Digital Foundry Article Technical Discussion [2023]

[arandomguy]

So remember how this whole discussion came about: vram limitations. So while Nvidia's markup of $100 for an extra 8GB is egregious (from most reports I've read an additional 8GB of GDDR6, especially standard GDDR the 4060ti uses and not GDDR6X, is $20-$30), you instead want to tackle it by GPU vendors installing an SSD mount (to a heat source like a GPU card!), have the user purchase an additional nvme ssd and install it, and have game developers support this and install 200+GB of uncompressed textures to this SSD at every game install?

I think the advantages, as well as the odds of that happening, are basically nil. Even with Nvidia's markup, the cost of just adding 8, 12 or even 16GB of vram to cards would likely be more economical, but also more useful - 8-16 gigs of 200+GB/sec vram over ~5GB/sec SSD. We would have Directstorage 1.1 being used widely and 16GB vram as the entry level before SSD's on GPU's would be something developers would be targeting. And once you have GPU decompression, then you lose the point of having an SSD installed on the GPU, as pci-e has more than enough bandwidth to transport the textures to the GPU in their compressed state (and uncompressed too, as the fastest nvme bandwidth is a fraction of pcie).

Nvidia's VRAM cost problem is in part due to opportunity cost and not material cost. Nvidia GPUs are at present much more desirable for content creation workloads (not even close in comparison to AMD/Intel, the gaming brand/platform delta doesn't remotely compare) and that segment will largely pay based on VRAM. This means any lower SKUs with high VRAM effectively end up cannibalizing higher up the stack.

You can kind of see this but they're likely going to do as much as possible to decouple and add VRAM (real or effective) in a way that is good enough for gaming but still pushes users to pay higher up the stack for content creation.

People often bring up why the RTX 4090 is so cheap gen on gen relative to the rest of the stack, well one reason is that it doesn't have the massive VRAM improvement that the 3090/ti brought for that market segment.

 

[eastmen]

Fine, apologies for being a bit of a dick but I don't know else to communicate you're just adding an additional roadblock to GPU's using textures effectively, this wouldn't affect the problem of vram. I just don't see how this solves that problem in any way.

The only situation where this would bring an advantage that I can see is in cutting out the decompression step, which primarily would benefit CPU usage. So maybe you could have the textures decompressed upon game install and saved to the GPU's SSD, where they wouldn't have to be decompressed by the CPU during gameplay (?). So maybe for something like TLOU, you wouldn't have these large CPU spikes when traversing between areas. Except in that case, now you've added an I/O bottleneck where these uncompressed textures have to be pulled from the SSD in their uncompressed state. It also begs the question that if you're fine with extending your install time of games so that they're writing 200+GB of completely uncompressed textures to disk, why would that be an easy ask for the end user if the SSD is one on the GPU and not on your motherboard? It's still storage space required which will be in competition with other games, just like any SSD.

I would have thought it would save cpu usage but also main system bus usage freeing up bandwidth. I guess a limitation would be the intial install but you might be able to hold a few games at a time in the nvme if disk size is large enough.

Directstorage makes far more efficient use of SSD's in no small part due to bypassing a lot of the legacy cruft with the Windows filesystem, but a significant benefit of the GPU decompression in 1.1 (and in the PS5/Xbox) is twofold: It massively reduces CPU usage for the decompression step, but as well it massively increases the throughput of the SSD's since the effective bandwidth from the SSD can be double/triple it's rated gb/s based on how efficiently the type of stored texture is compressed. If you store the texture uncompressed on the SSD you lose that benefit.

So couldn't you still use Direct storage's gpu decompression with compressed textures and data on the nvme attached to the graphics card. You'd still gain the benefit of by passing main system ram and the pci-e bus.

So remember how this whole discussion came about: vram limitations. So while Nvidia's markup of $100 for an extra 8GB is egregious (from most reports I've read an additional 8GB of GDDR6, especially standard GDDR the 4060ti uses and not GDDR6X, is $20-$30), you instead want to tackle it by GPU vendors installing an SSD mount (to a heat source like a GPU card!), have the user purchase an additional nvme ssd and install it, and have game developers support this and install 200+GB of uncompressed textures to this SSD at every game install?

Yes , I was just figuring that having the data local would act more like l1/l2/l3 cache in a cpu. There are certainly hurdles you'd have to get over like you said.

