-
Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.
- Status
Coulomb_Barrier
Member
Where is this Navi info from? Moore'sLawisDead?
Dodkrake
Banned
Microsoft has a way with words, their marketing team is strong. It looks like most Xbox only fans seem to be buying into the 12GB/s narrative, even though that is NOT what Microsoft said (but worded to purposely mislead). This information is from https://news.xbox.com/en-us/2020/07/14/a-closer-look-at-xbox-velocity-architecture/.
First, let's look into the actual I/O performance. There will be omissions as I don't want to copy the whole text, but please check the original article as I want to keep the context intact:
Custom NVME SSD
Translation: As opposed to normal PC SSD's, this one is designed to deliver consistent 2.4 GB/s raw I/O throughput. That's it. That's as much as you're gonna get.
Hardware Accelerated Decompression
Translation: If you assume a 2:1 compression ration, which is perfectly possible, you will get a 4.8 GB/s compressed I/O throughput. I will assume this is on average, so the peaks will likely be higher, up to whatever their decompressor allows, and some data will not compress as well, but that's it. Your average is 4.8 GB/s, not more, not less.
Sampler Feedback Streaming (SFS)
Translation: First, let's clarify again - The Series X still has a raw I/O throughput of 2.4 GB/s. Now that we've clarified that, let's clarify the funny word play. Assuming the numbers they mentioned are correct, on older systems you'd access 1/3 or less of the loaded textures, and with the new tech, you boost utilization by 2.5 of the effective raw bandwidth. This means that if you didn't have this tech, your available bandwith would be equivalent to 0.96 GB/s, as you'd be loading that data you don't need. This is pretty much what we saw with their old gen game switching (video below), where the average time to switch between game A and B was of 6 seconds, or 5.76 GB/s of RAM being loaded (games were programmed for 5.5 GB/s AFAIK).
So, to recap, the capabilities are:
Edit: Sorry, forgot the video I mentioned above
First, let's look into the actual I/O performance. There will be omissions as I don't want to copy the whole text, but please check the original article as I want to keep the context intact:
Custom NVME SSD
Translation: As opposed to normal PC SSD's, this one is designed to deliver consistent 2.4 GB/s raw I/O throughput. That's it. That's as much as you're gonna get.
Hardware Accelerated Decompression
Translation: If you assume a 2:1 compression ration, which is perfectly possible, you will get a 4.8 GB/s compressed I/O throughput. I will assume this is on average, so the peaks will likely be higher, up to whatever their decompressor allows, and some data will not compress as well, but that's it. Your average is 4.8 GB/s, not more, not less.
Sampler Feedback Streaming (SFS)
Translation: First, let's clarify again - The Series X still has a raw I/O throughput of 2.4 GB/s. Now that we've clarified that, let's clarify the funny word play. Assuming the numbers they mentioned are correct, on older systems you'd access 1/3 or less of the loaded textures, and with the new tech, you boost utilization by 2.5 of the effective raw bandwidth. This means that if you didn't have this tech, your available bandwith would be equivalent to 0.96 GB/s, as you'd be loading that data you don't need. This is pretty much what we saw with their old gen game switching (video below), where the average time to switch between game A and B was of 6 seconds, or 5.76 GB/s of RAM being loaded (games were programmed for 5.5 GB/s AFAIK).
So, to recap, the capabilities are:
- 2.4 GB/s raw I/O throughput
- 4.8 GB/s compressed I/O throughput.
- A boost of 2.5 times compared to old gen tech, meaning you can fully utilize the numbers above, as opposed to 1/3 of the numbers available to last gen games
Edit: Sorry, forgot the video I mentioned above
Last edited:
K.N.W.
Member
Thanks for the clarification, but I don't think I've been confused, rather it was someone trying to feed us misinformation
D
Deleted member 775630
Unconfirmed Member
What people are saying is that if you wanted to use the old tech you would need a 12GB/s SSD, now a 4.8GB/s SSD is enough.
https://www.guru3d.com/news-story/a...ercent-more-than-the-geforce-rtx-2080-ti.html
from the talk of clusters etc and comments, obviously still hearsay until we see final spec but looks very likely
EDIT - seems as though its from MooreLawIsDead
Last edited:
geordiemp
Member
The 2.5 x is misleading as it does not establish a proper base to compare the 2.5 x .
On last gen you loaded a level or maybe 10-15 seconds of prefetch or you loadedd the whoe level.
