bitbydeath
Member
GPU Bottlenecks and split memory no doubt.
-
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
GPU Bottlenecks and split memory no doubt.
hemo memo
You can't die before your death
Oh god .. the stupidity in this article is just beyond my understanding.
It is Kotaku.
Hawking Radiation
Member
Sonys not interested in PR buzzwords.
They let the games do the talking.
From with country you are?
I ain't goin to expose my PS5 to an excessive amount of radiation just to satisfy your curiosity
Last edited:
IllegalLemon
Member
Actually Zen and Zen+ had 4 cores ler cluster which is why the top threadripper released on Zen+ had 32 cores. Amd limits these clusters to 8 maximum as putting more more would simply made the cpu way too large. Zen2 transcended that limitation, introducing clusters with a maximum of 8 cores each, now they could fit 16 cores in a cpu using 2 clusters of 8 cores each instead of the 4 clusters of 4 cores each the 1950x had 2 gens before and the overall footprint of the cpu was reduced in size as well, making the cpu smaller while packing the same amount of cores as the 1950x into the 3950x with each core being physically larger as well. This made ultra powerful , mega-tasking powerhouses such as the 3990x possible as 8 clusters could then be used and with 8 cores each, this gave a combined total of an astonishing 64 cores. Zen3 Unified the L3 caches to improve latency as it now is shared across clusters and cores improving ipc in the process, the new 7EUV process used further improves efficiency allowing for significantly higher clocks compared to the 7DUV based Zen2 before it and rumor has it Amd has been hard at work trying to create smt tech that allows for 4 threads per core instead of the usual two threads per core. When we'll see this in a consumer cpu, time will tell.
Last edited:
SSfox
Member
Smartshift also meant to give extra 10% performance gain (source is amd themselves), not sure if XSX uses Smartshift tho.
Last edited:
Bo_Hazem
Banned
Well, you went from disagreeing to agreeing
. It's 72CU and CPU probably Zen3 at 3nm with PS5 Slim with a boost to like 2.5GHz based on RDNA3/4.No. They decided to use the brute force instead.
Last edited:
Lysandros
Member
Yes, if true a unified L3 cache would have much higher impact on instruction per clock than a mere 2-3% higher frequency. If i remember correctly AMD claimed up to 20% IPC increase from this feature for Zen 3.
FranXico
Member
Is this from the same person who wrote that people who want a PS5 should feel bad because of injustice in the world?
Oh god .. the stupidity in this article is just beyond my understanding.
Garani
Member
If not the same, they work for the same questionable employer.
FunkMiller
Banned
Sounds great.... but I think they should just go full blown Spider-verse with it. Alternate New Yorks and all that jazz.
Barakov
Member
"Propaganda"
Oh god .. the stupidity in this article is just beyond my understanding.
azertydu91
Hard to Kill
Wait you have access to radioactive materials? I feel likeit is a bigger problem than showing the ps5's SOC.I ain't goin to expose my PS5 to an excessive amount of radiation just to satisfy your curiosity
Because of the term 'console war', it makes sense."Propaganda"
Last edited:
azertydu91
Hard to Kill
I think it is unlikely considering what they teased at the end of both games.Maybe a Dlc for MM?
draliko
Member
Mine all the time randomly, near every 20 min for 30seconds
Aceofspades
Banned
During installs only, mine is inaudible even when spinning.
Rudius
Member
200 million PS5s sounds very unrealistic at first, but if they somehow manage to do slim versions that are really cheap and push them to developing markets, perhaps it's doable
Captain Hero
The Spoiler Soldier
I have the pulse 3D and it’s great the audio is so clear and the 3D audio in DS is fantastic
Last edited:
Bo_Hazem
Banned
Yeah it's very clean with impressive/finely tuned bass and amazing locality. Demon's Souls is the best tempest showcase so far. Sound quality and performance remains the same wireless or shut down and using it wired. According to IGN that SteelSeries Arctis 7P is on par with Pulse 3D. Both work on PC and even the switch.
Grinchy
Banned
Man, I keep it at the default 50% and sometimes the explosions and gunfire in COD blast my eardrums off
Oh and something cool about the audio in Demon's Souls is when you're in the Nexus and you are walking away from someone and they're doing their "I wasn't done talking yet" lines. You can actually hear their voice becoming distant in the direction you're leaving and it sounds so good!
