Support NeoGAF

[Loxus]

It's both. In the end it's about the margin. GPUs had a ridiculous margin priced in from the get go which caught a lot of the bullshit so far but the RAM crisis was just one hit too many.
If you are limited by parts or production capabilities you choose the most profitable path which is AI hardware.
And of course this doesn't change the fact that the ridiculous price increases for RAM and parts in general hit everyone, including Sony which has to compete with the likes of Nvidia for parts and production capacity now. With the current prices and price trajectories and the leaked specs 700 bucks wouldn't be doable without heavy subsidies and there's no reason for that when you have no competition.

I've been saying this for awhile. These DRAM manufacturers are trying to cash out on the AI demand due to the fact that they can over charge the AI hyperscalers wyh HBM. And what's funny is HBM was first created for consumer GPUs.

Even some journalists are beginning see figure it out.
AI memory crunch forces DRAM market into 'hourly pricing' model, report claims — small and medium-sized businesses fighting for survival
All this raises a question about what happens next if enough SMEs exit the market because they can't afford the premiums. If smaller buyers collectively pull back, tight capacity could soon become oversupply, potentially exposing the shortage as "illusory," says DigiTimes.


And I don't understand why you believe Sony can't secure RAM when they are currently doing it with the PS5. They are surely in the 100 top-tier buyers.

The report describes a market now split between roughly 100 top-tier buyers with the leverage to secure supply and more than 190,000 small and mid-size enterprises fighting over what remains.

The report says that cloud service providers, leading automakers, and smartphone giants Apple and Samsung hold enough financial clout to resist price hikes and maintain priority allocation from memory manufacturers. Samsung, SK hynix, and Micron cannot afford to jeopardize those relationships, so these large customers get served first, while also increasingly requiring prepayment or cash transactions before confirming orders — terms that smaller firms with little bargaining power will struggle with.

 

[Loxus]

You think 9070XT class GPU will reach 5090 levels when PSSR+RT is enabled vs. DLSS+RT on 5090?

If we covert 9070XT power to nvidia equivalent we end up with 5070ti:

That's the difference for raster. Here RT:

If we assume that RDNA5 is more peformant in RT than Blackwell (possible of course), do you think it's enough to cover almost 2x difference in power?

The relative performance are based on raw specs.

 

[Bojji]

The relative performance are based on raw specs.

It's based on their (TPU) reviews of GPUs.

 

[Loxus]

It's based on their (TPU) reviews of GPUs.

Which is based on raw specs.
It would be dumb of them to use DLSS against GPUs that can't use any ML upscaling.

 

[Bojji]

Which is based on raw specs.
It would be dumb of them to use DLSS against GPUs that can't use any ML upscaling.

When you add DLSS into the mix you will still achieve the same performance delta between 5070ti and 5090.

If you are talking about using pssr2 on PS6 vs native 4k image on 5090. You will still get almost 2x better results on 5090 by using performance DLSS (partity in native resolution vs. PS6).

 

[MrSmooth]

Genuine question, I thought that one of the pillars of project amethyst was this universal compression thing, meaning less memory would be needed?

 

[Loxus]

When you add DLSS into the mix you will still achieve the same performance delta between 5070ti and 5090.

If you are talking about using pssr2 on PS6 vs native 4k image on 5090. You will still get almost 2x better results on 5090 by using performance DLSS (partity in native resolution vs. PS6).

Here from the relative performance RT we see the 5090 doing 68fps at native 4k. The PS6 should be able to do that with RDNA5 Radiance Cores and PSSR3 utilizing the neural arrays.

I don't know why that is so hard to believe when the 9070XT is already doing 35fps native.
Ray Tracing Performance

Then here we see the 68fps at native 4k and 96fps with DLSS enable.
DLSS 4

 

[StereoVsn]

You think 9070XT class GPU will reach 5090 levels when PSSR+RT is enabled vs. DLSS+RT on 5090?