If you can load next 5 seconds of gameplay, then memory requirement is less. Maybe 2.5 x less, but thats compared to loading the level or loading 15 seconds ?
Statement is vague on purpose.
Snowdonhoffen
Member
Maybe I'm not remembering correctly, but wasn't there supossed to be a deep dive about XVA? Hope this wasn't it... It's like they were just making a summary of past months info.
Lort
Banned
its very clear, on aveage 2/3 of a texture load are never used, they were loaded but did not need to be.
By providing smaller blocks and realtime feedback of actual runtime game needs you can avoid loading over 60% of texture data.
If the xbox one x had this, it could get by with about 6 gigs of total ram rather than 12... with still the same quality textures rendered.
Last edited:
Thirty7ven
Banned
The problem is thinking this only applies to XSX and in comparison to PS5.
I think it’s smart of MS to market it this way, but it’s also a clear sign they haven’t been getting the traction they thought they would because they keep coming out with the same info just in different colors.
Lort
Banned
Lol what a beautiful pictorial way to show you have no idea what SFS is...So, basically what SFS does is the same as what Cerny was talking about? Refer to pictures attached.
I guess SFS is to do with loading data .. and your pic shows data being loaded ...
of course none of that actually relates to what SFS is and why its such a huge advantage.
SFS looks what what is being rendered and automatically triggers a SSD read and hardware decomposition .. so rather than a program requesting all the data that may be needed to be loded in ... the game in realtime triggers the read requests.
Imagine a cube spinning with a texture on each side .. the ps5 like all traditional games will load all sides of the cube so they are ready to be rendered at any time ... thexbox sfs waits for one side to slighly visible then loads the texture... thus meaning a huge amount less RAM usage and 2.5 times faster loading.
Games usually stream in a new area as the player enters they dont stream in partial textures as they are being required for rendering.
Last edited:
Dodkrake
Banned
No, that's not what people are saying. What people are saying is that the effective throughput is 12GB/s and comparing that to PS5's 9GB/s, saying somehow the Xbox has higher throughput and can better utilize. This is a lie in both accounts.
Edit: Case in point, the post before mine
Last edited:
D
Deleted member 775630
Unconfirmed Member
I deliberately never discussed the PS5 because I don't know enough about it. Does it have an exact same solution as SFS? Or something that's similar but not equally as efficient? If so how efficient is PS5's solution? Do they even need that since their SSD is already so fast? So many unanswered questions for me when it comes to PS5, which is why I always make the comparison with old systems.
True, what they say is that IF PS5 doesn't have a similar solution it will only utilize 3.6GB of the passed on 9, which is still less than 4.8. But like I said in my post above, I don't know if PS5 has something similar to SFS
Last edited by a moderator:
geordiemp
Member
SFS is a filter that allows a larger mimmap to be blended when it arrives LATE.
Do you think XSX is going to run with small textures and load high res ones only when on screen.
LOL
geordiemp
Member
No Xbox could not, its latency was huge. You had to load allot of data in advance, usually 10-15 minutes of gameplay/ Last gen consoles streamed from HDD at 20-50 mbs and had high latency.
Your now saying the prefetch amount is now 16 ms and 1 frame. REALLY ?
MS are really good at confusing and giving vague information.
Instant, 2.5 x something....something....some understand, somethink its now 15 GBs lol - MS know what they are doing, confusing the easily confused.
Last edited:
Dodkrake
Banned
Which is wrong in all accounts.
sircaw
Banned
Sorry Babes!Ohmylord! Did you just assume my gender here???? I'm flabbergasted!
It's she
Ps, your still fat.
Last edited:
Dodkrake
Banned
For context:
To load / create a model you need Shaders and Textures.
If you cull the shader because you don't need it, it means you are also not loading the texture, effectively only loading what you can see.
The SSD speeds allow for the loading of Mips as needed, so you don't need the full mip map on any circumstance, as you can procedurally load it.
You don't even need something like SFS to smooth the transition because the raw speed of the SSD is more than doubled on the PS5.
So no, there's no advantage in the Xbox, it's still a pretty big difference in IO Throughput, no matter how people spin it.
To load / create a model you need Shaders and Textures.
If you cull the shader because you don't need it, it means you are also not loading the texture, effectively only loading what you can see.
The SSD speeds allow for the loading of Mips as needed, so you don't need the full mip map on any circumstance, as you can procedurally load it.
