Nvidia RTX 4060 Ti specs confirmed by die shot and retail listing

[Porkous]

I don't think 8GB of vram is "controversial". For this money you can nearly buy a Playstation 5 AND THAT HAS 16gigs of VRAM.

One could say the SSd speeds are so fast on PS5, that the SSd could be a huge page pool memory of 812GB


A complete twist to this alternative, of a PS5, is the Rog Ally handheld with it's z1 extreme. Supposedly, its performance is that of about 85% (8.6 TFLOPs) of PS5 (according to AMD slides)...
I reckon PS5 consumption is 200w... Let's see how is z1 extreme this powerful...
This one has 16Gb as well.


Edit. Of course, those TFLOPS are if the chip is running, probably, at full speed TDP. So the power would be about 3 times less in reality. Take it with a grain on salt.

So probably about the power of GTX 1050? give or take.
Or rather 1650?

AMD Radeon 780M Specs

AMD Phoenix, 2700 MHz, 768 Cores, 48 TMUs, 32 ROPs

www.techpowerup.com

 

[yRaz]

In FY2022, Sony took hardware revenues in its gaming sector of roughly $6.4 billion. AMD's financial year is Dec-Dec (Mar-Mar for Sony) but it saw a revenue of $6.8 billion for the same sector. Approximately $3.8 billion of that was Sony, so this paper napkin exercise puts half the price of a PS5 just to cover the cost of the APU. So yes, it does indeed have 16 GB of GDDR6, but the unit itself makes no, or next to no, profit. One can't expect that with a graphics card.


In both a PS5 and a PC, the GPU only works with specific memory -- in the case of the former, the unified RAM, and the onboard local RAM with the latter. The fact that all bar one PC games store a copy of the assets in system memory doesn't affect the actual VRAM usage.


Even with bulk discounts, 8 GB of high-speed GDDR6 is unlikely to cost $20. Compared to the rest of the DRAM market, such products are relatively low-volume sellers (in comparison, not in raw figures) -- Jon Peddie estimated that 38 million discrete graphics cards were shipped in 2022, whereas 286 million PCs were shipped in that period and this doesn't account of discrete DIMM sales, too.

But even if it was super cheap, adding on more RAM does have further costs. To have 16 GB on a 128-bit bus means using 8 modules in clamshell mode. So the PCB needs to account for the additional traces and the power delivery system to RAM has to account for double the demand. Minor things, of course, but when graphics card sales have been constantly declining for the past decade, vendors aren't going constantly absorb additional costs when everything else has increased in price.

While the significant markup of high-end graphics cards can cover such things, every cent counts at the other end of the scale. AIB vendors typically have very low operating margins and as Intel's graphics revenues have shown, it's easy to make a substantial loss in this particular market.

As frustrating as it is to see yet another 8 GB card come out, AMD, Intel, and Nvidia are probably feeling very jittery about sales right now and are going to keep costs down as much as they can.

The PS5 example was a gross over simplification but GDDR6 isn't expensive. But going to VRAM usage, we can't actually get a proper number of how much is being used. We can see how much is being pre-allocated but we can't see what the GPU is using. In games we have assets being stored in system memory and in VRAM. Through some creative engineering Sony almost got around this entirely with the speed of their storage.

As far as putting more memory on a card, I feel it's nVidia's fault entirely for choosing a monolithic die and limiting the bus width. There are problems with the IO section that AMD has gotten around with their chiplet designs, but the fact of the matter is that my 8800GT from 2007 has a 256 bit bus width. nVidia needs to get their **** together, $400 for a card with a 128bit bus and 8 gigs of VRAM is entirely unacceptable. I'm not even against 8 gig cards, I'm against the price.

We can talk about the price if you want but my sources have told me that 16 gigs of GDDR6 goes for around $50. On a $200 card you can't hide that cost but we're talking about a $400 card. In one of my other machines I have a 2gig 1050ti and that has acceptable performance for a secondary rig. To be perfectly honest, I don't know how much longer that 1050ti is gonna last because it has some pretty bizzare performance issues in newer titles. I did only pay $70 for that card, though. So for $70 I'm willing to forgive A LOT.

I'm not mad about 8 gigs of VRAM, I'm mad at 8 gigs for the current price especially since the 6700xt is readily available for ~$320 with 12 gigs. What's the 4060ti going to be? Another card where nVidia can tell us about how great DLSS is?

The 4090 is a performance monster, it's really the only good card to come out of this generation. The weird part is that it's "too fast" and people are returning them after looking at the price tag. Somehow nVidia has made some of the best cards we've seen and people don't want them because they aren't price appropriately.