If we covert 9070XT power to nvidia equivalent we end up with 5070ti:

That's the difference for raster. Here RT:

If we assume that RDNA5 is more peformant in RT than Blackwell (possible of course), do you think it's enough to cover almost 2x difference in power?

Wouldn't PS6 be also bandwidth constrained with 160-bit bus and also lower cache amounts vs say an 5080 (not to mention 5090)?

 

[Loxus]

Genuine question, I thought that one of the pillars of project amethyst was this universal compression thing, meaning less memory would be needed?

Yes, but it still needs something to differentiate it from previous generations.

AI and LLM (which utilizes lot of memory) might be the differentiator with PS5 differentiator being RT.

And we are mostly only going by MLiD, who leaked most of the PS6 stuff.

 

[Loxus]

Wouldn't PS6 be also bandwidth constrained with 160-bit bus and also lower cache amounts vs say an 5080 (not to mention 5090)?

Not really based on this.

 

[StereoVsn]

Not really based on this.

Wonder what's the chance this will turn out like they are describing.

 

[Cumulus]

And I don't understand why you believe Sony can't secure RAM when they are currently doing it with the PS5. They are surely in the 100 top-tier buyers.

You lead too many dialogues at once. This was not about not being able to secure parts at all this was about your 700 buck pricetag for the PS6 which is simply impossible without heavy subsidies with the supposed specs as long as this AI driven hardware craze goes on.

 

[simpatico]

RAM hasn't been the console concern in 2 gens. It's CPU.

 

[Bojji]

Here from the relative performance RT we see the 5090 doing 68fps at native 4k. The PS6 should be able to do that with RDNA5 Radiance Cores and PSSR3 utilizing the neural arrays.

I don't know why that is so hard to believe when the 9070XT is already doing 35fps native.
Ray Tracing Performance

Then here we see the 68fps at native 4k and 96fps with DLSS enable.
DLSS 4

But you are comparing native 4k on 5090 vs 1080p on PS6. You can just use DLSS and see 2x performance uplift on 5090. This is nowhere near PS6=5090 that mild claims.

Wouldn't PS6 be also bandwidth constrained with 160-bit bus and also lower cache amounts vs say an 5080 (not to mention 5090)?

Depends how well universal compression works. PC GPUs still have large L2/L3 caches and massive native memory bandwidth (2TB/s for 5090).

 

[Loxus]

Wonder what's the chance this will turn out like they are describing.

The progress they did with RDNA4 should tell you.

 

[Zathalus]

I argued against the PS6 being 5070 level, because that king of uplift seems low. Equalling the 5090 in any kind of game is simply not going to happen. Yes, Radiance cores, but Nvidia has had the equivilant of that since Turing, albeit slower. RDNA4 still lack OMM, SER, Ray Traversal Hardware, and several other RT features that Blackwell has. Going from that to utterly dominating Blackwell in one generation is very unlikely. Extremely unlikely going off AMDs track record.

Personally I think the 5070ti is probably going to be the closest match in terms of compute and RT. Just with less VRAM. The cancelled 24GB 5070ti Super would probably have been by far the closet match in terms of GPUs right now. Maybe a bit more powerful, but probably within 10% or so

 

[Ellis]

And UE5 will still stutter like a motherfucker.

Thankfully this is far less of an issue on console. Still bad enough enough though.

I can't wait for that pos engine to die.

 

[slrbomber]

Not really based on this.

This wont work well on all kinds of data!! Sparse datasets, yes, but other non sparse datasets I really wonder how well it will work. Raw bandwidth increase is always a safe bet. I trust Cerny but this sounds like a huge risk.

 

[Loxus]

This wont work well on all kinds of data!! Sparse datasets, yes, but other non sparse datasets I really wonder how well it will work. Raw bandwidth increase is always a safe bet. I trust Cerny but this sounds like a huge risk.

It says all available data within the GPU though, so who knows.

I guess we'll have to wait and see.