Digital Foundry Article Technical Discussion [2023]

[pjbliverpool]

Memory allocation is probably the easiest thing Nvidia could fix for their next GPU line. But people are really feeling the mediocrity of the 4000 series of GPUs. I don't know if it's technologically possible for Nvidia to improve the situation with greater power increases per gen. But they should atleast get their product prices down.

This is a massive misjudgement of the situation IMO. Technically the 4000 series are spectacular. The issue here is that NVIDIA are trying to sell each SKU into a higher performance tier than its suited for at the same or even higher prices than the previous gen GPU of the same performance tier. Essentially paying more for less.

If something like following were the case then no one at all would be complaining and instead we'd be hailing the 4xxx series as being as great as Ampere or better.

4080 @ $799
4070Ti -> 4070 @ $549
4070 -> 4060Ti @ $449
4060Ti -> 4050Ti @ $329
4060 -> 4050 @ $199

 

[Opeth]

It's not even restricted to 4k. Outliers sure, but in any circumstance seeing it losing to the previous gen at 1080/1440p is truly shocking.

View attachment 8940

View attachment 8941

Abysmal! And Rich will still conclude on a positive note I'm sure I love Rich but he is very kind on GPU reviews I feel

 

[Flappy Pannus]

Abysmal! And Rich will still conclude on a positive note I'm sure I love Rich but he is very kind on GPU reviews I feel

Tweets by DF, and discussion on DF Direct on it so far don't really indicate that. I think they'll be more measured in including the advantages of DLSS3 than other outlets probably, but the impression I get is they're not exactly enthused.

Edit: Welp, here you go.

There's seven pages here of RTX 3070 Ti analysis and data in this review, but the more I used the card, the more I realised that this is massive overkill for a GPU where the salient points can be summed up in one paragraph. The RTX 4060 Ti is essentially a hybrid RTX 3060 Ti/RTX 3070 with the added bonus of DLSS 3, a good media block and extra efficiency, but beyond that it's a missed opportunity and a disappointment. There are some positive points to highlight, but they're crushed by the sense that Nvidia isn't delivering good value to the sector of the market that most demands it.

Nvidia believes that the total addressable market is so dominated by 8GB cards that developers will continue to support them and that it's not an issue at 1080p. I disagree on this and think the situation we've seen on so many 2023 games will continue. Cross-gen is over and all of the titles I just mentioned have had issues precisely because PlayStation 5 was the main target platform. And for a $400 GPU, I don't just expect console parity - I expect better. On the GPU side, the 4060 Ti can do this. Its raster power exceeds the consoles, its RT performance is way ahead. However, for console-grade textures and room for the data required for RT, 8GB is already straining the limits. I have serious reservations about the future-proofing of the card in a world where the far less capable RTX 3060 12GB performed better than 3060 Ti in all of the above titles at launch. Put simply: if we're having problems now, what about next year... or the year after that?

The 8GB framebuffer also deals even more damage to the upgrade path from the Pascal or Turing cards. Maybe you're looking for a replacement to a GTX 1070 or RTX 2060 Super? Well, if you had VRAM problems with those products, you'll have them with the RTX 4060 Ti too. Nvidia's solution is to offer a 16GB alternative, but unfortunately the firm has positioned it with an astonishing $100 price premium - a 25 percent price rise which simply cannot be justified and of course, it then makes you wonder whether you should spend a further $100 on top of that instead for the RTX 4070, which as the benchmarks reveal is typically 35 percent faster (and much more in some cases at 4K).

 

[Flappy Pannus]

This is a massive misjudgement of the situation IMO.

I think you may be focusing too much on the 'mediocrity' comment - this is a 'mediocre' lineup when you factor in the value proposition of these cards, products are only as good as their price. Technically sure, there's nothing inherently wrong with Nvidia's designs here, they're dusting AMD on wattage per frame, nobody is providing a more efficient product. RDNA3 is indeed a 'mediocre' generation as it barely moves the needle in any aspect against the previous, ADA makes some significant strides.

I took "People are feeling the mediocrity" to just mean that the product lineup is not that enticing for what they're asking, so not sure you're really saying anything different here.

Technically the 4000 series are spectacular. The issue here is that NVIDIA are trying to sell each SKU into a higher performance tier than its suited for at the same or even higher prices than the previous gen GPU of the same performance tier. Essentially paying more for less.

If something like following were the case then no one at all would be complaining and instead we'd be hailing the 4xxx series as being as great as Ampere or better.

4080 @ $799
4070Ti -> 4070 @ $549
4070 -> 4060Ti @ $449
4060Ti -> 4050Ti @ $329
4060 -> 4050 @ $199

And note even this ideal scenario still has memory bus width being reduced gen-on-gen for most of the models, so this is still accounting somewhat for the increased manufacturing costs.