I'm no major audiophile or anything, but while I don't think the audio quality itself is as good as a nice pair of Sennheisers or something, for $100, these do a pretty good job. The 3D audio effect works quite nicely and these are early days.
Last edited:
Bo_Hazem
Banned
Man, I keep it at the default 50% and sometimes the explosions and gunfire in COD blast my eardrums off
EA games are loud as well, so it depends on the game.
Razvedka
Banned
They offered an entire Masters program centered around Java? That's incredible.
Razvedka
Banned
Btw I will try to ignore the fact HTML is having the same treatment in this discussion as a language like C ++ has
Tbh, CSS3 and HTML5 aren't even entered into my calculus when it comes to technical skillsets, but that's largely because as a web app guy it's simply "assumed" you understand and excel at it. The reality is most pure programmers I know or who have worked with, talented guys, have lost their shit at CSS3 because "it doesn't make any sense".
So if there's a non-web guy who has tackled these and actually is good with them then I'm not going to thumb my nose at that guy. He sat his ass down and figured it out.
Physiognomonics
Member
Do you think you really like your bestfriend? Maybe you are projecting.Python is shit. I know that it triggers a lot of people, but when I see banks asking Python competencies for their traders, I feel vindicated. I did think it was buzz-bloatware 15 years ago, I do think it now. And my best friend makes a living with python, mind you
Last edited:
Razvedka
Banned
Python is shit. I know that it triggers a lot of people, but when I see banks asking Python competencies for their traders, I feel vindicated. I did think it was buzz-bloatware 15 years ago, I do think it now. And my best friend makes a living with python, mind you
I haven't been super impressed with Python. When I was doing my OSCP (don't hold your breath, I failed it. Will be retaking later this/next year) alot of it was (of course) in Python and to me it always felt like an inferior Node.js, which of course isn't a valid assessment given the age disparity between the two languages. But for most things Node.js is simply better vs Python, there's only a few areas I can think of where the latter has a serious edge over the former.
By in large I think it comes down to legacy systems and knowledge
inflation
Member
The best part of Python is its ecosystem. During the years of maturing, you can find all kinds of packages in nearly every area, whether are written in Python or not. So it makes Python the perfect glue language to do something quick and dirty.
I love about it when scripting, because it’s battery-included, so I don’t need to download 20 misc packages from npm. People outside of engineering like science folks also love it, because of the fantastic community built around it, as an alternative to Matlab.
v_iHuGi
Banned
Split Memory is sub optimal, most games will be using >10gb especially at 4K which will cause huge huge bottlenecks going forward.
GPU isn´t looking any better too but the situation isn´t as dire as RAM.
Riky
$MSFT
Series X doesn't use split ram as in the way say PS3 did, Devs see it as one pool this has been confirmed several times, there is no evidence from anyone that any game so far uses more than 10gb of Vram, link if you can find one. Even the slower portion of ram is faster than the Xbox One X ram which ran games at 4k.
That 10gb also has a massive bandwidth advantage over PS5 on top.
That 10gb also has a massive bandwidth advantage over PS5 on top.
SSfox
Member
It really make me mad seeing this, it's one of the many ps2 games I would have loved to replay in ps5 with BC, but unfortunately Sony doesn't seem to care about ps123 BC :/
TheThreadsThatBindUs
Member
Yes, in some ways it's better, but in some other significant ways (i.e. common bus, cannot access both pools simultaneously) it's meaningfully worse.
Correction. Devs see it a one contiguous address space. Not the same thing. That's not special or unique. That's how all unified memory architecture systems work.
The problem is that anytime the GPU needs to read data in the slower pool, they're limited to a significantly smaller bandwidth than even the PS5.
10GB is not a lot and will become even more limiting as the gen progresses. It's RAM not a software managed cache. So devs need to ensure all performance-critical data lies in the fast pool. Not all devs are created equal and on projects where devs are pushing against tight project deadlines with the need to target a stupid number of platforms, the spilt memory setup on XSX and XSS is just another headache devs don't really need or want to have to optimize for.
This is completely clueless. Every game released ever uses the full amount of RAM available for every system games will run on. It would be obtuse to have a game running on a system and not be filling the RAM completely with every bit of requisite data you possibly need. Otherwise you're simply not making effective use of the hardware resources provided. This is just fucking common sense.
