD to 5.5GB/s.. which is what they settled on as being appropriate for next gen games.. If the PS5 had 32GB of RAM... there would not be the 5.5GB/s SSD that is in there today.. they would have an SSD... but it wouldn't be as robust.
 
Last edited:
Of course it doesn't and I never claimed this. I am talking about system latency from storage to VRAM.

I never said it can or will replace VRAM. I am strictly talking about latency from storage to main memory. It's also strange for you to describe tech that serves to reduce latency a "nice trick".

You keep mentioning latency while ignoring earlier posts that have specifically explained why there should be no appreciable latency difference between a PC and PS5 in getting data from SSD into VRAM. And in fact how their can be considerably less latency when properly using the sys RAM to pre-cache data from the SSD.

It absolutely increases effective bandwidth because effective bandwidth considers latency into the equation. Games like Demon Souls, Ratchet with specific portal transitions, and Last of Us Pt. 1 would be impossible to run on PS5 at that level of fidelity without the hardware decompression; the CPU would get demolished. Perhaps you meant to say "it doesn't do anything to increase bandwidth" - and my response would be: I never said it did.

Ok, so we seemingly now agree that the decompression block does nothing to increase the PS5's peak IO capability. But what it does do, is offload work from the CPU, meaning that when used, there is more CPU power available for other things when data transfers are taking place. So as a system, it does increase the PS5's overall capability.

So I hope that you can agree that if the work being done on the PS5's decompression block were to be done on some other unit in a PC that has a sufficient surfeit of performance vs it's PS5 counterpart, then that would not result in any specific peak IO performance advantage for the PS5 vs said PC. It would however give the PS5 an efficiency advantage in terms of the level of compute resources needed to reach that peak IO performance.

General observation - It is very strange to see people's excitement about the future possibilities of DirectStorage, while simultaneously downplaying the PS5's i/o architecture.

I don't think anyone here is downplaying the PS5's IO at all. Everyone is quite aware of how it works and what it can do. The issue is with it being overplayed relative to what PC's can be capable of (seemingly based on a single bad port) and because this is a technical forum which deals in facts rather than hyperbole, you get people trying to restore technical balance back into the discourse. At no point for example has anyone challenged the PS5's peak throughput of ~11GB/s (with RDO), or the fact that it's decompression block offloads the decompression work from the CPU which potentially frees up a lot of CPU performance, or how relatively easy the PS5 is to develop for.

The issues start when people begin using the above facts to make misleading or outright inaccurate claims about what "can't be done on PC's" because of "fundamental architectural inferiorities" while ignoring technologies in the PC space that address those shortcomings and alternate way's of doing things on the PC which are better suited to it's particular architecture.

You're wrong and I would be mildly surprised if many others here agree with you on this. I don't care if you have 1,000gb of VRAM, preloading would still be a wasteful practice if you had a capable memory subsystem and engine to ensure data can be sent to VRAM when it's needed. You talk as if memory, specifically VRAM, is infinite. As if decently spec'd GPUs haven't been incapacitated by sizeable increase in VRAM requirements. At the core, the idea behind PS5 i/o is to ensure the GPU is the bottleneck and not the ability to move data around as needed. You need not look further than the 3070 to determine whether this was a smart design choice.

You seem to be operating under the misapprehension that the PS5 does not need to cache data in VRAM/main memory. And that it's IO is fast enough to call data directly from the SSD when it's needed by the GPU. That's wrong. The bandwidth and latency from the SSD are no where near sufficient to allow this, and the PS5, just like the PC needs to anticipate what will be needed and then pull data in from storage to cache it in VRAM ready for it to be brought in from there into the GPU's caches and ultimately registers to be worked on.

Obviously the faster you can bring data in from storage, the less you have to cache, for a given scenario, but you're still going to be limited to what is available in your VRAM (and GPU's caches at any given time). For example if the PS5 had 200GB of VRAM and could literally store the entire uncompressed game in VRAM, that would open possibilities that simply aren't available to it in it's current form.

Your comments above about the 3070 are a great example of your misapprehension. The issue with the 3070 is nothing to do with IO, it's to do with it's lack of VRAM. The 3070 can bring data into it's VRAM from system memory far faster than the PS5 can bring data into it's VRAM from the SSD. Yet when it runs out of VRAM, that doesn't help it, and the same applies to the PS5 which is why having larger caches closer to the GPU is always a good thing

Oh I see now, you're talking about system RAM. Well since you've already written me off as some sort of die-hard console fan, perhaps you will accept the messaging if it's coming from the Godfather of Digital Foundry, for which this thread is dedicated to:


I'm really not sure what you're trying to say with this. Remji was talking about the benefits of having more VRAM (which in the PS5 is equivalent to it's unified ram) and yet you're linking to a video explaining why System RAM can't replace VRAM?

That video is literally reinforcing what he's arguing because it's saying you can't replace VRAM with a larger, slower memory pool further up the IO chain..... you know... like an SSD. So you still need that fast VRAM to cache data, and if you run out of it, the larger slower pool isn't going to be able to substitute.

You can however speed up how quickly you can get data into VRAM which in turn reduces the amount of data you need to cache in VRAM. The PS5's IO can do that pretty quickly. The system RAM in a PC can do it even faster. What Richard is saying though, is that it still doesn't replace the need to have enough VRAM for whatever you're trying to render.
 
@Dictator nice comparison video. Did you notice any weirdness due to denoising? Also is there a good description somewhere of how Image Based Lighting probes work?
 

Pretty amazing stuff.
giphy.gif
 
Status
Not open for further replies.
Back
Top