Those kinds of prices would have glowing reviews and are far more in line with recent releases. The 4060ti at $329 would still perhaps receive some critique for 8GB considering the 3060 at the same price had 12 and it's more critical now than 2 years ago, but otherwise this lineup would be widely praised regardless of outlet. Goddamn a 4070 at that price would be in my system now, or even a 4070ti. There would be stories of sold out inventory and camping out in front of stores, instead of say, a single person.

Obviously a fantasy though as this point and from the latest earning reports, it doesn't look like the market is going to scare Nvidia into making significant changes to their gaming product matrix anytime soon.

 

[Silent_Buddha]

Obviously a fantasy though as this point and from the latest earning reports, it doesn't look like the market is going to scare Nvidia into making significant changes to their gaming product matrix anytime soon.

And considering the current incredibly high demand for AI processors, NV aren't going to sweat it even if consumer GPU sales were to tank due to high prices. That just means more wafer starts can be moved to AI chips.

Regards,
SB

 

[Flappy Pannus]

And considering the current incredibly high demand for AI processors, NV aren't going to sweat it even if consumer GPU sales were to tank due to high prices. That just means more wafer starts can be moved to AI chips.

Regards,
SB

 

[function]

It's not even restricted to 4k. Outliers sure, but in any circumstance seeing it losing to the previous gen at 1080/1440p is truly shocking.

View attachment 8940

View attachment 8941

I wonder if the PCI-e bus is anything to do with this (particularly Hogwarts)? Sure, when things are going well PCIe4 x 8 is fine - you lose effectively nothing from your half width bus. But when you spill out in main ram because 8GB VRAM shit the bed, traffic across PCIe could skyrocket. That hurts majorly even on full width cards.

I'm still on PCIe 3, and intend to hang on another year or two. That leaves me at effectively PCIe4 x 4 with a 4060Ti, and with no Re-bar it'd hurt even more. I know this card isn't really targetted at me, but by god £400 for an 8GB, half width PCI card? I can still feel offended on behalf of the PCMR, nVIDIA.

 

[neckthrough]

And note even this ideal scenario still has memory bus width being reduced gen-on-gen for most of the models, so this is still accounting somewhat for the increased manufacturing costs.

Is it though? The reduced bus width is compensated by a massive increase in L2$ which again consumes a bunch of TSMC's precious sand.

 

[Remij]

Obviously a fantasy though as this point and from the latest earning reports, it doesn't look like the market is going to scare Nvidia into making significant changes to their gaming product matrix anytime soon.

No company is above "F around and find out".

 

[Flappy Pannus]

No company is above "F around and find out".

We'll see, the AI race is definitely heating up and Nvidia will have more competitors, but it's also a software issue too - it's not just producing a GPU, and on the software side Nvidia has a huge head start.

As I've said before, the only real threat I see on the horizon in the relatively near future is for AMD to finally flex it's X86 liscense advantage and bring APU's that can at least target the midrange. That may be happening in 2024, but it's also AMD, so definitely not holding my breath.

 

[pjbliverpool]

I think you may be focusing too much on the 'mediocrity' comment - this is a 'mediocre' lineup when you factor in the value proposition of these cards, products are only as good as their price. Technically sure, there's nothing inherently wrong with Nvidia's designs here, they're dusting AMD on wattage per frame, nobody is providing a more efficient product. RDNA3 is indeed a 'mediocre' generation as it barely moves the needle in any aspect against the previous, ADA makes some significant strides.

I took "People are feeling the mediocrity" to just mean that the product lineup is not that enticing for what they're asking, so not sure you're really saying anything different here.

Yes agreed its mediocre from a value proposition perspective. Unfortunately there's no alternative right now.

The 4060ti at $329 would still perhaps receive some critique for 8GB considering the 3060 at the same price had 12 and it's more critical now than 2 years ago,

Yeah the memory size was the main reason I dropped it down to the x050 tier where I think that capacity would be acceptable. The 3060 does muddy things a bit but its memory capacity was an aberration even last gen.

 

[Flappy Pannus]

Is it though? The reduced bus width is compensated by a massive increase in L2$ which again consumes a bunch of TSMC's precious sand.

Hence, 'somewhat'. If it was a purely 1:1 exchange in terms of cost, it wouldn't make sense to bother with L2 upgrades at all and just keep the bus width from the previous gen. I'm just saying even this highly optimistic scenario doesn't completely ignore the increased costs of maintaining that same bus width as the previous gen, they could have even reduced the L2 cache as well and these models would still be a significantly better value that what they're offering today.

While I would prefer that product stack instead, it terms of silicon cost, it's possible the most cost efficient method Nvidia would take to head off not all, but a significant portion of the criticism would be just to double the vram. 3060ti, 16GB, $399. 4070 TI, 24GB - even if you had to say, increase the MSRP by $50 that would still have resulted in a far more positive reception I would think. Wouldn't be an absolute glowing reception like it would be with just the cards being shifted down a tier, but it would quell the biggest critique.