You don't even need something like SFS to smooth the transition because the raw speed of the SSD is more than doubled on the PS5.
So no, there's no advantage in the Xbox, it's still a pretty big difference in IO Throughput, no matter how people spin it.
Lort
Banned
Games right now stream some textures just as they are needed .. thats why games often have texture pop in. You dont need to load data 15 minutes beforehand ... you load all the data you need right now and stream in as the player moves around the world .. which is then limited by bandwidth ( the amount of data that you can stream in) and latency ( the time taken to get the data after you request it).
The cerny pics show the increase in bandwidth reduces the size of the RAM cache required in memory ( whoch applies to the xbox as well but to a lesser extent since its slower), SFS reduces the number and size of textures that need to be loaded in. SFS tackles the latency limitation, the raw speed of SSD ( for both consoles )tackles the bandwidth.
The relationship between bandwidth and latency becomes quite complex ... needless to say SFS reduces the effective advantage the ps5 raw throughput offers.
Last edited:
D
Deleted member 775630
Unconfirmed Member
Then how come that textures were being loaded and only 1/3 was being used?For context:
To load / create a model you need Shaders and Textures.
If you cull the shader because you don't need it, it means you are also not loading the texture, effectively only loading what you can see.
The SSD speeds allow for the loading of Mips as needed, so you don't need the full mip map on any circumstance, as you can procedurally load it.
You don't even need something like SFS to smooth the transition because the raw speed of the SSD is more than doubled on the PS5.
So no, there's no advantage in the Xbox, it's still a pretty big difference in IO Throughput, no matter how people spin it.
Lort
Banned
The xbox loads only the part of the high res mip map it needs .. so the effective textures stored in memory size goes from 16gig to 24-32gig.For context:
To load / create a model you need Shaders and Textures.
If you cull the shader because you don't need it, it means you are also not loading the texture, effectively only loading what you can see.
The SSD speeds allow for the loading of Mips as needed, so you don't need the full mip map on any circumstance, as you can procedurally load it.
You don't even need something like SFS to smooth the transition because the raw speed of the SSD is more than doubled on the PS5.
So no, there's no advantage in the Xbox, it's still a pretty big difference in IO Throughput, no matter how people spin it.
SatansReverence
Hipster Princess
lepermessiah
Member
Jesus H Christ, now the velocity architecture is possibly superior?
It's certainly no slouch but NO!
EMPIRICALLY NO!
LESS PRIORITY LEVELS, LESS CHANNELS
But I Am AuThoRitY!
NO!
It's certainly no slouch but NO!
EMPIRICALLY NO!
LESS PRIORITY LEVELS, LESS CHANNELS
But I Am AuThoRitY!
NO!
Dodkrake
Banned
Because your effective throughput was 100X less, so you needed to load everything the player may need to see in the next 30 seconds / 1 min / whatever time.
Imagine a rock
In current gen, the player can either stand still for 1 min, move at a slow pace and require a new mip in 1 min, or run fast and require that in 10 sec. Then, after those 10 seconds (worst case) the player can turn around and look at the other side of the rock. So, you need to account for worst case, so you load all mips for the next minute (lets say 4 levels) + all textures of every single object from all angles, as you need to account for that worst case scenario.
In new gen games, since you can load 1 sec ahead, you don't need to account for even 10 seconds of loading, so you don't need to load all mips for that rock, nor you need all angles.
Thirty7ven
Banned
The only bit that is specific to XSX are texture filters that help avoid artifacts and pop induring the process of blending the lower res texture to higher res one. It’s a cool bit, but it’s not really a game changer by itself.
Last edited:
sircaw
Banned
Listen up Mountain man, enough of your cheap insults, Tilapia is one of the most consumed fish in all of Africa, with out this little HERO, there would be mass starvation and death.
His gives his life so that others may live.
A true warrior in the grand scheme of things.
Show respect, HEATHEN.
Lort
Banned
SFS requires exclusive hardware and integration to perform the selective realtime SSD reads. MS has been clear about that and Sony has never claimed to have that capability.
You claiming something without justification or rationale that even Cerny never has.
In short your wrong and i suggest you read my ... or say Microsofts descriptions of how it works. They do mention they ALSO have hardware high quality filtering to slowly transition texture loads.... but that is clearly and obviously not what SFs is.
geordiemp
Member
No it does not, sofware does not make an SSD chip behave less latent, it just can bypass IO bottlenecks in other software.
The Chip latency is what it is, you cannot magic it away with marketing.