I remember Linus talking about his NCIX days where he had a phonecall with nVidia and they wouldn't tell him the price of the card. He said nVidia told him, "what does it matter, they're going to sell." to which he replied, "Well I need to know the price so that I know how many I can order" The era of nVidia cards sell simply because they're nVidia cards is over

 

[ScottSoapbox]

I predict a lot of “Super” versions of cards with more VRAM launching in the near future.

 

[neeyik]

But going to VRAM usage, we can't actually get a proper number of how much is being used. We can see how much is being pre-allocated but we can't see what the GPU is using. In games we have assets being stored in system memory and in VRAM.

Actually, we can! Keep your eyes peeled for something on that very topic, very soon

As far as putting more memory on a card, I feel it's nVidia's fault entirely for choosing a monolithic die and limiting the bus width. There are problems with the IO section that AMD has gotten around with their chiplet designs, but the fact of the matter is that my 8800GT from 2007 has a 256 bit bus width.

While AMD has gained the advantage of being able to have a 384-bit bus on a 300 mm2 die, by using chiplets, Nvidia's decision to stay with a monolithic design doesn't mean they couldn't have gone with a larger bus with the AD104/6 -- the AD102, for example, has the same bus width as the Navi 31 and while it's roughly 20% larger than the full Navi 31, it's smaller than the GA102 and TU102, both of which also have a 384-bit bus.

They did it to make the chips for the models that generate smaller profit margins (well in the world of normal pricing...) have better wafer yields, by upping the L2 cache massively to utilize the fact that SRAM scales far better than IO logic with node shrinks.

We can talk about the price if you want but my sources have told me that 16 gigs of GDDR6 goes for around $50. On a $200 card you can't hide that cost but we're talking about a $400 card.

How many 16 GB cards have been $400 at launch, though? The cheapest Radeon with that amount was $530. There was one, of course, that was under that price -- Intel's A770, at $330, but its entire graphics sector has made a loss every single quarter. Samsung, Micron, and SK Hynix are all suffering a big downturn in revenue and profits; they're really not going to be selling 18-20 Gbps GDDR6 modules for a few dollars, when the entire market for it is on a huge downturn. They won't sell any more modules, even if they did, as the demand for them is decreasing.

I'm not mad about 8 gigs of VRAM, I'm mad at 8 gigs for the current price especially since the 6700xt is readily available for ~$320 with 12 gigs. What's the 4060ti going to be? Another card where nVidia can tell us about how great DLSS is?

Don't forget that all evidence points to the RX 7600 having 8 GB too. AMD is definitely enjoying the fruits of its RDNA 2 lineup now but I do wonder if it's made a mistake in dropping the prices so low. Unless the new models are notably faster than their predecessors, the 6000 series is going to cannibalize sales -- partly because of the price, but also because VRAM amount is the current hot topic.

The era of nVidia cards sell simply because they're nVidia cards is over

Almost certainly, yes. Not for the first time in its history, Nvidia has made a mess out of the Ada generation, by not reading the room very well. Shame really, as the actual products are really very good.

 

[takaozo]

So a 8 GB card will do for 720p upscaled to 1080p, but will a 4060 keep up with the newest games?

Just look at the FPS for 3070 with RT enabled. I guess this will be the perf for 4060Ti.

Even the 4070 barely hit 60+ FPS.

 

[yRaz]

Actually, we can! Keep your eyes peeled for something on that very topic, very soon

That would be very interesting, I look forward to it.

While AMD has gained the advantage of being able to have a 384-bit bus on a 300 mm2 die, by using chiplets, Nvidia's decision to stay with a monolithic design doesn't mean they couldn't have gone with a larger bus with the AD104/6 -- the AD102, for example, has the same bus width as the Navi 31 and while it's roughly 20% larger than the full Navi 31, it's smaller than the GA102 and TU102, both of which also have a 384-bit bus.

I've been told that nVidia is no longer trying to target the desktop market. Their strategy is now aimed at the mobile and laptop markets. The 4060ti is a laptop chip. They cut the IO down to make it easier to fit on laptop PCBs where the is limited realestate. It's small bus isn't because they're trying to get over on everyone, it's because they never intended for it to be a gaming card in the first place. They have massive $30-40,000 AI GPUs and things like the 4090 are a "cut down" version of those professional GPUs. Essentially 4090s just use a GPU that doesn't meet the clock speed or power requirements of what's going into their professional level products.

How many 16 GB cards have been $400 at launch, though? The cheapest Radeon with that amount was $530. There was one, of course, that was under that price -- Intel's A770, at $330, but its entire graphics sector has made a loss every single quarter. Samsung, Micron, and SK Hynix are all suffering a big downturn in revenue and profits; they're really not going to be selling 18-20 Gbps GDDR6 modules for a few dollars, when the entire market for it is on a huge downturn. They won't sell any more modules, even if they did, as the demand for them is decreasing.