I think the advantages, as well as the odds of that happening, are basically nil. Even with Nvidia's markup, the cost of just adding 8, 12 or even 16GB of vram to cards would likely be more economical, but also more useful - 8-16 gigs of 200+GB/sec vram over ~5GB/sec SSD. We would have Directstorage 1.1 being used widely and 16GB vram as the entry level before SSD's on GPU's would be something developers would be targeting. And once you have GPU decompression, then you lose the point of having an SSD installed on the GPU, as pci-e has more than enough bandwidth to transport the textures to the GPU in their compressed state (and uncompressed too, as the fastest nvme bandwidth is a fraction of pcie).

Yes but to be fair this discusion of vram has been going on for awhile. I remember when I was deciding on the 3080 people doubted 10gigs was enough. So I am not sure if the issue is going away. Thank you for explaining it more depth.

 

[Flappy Pannus]

Nvidia's VRAM cost problem is in part due to opportunity cost and not material cost. Nvidia GPUs are at present much more desirable for content creation workloads (not even close in comparison to AMD/Intel, the gaming brand/platform delta doesn't remotely compare) and that segment will largely pay based on VRAM. This means any lower SKUs with high VRAM effectively end up cannibalizing higher up the stack.

Yes I know, I've said the same in other posts. I'm just saying that even just from a material cost perspective, this wouldn't be an attractive option for Nvidia's margins on gaming cards alone even disregarding their segmentation with the pro cards, or cost effective for end users either even if Nvidia kept the increase to the bare minimum. Adding another SSD + heatsink, even as cheap as they are, large enough (1TB+) to avoid lengthy re-uploads of uncompressed textures when you switch games, to avoid the issue of an 8GB card...when the 16GB option is $100 more.

It absolutely shouldn't be $100 more, it should be the same $399 of course, but ultimately this scheme looks to solve a problem (that it can't) vs the better solution that will be coming in a couple of months already. By the time (again, if this even had some technical reasoning) this solution could get any dev support we'll be standardized on GDDR7+ and have higher densities across the board regardless.

Nvidia wants that segmentation with vram no doubt, but they're not completely oblivious to the concerns of the gaming market, as offering a $500 GPU with 16GB shows. Not great, but even they recognize 8GB cards are a dead end, regardless of how they may want to segment their lineup by vram.

 

[Henry swagger]

I hate this tlou port has caused this vram obsession.. vram wont make your gpu faster with ugly textures

 

[Silent_Buddha]

I hate this tlou port has caused this vram obsession.. vram wont make your gpu faster with ugly textures

Possibly not but it will allow potentially faster and better use of RT and reconstruction techs (DLSS, for example) since both of those require more memory over not using them. That is something that is particularly painful for cards like the 3070 or 4060 Ti.

Regards,
SB

 

[Flappy Pannus]

I hate this tlou port has caused this vram obsession.. vram wont make your gpu faster with ugly textures

It's not just TLOU though. The problem is all the other games that were right on the cusp of 8GB, and most of those were cross-gen ports. Now that we're getting ports from PS5/SX 'native' titles, as well the two prominent features that Nvidia is promoting - RT and DLSS3 frame gen, also take up considerable amounts of memory - and they can't scale as easily as textures can. A Plague's Tale for example is routinely brought up for its effective texture mananagement vs. TLOU, and it's true - it has room to spare on 8GB cards, even at high resolutions with excellent detail. Until you turn on RT and frame gen...and then 8GB cards are tapped out.

It's one thing for a raster-only game like TLOU to blow through the vram budget that the majority of cards have on the market it's targeting and not provide scaling options where the only choice is massive stuttering or basically, textures not loading properly. It's another thing for a new, $400 product released 2 years into a console's lifecycle - where that entire console also costs $400 - to still be stuck with 8gb, a standard set in cheaper cards 7 years ago.

It's all about what's a reasonable ask. A vram 'obsession' would be expecting vram to scale with GPU density and demanding 24-32GB on mainstream cards, that's not reasonable. Expecting the same bus width on extremely expensive new processes where every mm counts is perhaps not reasonable. Expecting 16GB on mainstream cards in 2023 after seven years of vram stagnation for the majority of the market, especially when the vram demands are increased due to the very features the leading GPU manufacturer promotes? That's reasonable.

Wish we heard more about Samsung's GDDR6W since November, but I guess we're still in the 'announcement' stage and not 'we've made this cost effective' stage. But damn that looks to solve both the bandwidth reductions from reduced bus width and vram capacity in one go.

 

[Henry swagger]

It's not just TLOU though. The problem is all the other games that were right on the cusp of 8GB, and most of those were cross-gen ports. Now that we're getting ports from PS5/SX 'native' titles, as well the two prominent features that Nvidia is promoting - RT and DLSS3 frame gen, also take up considerable amounts of memory - and they can't scale as easily as textures can. A Plague's Tale for example is routinely brought up for its effective texture mananagement vs. TLOU, and it's true - it has room to spare on 8GB cards, even at high resolutions with excellent detail. Until you turn on RT and frame gen...and then 8GB cards are tapped out.