Dodkrake
Banned
No it doesn't. It means that in current gen you'd need 13.5 times 2.5. yeah, it's 13.5 for games, not 16.
Also, geometry culling achieves the same goal. You're not loading bloody texture maps for geometry you're not rendering.
Thirty7ven
Banned
See this post? It’s all about what somebody wants to believe and appealing to authority which in this case the authority is MS. But then if use the same reasoning I reach the conclusion less CUs are better because Cerny said so, and XSX doesn’t have geometry culling because MS didn’t talk about it. See it helps when you question these things and try to understand them instead of adhering to narratives.
You don’t even know what sampler feedback is. Go read on it, then go read on Sampler Feedback Streaming, and then go read on what’s actually specific to XSX and for what purpose. The only thing specific to XSX are hw texture filters to prevent artifacts and pop in.
If you want to remain ignorant, I won’t stop you. The world will just move on.
Last edited:
Lort
Banned
So in order to show im at least somewhat agnostic here is some talking points some you might want to argue if you want to debate xbox fans in a more informed fashion.
SFS is designed for textures and may not work as intended on meshes or voxel / texel style engines such as nanite. Ps5 bandwidth applies to everything.
Raw bandwidth of SSD would usually reduce the latency by half .. so its possible ps5 SSD latency is half xbox.. but dependS on a LOT of factors and potentially xbox latency could be lower ... noone knows atm.
SFS enables programmers to forget about management of textures which is awesome for open world games or smaller developers. Triple A developers should be able to get more effective texture usage on the ps5 where texture use is predictable ( ie cut scenes or linear games).
Xbox has more CU which is better for ray tracing but can be harder to fully utilise, however games with async compute can easily fill unused CU counts.
xbox raw bandwith is helpful for high res textures .. ps5 can load them faster but xbox can render them faster...
All in all both are going to look awesome and multiplatform triple a games will probably look exactly the same on both .. with maybe a few frames faster on xbox or unnoticeably small res bump... and a slightly short blank load screen on the ps5.
so whatever console you want to “win” just buy it and have fun....
SFS is designed for textures and may not work as intended on meshes or voxel / texel style engines such as nanite. Ps5 bandwidth applies to everything.
Raw bandwidth of SSD would usually reduce the latency by half .. so its possible ps5 SSD latency is half xbox.. but dependS on a LOT of factors and potentially xbox latency could be lower ... noone knows atm.
SFS enables programmers to forget about management of textures which is awesome for open world games or smaller developers. Triple A developers should be able to get more effective texture usage on the ps5 where texture use is predictable ( ie cut scenes or linear games).
Xbox has more CU which is better for ray tracing but can be harder to fully utilise, however games with async compute can easily fill unused CU counts.
xbox raw bandwith is helpful for high res textures .. ps5 can load them faster but xbox can render them faster...
All in all both are going to look awesome and multiplatform triple a games will probably look exactly the same on both .. with maybe a few frames faster on xbox or unnoticeably small res bump... and a slightly short blank load screen on the ps5.
so whatever console you want to “win” just buy it and have fun....
D
Deleted member 775630
Unconfirmed Member
I agree, and get what you are saying. But does the PS5 have the capabilities to chose which parts of a mip it should load? Because that's where the efficiency is for the XSX. It can choose which sub portions of the mips you should load, while the PS5 stil loads the complete mip?
Rudius
Member
Lort
Banned
Geometry culling is common practice and nothing new the textures are in memory already before the geometry is culled. You are absolutely loading texture data for geometry that becomes culled. Have you ever waited for a level to load .. its loading the textures.... then a small amount of extra textures are streamed.
13.5 x 2.5 = 33.4 gbytes .. i was being conservative by saying effective 24-32 gigs equivalent ram usage.
CurtBizzy
Member
The SFS 2x multiplier is only being used for textures, SFS 2.5x does not include the overal memory
you have more memory due to less textures being stored in ram
Dodkrake
Banned
Yeah, super common, but you are culling before sending the information in to your GPU. You effectively don't load the shader nor the texture.
Anyway, you are again misrepresenting what they said, so there's no point in arguing. Keep living in your fantasy land and brace yourself for disappointment.
DrDamn
Member
Had some questions on SFS which I think are just starting to be answered on this page. The x2.5 figure needs context.