We are years passed the last launch. It's been fun watching Intel's GPUs progress, but it's interesting that you said "demand for the modules is decreasing". I think the absolute opposite is true. Consumer demand for more GDDR is there but people like nVidia aren't giving consumers what they want. We could bring the price down but I think the real issue is that we're stuck in a feedback loop of "nVidia isn't buying enough so the $1/GB is going up". I feel everyone from the manufactures, developers and consumers want more VRAM. The 8gig issue is a bit overblown but I'm happy to jump on the bandwagon. The thing is, we are reaching the end of life for 8 gig cards. It's not close but it's not far away either.

Don't forget that all evidence points to the RX 7600 having 8 GB too. AMD is definitely enjoying the fruits of its RDNA 2 lineup now but I do wonder if it's made a mistake in dropping the prices so low. Unless the new models are notably faster than their predecessors, the 6000 series is going to cannibalize sales -- partly because of the price, but also because VRAM amount is the current hot topic.

Well we'll have to see how it's priced. I think most consumers are willing to forgive a lot if the price is right. I'm still mad at AMD over the 6500xt only having 4 PCI-e lanes. As a budget card it was naturally going to be used with older hardware making the PCIe 3.0 vs 4.0 a major issue. I understand that multilayer PCBs are expensive but making it an 8X card instead of 4 would have brought performance up significantly and added very little to cost. Another one of those "just add $5-10 to the cost and give us a good product"

Almost certainly, yes. Not for the first time in its history, Nvidia has made a mess out of the Ada generation, by not reading the room very well. Shame really, as the actual products are really very good.

As much fun as I'm having riding the nVidia hate bandwagon I really wish they didn't mishandle Ada. There is a lot of really cool tech in those cards and a lot things to get excited about, it's just the price hurts to even think about. Looking at the price-to-performance premium on those cards, nVidia is charging about $100 for DLSS. DLSS is awesome, It's one of the coolest things to come out in years, but it's not worth $100. And the way nVidia is handling it, DLSS does not excuse the underlying issues with how they're marketing Ada. Ada could have been one of the best product launches in nVidia's history if they didn't mishandle it so horribly.

 

[neeyik]

I've been told that nVidia is no longer trying to target the desktop market. Their strategy is now aimed at the mobile and laptop markets. The 4060ti is a laptop chip. They cut the IO down to make it easier to fit on laptop PCBs where the is limited realestate. It's small bus isn't because they're trying to get over on everyone, it's because they never intended for it to be a gaming card in the first place.

Not sure where that's been said, but I'm not convinced by that argument at all. Nvidia isn't struggling in the laptop GPU market and unit sales in that sector are falling off a cliff, so why would they be targeting it further? Besides, the AD106 actually isn't super small -- at a little under 200 mm2 in size, it's roughly the same as the TU117 and GA107, both of which were commonly used in laptops. AMD's Navi 24, on the other hand, is half the size of all those chips; now that's a chip that's properly aimed for laptop market!

They have massive $30-40,000 AI GPUs and things like the 4090 are a "cut down" version of those professional GPUs. Essentially 4090s just use a GPU that doesn't meet the clock speed or power requirements of what's going into their professional level products.

Those AI GPUs aren't the same architecture as the AD102 (they use Hopper) and it's not just the top-end Ada chips that are used in Tesla and Quadro products, the AD104 is too. Nvidia does design its main architecture to be used in as many sectors as possible, of course, and has done for a while. AMD does too, but to a lesser extent -- RDNA is primarily a gaming-focused architecture, though it does get used in the Radeon Pro range; CDNA is AMD's Hopper-equivalent.

The thing is, we are reaching the end of life for 8 gig cards. It's not close but it's not far away either.

Indeed!

 

[yRaz]

Not sure where that's been said, but I'm not convinced by that argument at all. Nvidia isn't struggling in the laptop GPU market and unit sales in that sector are falling off a cliff, so why would they be targeting it further? Besides, the AD106 actually isn't super small -- at a little under 200 mm2 in size, it's roughly the same as the TU117 and GA107, both of which were commonly used in laptops. AMD's Navi 24, on the other hand, is half the size of all those chips; now that's a chip that's properly aimed for laptop market!

I'm just repeating what I've been told and considering the subject I don't want to say too much or I put my source at risk. nVidia is expecting an increase in sales in the mobile segment, the AD106 is primarily meant for the 4070m with them "recycling" dies that aren't up to spec into the 4060 and 4060ti. We might see some Tegra like products soon as companies are showing an interest in the 'steam deck' form factor. I'd be thrilled to have a high-end steam deck that I can dock like a switch and take wherever I want.