The issue with XSX's split memory setup is that any data residing in the 6GB slow pool needs to be buffer data and not the stuff your GPU needs to hit often, otherwise you'll be seeing GPU stalls galore. This is something devs need to carefully manage, unlike on PS5.
LucidFlux
Member
Wow so much new and exclusive information packed into a nearly 30 min video. I know some of you already mentioned there was nothing new of interest here but I wanted to break this down to show you how convoluted this guys videos are...
Mentions it will take time to fully utilize a feature:
3
Custom RDNA2/RDNA2 Mentions:
11
Geometry Engine mentions:
12
Cache scrubbers:
5
Number of times Road to PS5/Cerny was mentioned:
9
API/DK/GDK mentions:
10
Times he mentioned his previous leaks (and then repeated them):
10
Times PS5 was compared to Xbox/PC even though he goes out of his way at the end to mention "this is not an xbox has this and xbox doesn't have that video":
19 -LOL
Potentially new information section:
8MB of unified L3 cache in CPU: previously mentioned, not actually new and still not confirmed by Sony but highly suspected.
CPU reserves for system/OS: 1 dedicated core, "generally" - Pretty obvious guess, the qualifier "generally" he uses here doesn't shout confidence in this number and also leaves room for other possibilities.
Geometry Engine: Regurgitates all known information in a shiny new package, but the huge nugget he dropped here is when the GE can cull geometry and according to him it's "super duper" early. Then goes on to obviously state how Nanite was leveraging the GE to "control geometry" - another huge revelation. It actually controls geometry. Woah!
Cache Scrubbers: Again repeats Cerny then says he was told this was a "pretty large" speed up of GPU - So specific...
Tempest Engine: Cites Cerny again, mentioning Tempest can be used for other game calculations such as physics then mentions that according to a developer working on 2D game their audio used ~20% of tempest engine. Doesn't tell us a whole lot but may actually be the only new piece of information here besides the cpu reserves.
GPU Frequency: PS5 always seems to stay at peak frequency, 95% of the time according to "developers". If it dips it's for less than a single frame.
So he has access to a developer who is running GPU profiler tools on the PS5 to check on clock frequency throughout a frame but can't get any more information about GPU Cache sizes, structure, how often the cache scrubbers purge data or increase performance other than it being a "pretty large" speed up of the GPU then I just don't buy it.
TLDR Conclusion
I finally know who RGT's source is and it's been right in front of us this whole time. It's Mark Cerny. He not so subtly hinted this by name dropping him and his Road to PS5 bible nearly 20 times in just this video. All of his information can be gleamed from Cerny's talk, sprinkled with other publicly known information from AMD as well as obvious conclusions that can be drawn from the known information all wrapped up nicely with some generalized qualifiers just in case some numbers miss the mark.
I'm sorry but when you feel the need to qualify yourself by mentioning 10 of your previous leaks in a single video it just means you don't have anything of substance here, otherwise you'd let the new information stand on its own.
Does anyone know if Cerny has a YT channel or podcast? I'd rather just listen to the source.
Razvedka
Banned
Why would you need to download anything from NPM necessarily? Node.js is the full V8 JS engine + its built in goodies that make Node, Node. I've written entire programs in it without downloading stuff from NPM. Perhaps I'm showing my youth here by asking such a question lol.
And I agree, Python has strong legacy support with most stuff out there. The other side of the coin though is that NPM is viewed as 'best in breed' for package management and iirc there are more available packages for Node.js then any other language currently. So its a huge ecosystem.
@OnTopic
I'd be interested in seeing figures for cache scrubber impact on GPU performance, but given PS5 is so customized I guess that would be a hard thing to do. 'Generic GPU with cache scrubbers vs Generic GPU without' sort of thing.
Last edited:
DonkeyKong
Member
This split ram thing is a bit confusing, 10Gb runs at 560GB/S and 6Gb runs at 336GB/s, only the GPU sees the faster ram, right? so does the GPU always access the higher speed ram or does the whole speed drop to 336GB/s when the slower ram pool is also being accessed? Who knows?
Mahavastu
Member
No, the memory is unified and BOTH, the CPU and the GPU are able to access the full 16GB of the memory, no matter if fast or slow.