It's one thing for a raster-only game like TLOU to blow through the vram budget that the majority of cards have on the market it's targeting and not provide scaling options where the only choice is massive stuttering or basically, textures not loading properly. It's another thing for a new, $400 product released 2 years into a console's lifecycle - where that entire console also costs $400 - to still be stuck with 8gb, a standard set in cheaper cards 7 years ago.

It's all about what's a reasonable ask. A vram 'obsession' would be expecting vram to scale with GPU density and demanding 24-32GB on mainstream cards, that's not reasonable. Expecting the same bus width on extremely expensive new processes where every mm counts is perhaps not reasonable. Expecting 16GB on mainstream cards in 2023 after seven years of vram stagnation for the majority of the market, especially when the vram demands are increased due to the very features the leading GPU manufacturer promotes? That's reasonable.

Wish we heard more about Samsung's GDDR6W since November, but I guess we're still in the 'announcement' stage and not 'we've made this cost effective' stage. But damn that looks to solve both the bandwidth reductions from reduced bus width and vram capacity in one go.

I would have no problem if the games that demanded high vram actual looked as good plaue tale.. tlou has ps3 textures which i d'not like

 

[Silent_Buddha]

I would have no problem if the games that demanded high vram actual looked as good plaue tale.. tlou has ps3 textures which i d'not like

You're far too focused on TLOU PC. We're talking about other titles with RT and DLSS on modern 8 GB cards where the memory becomes insufficient to have both RT and DLSS enabled (both increase memory usage), basically memory becomes a bottleneck and the cards aren't able to perform to the full capabilities of the GPU. In the case of the 8 GB 4060 Ti, I think some titles with RT were actually shown to run slower once DLSS 3 was enabled due to memory pressure.

Regards,
SB

 

[SmooTh]

 

[Below2D]

Disgusting card.

 

[SmooTh]

 

[jlippo]

Disgusting card.

Card is ok, the price is disgusting.
I'm sure it would do fine at 200$.

 

[Overstander]

So the Inside Gamer Leak about the "Q" hand held was spot on... Right down to the name..
Gives their PS5 Pro story a bit more trust i guess..
And we will get a proper benchmark finally with Ratchet & Clank Rift Apart.

Too bad they did not release hardware req for now.
And no word on Direct Storage either ...
And since Nixxies is doing the port People will likley be more causious calling it a bad port if a Ryzen 5 3600 and a RTX20xx Card is not enough to match PS5s Modes...

 

[davis.anthony]

So the Inside Gamer Leak about the "Q" hand held was spot on... Right down to the name..
Gives their PS5 Pro story a bit more trust i guess..
And we will get a proper benchmark finally with Ratchet & Clank Rift Apart.

Too bad they did not release hardware req for now.
And no word on Direct Storage either ...
And since Nixxies is doing the port People will likley be more causious calling it a bad port if a Ryzen 5 3600 and a RTX20xx Card is not enough to match PS5s Modes...

In the RT games modes an RTX 2060 should be more than enough to keep up with PS5, with RT disabled an RTX2060 shouldn't get close to PS5.

But the benefit of the RTX20060 is DLSS which nearly every Nvidia owner will use in reality so it'll likely perform the same with no RT and offer better image quality.

 

[Globalisateur]

So the Inside Gamer Leak about the "Q" hand held was spot on... Right down to the name..
Gives their PS5 Pro story a bit more trust i guess..
And we will get a proper benchmark finally with Ratchet & Clank Rift Apart.

Too bad they did not release hardware req for now.
And no word on Direct Storage either ...
And since Nixxies is doing the port People will likley be more causious calling it a bad port if a Ryzen 5 3600 and a RTX20xx Card is not enough to match PS5s Modes...

We already have benchmarks with a focus on CPU / I/O comparisons: Spider-man and TLOU. People not liking the results (notably in the case of TLOU) doesn't change the fact that those are fair benchmarks. In both cases you need a very powerful CPU (about twice faster) to have a comparable PS5 performance. Overall no surprises here as we know PS5 I/O is worth about 9 zen 2 core CPUs.

For more GPU (raster) comparison we need to use a much less CPU bound game (with much less emphasis on assets decompression on the fly) like Death Stranding. Here you need a Nvidia 3060 / 3070 GPU depending of the scene to mach PS5 GPU performance.

There are others benchmarks, sure, but at least in the case of Spider-man and Death Stranding, we know both PC and PS5 versions are very optimized to run on their respective hardware so that make those comparisons quite fair.