Presumably that is for a given frame? So no need for parts textures on the other sides of objects, no need for parts of textures hidden by other objects? What if you move? What if objects move? Then you need those textures and they need to be loaded in really quickly? The x2.5 figure doesn't apply to all the other stuff you need to load either does it - models etc, just textures? It also won't make any difference if you are loading up a game/level which fits entirely into memory or doesn't need to load textures just in time?
Presumably that is for a given frame? So no need for parts textures on the other sides of objects, no need for parts of textures hidden by other objects? What if you move? What if objects move? Then you need those textures and they need to be loaded in really quickly? The x2.5 figure doesn't apply to all the other stuff you need to load either does it - models etc, just textures? It also won't make any difference if you are loading up a game/level which fits entirely into memory or doesn't need to load textures just in time?
icerock
Member
Bloomberg article is an interesting read in light of today’s development. They wrote about Sony struggling with the pricing as BoM had soared to $450. The key issue being the hike in DRAM prices. I haven’t kept up with the market to know where they are at. But, the important point is at the time, nobody knew about the existence of a digital SKU. You can reduce $25 from the BoM for the digital SKU, add the retailer cut, packaging/shipping and other logistics cost. Sony could eat ~$80-100 losses and have the digital SKU hit the $399 mark.
They can offset some of this by eating only ~$15-20 on the disc SKU by launching it at $499. Such a scenario doesn’t leave them too much in the red, while at the same time gives them an appealing $399 magical entry point into a new gen.
I’m very hopeful about a $399 Digital SKU although it’ll be mighty hard to find them initially. I still expect me and others who are going with a disc SKU to pay a $100 premium unless they are feeling very charitable.
They can offset some of this by eating only ~$15-20 on the disc SKU by launching it at $499. Such a scenario doesn’t leave them too much in the red, while at the same time gives them an appealing $399 magical entry point into a new gen.
I’m very hopeful about a $399 Digital SKU although it’ll be mighty hard to find them initially. I still expect me and others who are going with a disc SKU to pay a $100 premium unless they are feeling very charitable.
geordiemp
Member
So beacuse some testures POP in late on current gen, your assuming textures are requested , fetched and used within a frame ?
D
Deleted member 775630
Unconfirmed Member
To answer some of your questions, Microsoft said:
The x2.5 applies to the complete memory usage due to this change to textures. SFS only applies to textures and makes sure only the 1/3 that is needed is loaded, which results in a total of x2.5 more bandwidth that can be used because it's not only textures that need to be loaded. Simplified but take this as an example:
Memory bandwidth of 100:
* textures uses 90
* other stuff 10
Due to SFS it now is:
* textures uses: 30
* other stuff 10
Only 40 is being used, so we have x2.5 as much memory that is now free to use. You might be wondering, but why the 90-10 split? 2 reasons, first of all most bandwidth is being used by textures, secondly, the math works perfectly like this so I'm pretty sure Microsoft used the same distribution to get to these numbers.
Lort
Banned
On pc you can stream some textures from RAM to GPU ram .. on the consoles there is no seperate ram .. so in console space its either loaded into ram or not.
No current console can realistical stream any useful amount of texture in between the culling phase and the gpu texture and lighting phase .. which is what your implying. If what your implying was true ( loading textures from disk mid frame) on next gen consoles then the xbox could realistically have 1tb of “ram” for textures.
That wont work ( for numerous reasons) one being bandwidth lets say you have 10 gBytes a sec SSD... that divided by 60 fps = 166 mBytes per frame.. you could only address THAT each frame ( which uses the whole time to just load and 0 time to setup geometry, render shaders or pixels and 0 latency for an SSD read). Of course if you wanted to make the most of that bandwidth youd want to ve VERY selective of what you loaded .. you could sample the previous scene and precache data from that automatically... Wow i think i just invented SFS...!
Alex Scott
Member
You can shout SFS all you want, but the raw speed is still 2.4GB/s, and 4.8 compressed. How does SFS makes up for that i/o difference.
Alex Scott
Member
YEs it is
There is a reason that they did not not have a presentation like Sony after all to give to give a deep dive, All they can do it misinform the fans.
Lort
Banned
i should have said “preload all the data you need right now and stream in a very limited amounT” .. as per my other discussion i do NOT believe it is at all possible on current gen to stream any textures in mid frame and i still doubt its potential with the next gen.
D
Deleted member 775630
Unconfirmed Member
Because every byte of that 4.8GB is being used by the GPU thanks to SFS (in case of textures)
Last edited by a moderator:
- Status