The thing is, that the GPU needs higher bandwith than the CPU (on PC the CPU ram is way slower than on consoles), therefore as a dev you will try to keep the GPU specific stuff in the fast area and the mostly CPU stuff in the slower part. Stuff which both access (there are reasons why you want to have unified memory) you try to keep in the fast area as well.
The RAM is using the same lanes in the computer, so you access either the slow part or the fast part of the memory, not both at once. You can read this as accesses to the slow memory slows down access to memory overall...
Last edited:
EliteClarity
Member
If anyone from Sony is Watching, PLEASE can we mute game audio easier so I can Listen to Spotify while gaming. I Current can see no way to do it, apart from go to options and individually turn down/off each audio setting on a per game basis, and fuck me if I want the game audio, I have to do the same and turn it all back up again.
Yes, Yes, First world problems, but it it annoys me endlessly.
Yes, Yes, First world problems, but it it annoys me endlessly.
TheThreadsThatBindUs
Member
The two virtual "pools" of memory exist on 10x physical GDDR6 memory chips, which are a mix of 1 and 2GB densities.
The "fast pool", comprises the 4x 1GB GDDR6 chips together with the first 1GB of the 6x 2GB density chips.
The "slow pool" comprises the second 1GB on the 6x 2GB density chips.
Think of the 2GB chips as being partitioned in software, which the first GB accessible in the "fast pool" and the second GB accessible in the "slow pool".
The number of chips is what defines the memory bus width and thus bandwidth, with each chip having 32x pins with 1bit per pin.
So the fast pool has an interface width of = 10 chips x 32 pins per chip x 1 bit per pin = 320bit
Likewise the slow pool width is = 6 chips x 32 pins per chip x 1 bit per pin = 192bit
With 14Gbit/s GDDR6 chips that gives the bw figures of 560 GBytes/s and 336GB/s respectively for the two pools.
As you can tell, both pools are accessed physically through the same pins on each 2GB chip, i.e. a common memory bus, so both pools cannot be accessed at the same time.
So when accessing the second 6GB on the 2GB density chips, i.e. the slow pool, you're doing so though a 192bit interface, thus your bw is limited to 336GB/s only.
I'm pretty sure the GPU can access both memory pools, as otherwise that would be an oversight in design of the highest order. Effectively making the system a non-unified memory architecture with only 10GB of VRAM (putting it at a significant disadvantage to the competition). I'm pretty confident this is not the case.
Both the CPU and GPU can access both pools, but you really want your GPU reading primarily from the fast pool. Otherwise your performance will be tanking hard.
Insane Metal
Member
Is nobody going to get that PS5 APU die shot? What the fuck? Last gen we got that before launch IIRC or right after both consoles released.
We need to know if the PS5 has unified cache in its CPU. If that's indeed the case like RGT is saying it's a huge benefit for PS5.
We need to know if the PS5 has unified cache in its CPU. If that's indeed the case like RGT is saying it's a huge benefit for PS5.
MasterCornholio
Member
What I'm curious about is when the GPU needs more than 10GBs what advantage will the PS5 have in multiplats?
megreotsugua
Banned
EliteClarity
Member
Metroiddarks
Member
Don't get me wrong I respect a person who can use things like CSS in a correct way, the fact is not a programing language doesn't mean their work is less important.
Garani
Member
Do you think you really like your bestfriend? Maybe you are projecting.
He saw me through a tough time in my life, and was my Bestman as well. Nobody is perfect
Python is majorly loved in Physics and Math environments. Today's trading is very much based on physics being applied to the stock market (I know, it sucks, but that's what it is nowadays) so this is why trader's roles include Python. For the very same reason it has been injected into adaptive security principles. Problem is that adaptive security isn't about getting a log and doing something with it (you can use Perl, and it's a much stronger string manipulator). Adaptive security is a wider arena that encompass proper asset management (CMDB), configuration base lines, impact analysis, information pathways validation and intelligence feed. Put all together and THEN you apply it to your first line's security operations KABOOM! You have adaptive security.
Problem is that it's fucking expesive (resource wise) and you gotta have management that really buys into it. Once they get word that you are asking for a 20 to 30 people team, to start with, for a project that will start seeing the light of day not before 2 years have passed, everyone disappears.
- Status