 

[SmooTh]

 

[davis.anthony]

We already have benchmarks with a focus on CPU / I/O comparisons: Spider-man and TLOU.

I've not seen a single benchmark of those games that isolates I/O enough to actually produce useable data relating to I/O performance.

People not liking the results (notably in the case of TLOU) doesn't change the fact that those are fair benchmarks.

As per my above comment, I/O can't be isolated well enough to actually produce a meaningful test.

In both cases you need a very powerful CPU (about twice faster) to have a comparable PS5 performance.

This is just flat out false.

Overall no surprises here as we know PS5 I/O is worth about 9 zen 2 core CPUs.

This is also flat out false, try not to get so caught up in marketing hype.

For more GPU (raster) comparison we need to use a much less CPU bound game (with much less emphasis on assets decompression on the fly) like Death Stranding. Here you need a Nvidia 3060 / 3070 GPU depending of the scene to mach PS5 GPU performance.

Last gen game using a last gen engine so how is that relevant to current gen comparisons?

There are others benchmarks, sure, but at least in the case of Spider-man and Death Stranding, we know both PC and PS5 versions are very optimized to run on their respective hardware so that make those comparisons quite fair.

We know that with RT enabled PC pulls way a head of PS5 and we also know that RT isn't going anywhere.

 

[pjbliverpool]

We already have benchmarks with a focus on CPU / I/O comparisons: Spider-man and TLOU. People not liking the results (notably in the case of TLOU) doesn't change the fact that those are fair benchmarks.
In both cases you need a very powerful CPU (about twice faster) to have a comparable PS5 performance. Overall no surprises here as we know PS5 I/O is worth about 9 zen 2 core CPUs.

You do not need a CPU that's twice as fast as the PS5 CPU to have comparable performance in those games. That's 7800X3D / 13900KS territory in a lightly threaded scenario like a game, either of which comfortably destroy the PS5's performance in both of those games. The 9 Zen 2 cores comparison is senseless, because neither of these games are streaming data at anything close to the maximum throughput of the PS5's IO. We've seen from Alex's video that Spiderman at least peaks at around 500MB/s.

That's not to say you don't need a more powerful CPU on the PC side if you're doing CPU based decompression and real time BVH generation (as opposed to streaming BHV), so yes, in a game that does not leverage GPU based decompression that also has moderate to high levels of in game streaming, it's reasonable to expect higher CPU requirements on the PC side. In fact even with GPU decompression, PC's have more work to do on the CPU side, particularly where RT is involved so we can certainly expect R&C to require more than a 3600X to match the PS5's performance, particularly given that CPU 25% fewer cores than the PS5.

For more GPU (raster) comparison we need to use a much less CPU bound game (with much less emphasis on assets decompression on the fly) like Death Stranding. Here you need a Nvidia 3060 / 3070 GPU depending of the scene to mach PS5 GPU performance.

No you don't. You need a 2080 to roughly match it or a 2080 Super to slightly exceed it. Which is pretty much exactly where you would expect the PS5's GPU to land in a well optimised title that doesn't use RT (Ratchet & Clank does).

There are others benchmarks, sure, but at least in the case of Spider-man and Death Stranding, we know both PC and PS5 versions are very optimized to run on their respective hardware so that make those comparisons quite fair.

Assuming R&C does not use Direct Storage and GPU based decompression, it's fair to expect pretty high CPU requirements. It may also have high RAM requirements to sidestep some of the streaming requirements, both for CPU load reasons, and to accommodate users with slower storage mediums.

On the GPU side, if it's possible to match settings (and this may be quite difficult given the PS5 version uses DRS), then I'd expect the very light RT load will allow it to perform reasonably high up the Nvidia stack. Perhaps 2070S - 2080 level if VRAM doesn't become a bottleneck (which is may do for the same reasons as the high RAM requirements I mentioned above). However since each system has different strengths, it would also be fair to consider how the PS5 would perform at what will likely be the much higher RT settings that the PC version will offer vs those same cards.

In other words, where the mix of settings are tailored towards the console (console matched settings) I expect relatively high PS5 performance. But I also expect settings options which will run well on those Turing cards and their newer equivalents that would put some serious hurt on the PS5 (before we consider DLSS). Unfortunately that's not something we can test. All of these tests are always the best case scenario for the console.

 

[Nesh]

Man, everything they were showing from the previous spiderman games looked much more pleasing in that analysis even though Insomniac did improve some assets.
There is something about the lighting in Spiderman2 that just doesnt work very well.

 

[SmooTh]

.