Obviously though still in fantasy territory, but that I think has the lowest sacrifice to Nvidia's margins...for their cost of gaming chips. The problem is when you're getting into 24+ GB, you can potentially be threatening the market position of your AI-focused cards.

 

[eastmen]

With the massive increase in speed and massive drop in price isn't it time to just add an nvme connector to graphics cards. Just let the game stream in textures and fill up dozens of hundreds of gigs of textures from storage or main system memory and then let the graphics card grab what it needs from the on board nvme. That way once its loaded in you don't have to worry about consuming cpu resources or bandwidth from the system ram to cpu to gpu. Its just a slower gpu cache

 

[Flappy Pannus]

With the massive increase in speed and massive drop in price isn't it time to just add an nvme connector to graphics cards. Just let the game stream in textures and fill up dozens of hundreds of gigs of textures from storage or main system memory and then let the graphics card grab what it needs from the on board nvme. That way once its loaded in you don't have to worry about consuming cpu resources or bandwidth from the system ram to cpu to gpu. Its just a slower gpu cache

PCIE-16X from main ram provides far more bandwidth than nvme SSD's. If transferring textures from main ram across PCIE is a bottleneck now, then the solution is hardly introducing an even slower path that requires specific developer support - and potentially significantly increasing loading time as you'd still need the CPU to decompress all those textures to the GPU's NVME before they can be used, and what game would ship with 100+GB of textures? The only way the CPU is not involved is by using Directstorage, and the bottleneck there is actually getting developers to use it, not transferring those compressed textures across PCIE. As it stands now, the CPU would to decompress them, then send them to the GPU, where it would write them to nvme, then into vram. Makes no sense.

The cheapest, most efficient solution is obvious - more vram. It's just not the solution that retains Nvidia's high margins or poses no threat to their professional lineup.

 

[eastmen]

PCIE-16X from main ram provides far more bandwidth than nvme SSD's. If transferring textures from main ram across PCIE is a bottleneck now, then the solution is hardly introducing an even slower path that requires specific developer support - and potentially significant increasing loading time as you'd still need the CPU to decompress all those textures to the GPU's NVME before they can be used, and what game would ship with 100+GB of textures? The only way the CPU is not involved is by using Directstorage, and the bottleneck there is actually getting developers to use it, not transferring those compressed textures across PCIE. As it stands now, the CPU would to decompress them, then send them to the GPU, where it would write them to nvme, then into vram. Makes no sense.

The cheapest, most efficient solution is obvious - more vram. It's just not the solution that retains Nvidia's high margins or poses no threat to their professional lineup.

Would it take less time to send data from the nvme drive through the cpu to the gpu vs sending it from an nvme on the gpu to the on board ram ?

Games are getting bigger and will start shipping with more and more textures.

 

[Flappy Pannus]

Would it take less time to send data from the nvme drive through the cpu to the gpu vs sending it from an nvme on the gpu to the on board ram ?

But you still have to send it to the CPU, as the CPU needs to decompress the textures. Your solution only provides some potential marginal benefit if Directstorage 1.1 is utilized, and even there it still would be highly questionable as if the GPU is decompressing the textures, then have it do that from main system ram right into vram, significantly better latency and bandwidth.

You're just adding in an extra step. There is no potential gaming situation in this gen, and potentially next, where this makes sense vs. just utilizing 32GB of system ram which is quickly becoming the standard regardless. You're trying to solve a potential problem in a Directstorage world where we haven't even gotten one game to utilize GPU decompression yet.

 

[eastmen]

But you still have to send it to the CPU, as the CPU needs to decompress the textures. Your solution only provides some potential marginal benefit if Directstorage 1.1 is utilized, and even there it still would be highly questionable as if the GPU is decompressing the textures, then have it do that from main system ram right into vram, significantly better latency and bandwidth.

You're just adding in an extra step. There is no potential gaming situation in this gen, and potentially next, where this makes sense vs. just utilizing 32GB of system ram which is quickly becoming the standard regardless. You're trying to solve a potential problem in a Directstorage world where we haven't even gotten one game to utilize GPU decompression yet.

Why would it need to go to the cpu , you can store uncompressed textures on the nvme on the drive. 2TB nvme's are $70 bucks now

 

[Flappy Pannus]

Why would it need to go to the cpu , you can store uncompressed textures on the nvme on the drive. 2TB nvme's are $70 bucks now

I give up. It's a great idea.

 

[eastmen]

I give up. It's a great idea.

You don't have to be an ass. I am just asking questions. Or is it that you don't know how how performance would be affected with uncompressed textures. It's okay if you don't know as I don't either. Which is why I originally said

Would it take less time to send data from the nvme drive through the cpu to the gpu vs sending it from an nvme on the gpu to the on board ram ?

Games are getting bigger and will start shipping with more and more textures.

 

[Flappy Pannus]

You don't have to be an ass. I am just asking questions. Or is it that you don't know how how performance would be affected with uncompressed textures. It's okay if you don't know as I don't either. Which is why I originally said

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.

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 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).