Those AI GPUs aren't the same architecture as the AD102 (they use Hopper) and it's not just the top-end Ada chips that are used in Tesla and Quadro products, the AD104 is too. Nvidia does design its main architecture to be used in as many sectors as possible, of course, and has done for a while. AMD does too, but to a lesser extent -- RDNA is primarily a gaming-focused architecture, though it does get used in the Radeon Pro range; CDNA is AMD's Hopper-equivalent.

So that what was my mistake, they cut down the H100 die and repurpose it into lower end professional series/workstation cards. They disable parts of the H100 and sell them as lower teir quadro cards. I'm going back through my email to find this stuff. There was something interesting going on with the AD102 and the 4090's. I spoke at length with someone about what's going on at nVidia and it's A LOT. I'll get back to you on the 4090 story. Everything from manufacturing numbers to sales projections going out the next 5 years.

EDIT:
nvm on the 4090 story. Modified AD102 dies are just being tested for self driving vehicles. Apparently car manufacturers don't want to pay 30% of the vehicles value to have self driving tech. It's cheaper to add a 4090 to a car than an H100

 

[brucek]

more, it's the fact that microsoft and sony are losing money on console hardware... they can, coz they're making their cash on games, online service, and peripherals etc... it's nearly impossible to compare this to a GPU price on pc

but the unit itself makes no, or next to no, profit. One can't expect that with a graphics card.

Yes, I understand the business model is different from the manufacturer's side and that therefore from their perspective it's apples & oranges.

The thing is though from the consumer point of view, $399 is still $399. If the GPU manufacturers, aided perhaps by sloppy industry porting practices, have reached a point where it is going to take $599 or more of GPU to roughly match $399 of console (in delivered gaming experience, not raw/theoretical power on paper), even in the case where the entire rest of the computer is "free" / sunk cost due to the consumer already having the non-gaming part for other purposes anyway, it is eventually going to influence consumer behavior. Especially those who aren't already committed "PC Gamers" from past decision cycles.

The 2023 PS5 digital at $399 has maintained the same price level as the 2013 PS4. Less true for Nvidia, where at end of 2013 you could get the 770 for $329 or the 780 for $499, but today those same-level SKUs cost much more at $599 / $1,199.

 

[neeyik]

H100 die and repurpose it into lower end professional series/workstation cards. They disable parts of the H100 and sell them as lower teir quadro cards.

Quadro cards haven’t used HBM for years so GH100 aren’t being used in those; the A-range and Teslas maybe, though. Anyway, well off topic now so best just stick with discussing the 4060.

 

[Alfatawi Mendel]

Graphics card manufacturers release VRAM crippled (and expensive) cards. Tech press: POOR OPTIMISATION !!!

 

[Rolfil]

So in 3 months, it's a 1080p gpu med/high settings, without ray tracing...

yea I know I'm a little "extrem" but for the price... unbelievable, and I'm sure it'll be a top sell...

But if your buying a card you aren't going to replace for 2-3 years then 8gb isn't going to cut it. Everyone knows this. Also NVIDIA paid way too much to TSMC for their wafers which is why the price is so high. This is really going to hurt them. 4070's aren't selling and the lower end isn't worth buying.

There is a real opportunity for AMD here. They mostly put more ram on their cards. If they go with 16gb for at least the 7700 and higher cards and say 12gb on the 7600 range at lower prices, then even if they are only making smaller profits then they are going to gain market share.

If NVIDIA don't respond until next year then even Intel might give them a run for their money at the low-mid range cards.

 

[kira setsu]

But if your buying a card you aren't going to replace for 2-3 years then 8gb isn't going to cut it. Everyone knows this. Also NVIDIA paid way too much to TSMC for their wafers which is why the price is so high. This is really going to hurt them. 4070's aren't selling and the lower end isn't worth buying.

There is a real opportunity for AMD here. They mostly put more ram on their cards. If they go with 16gb for at least the 7700 and higher cards and say 12gb on the 7600 range at lower prices, then even if they are only making smaller profits then they are going to gain market share.

If NVIDIA don't respond until next year then even Intel might give them a run for their money at the low-mid range cards.

8gb isnt nearly enough now, my last gpu before I upgraded was a 980ti and it was solid until the ps5/xbone gen started plus I was playing at 1080p, now with a 4080 on a 4k tv I wanted to see how much vram certain games gobbled up, hogwarts legacy cranked with rt on was just below 15gigs of usage, I know most games wont hit that but yeah its possible now.(imho rt isnt worth it)

even if the new consoles have a "shared" 16gigs of memory then new gpu's should consider that the borderline required amount, mainly because a normal pc has other stuff going on that isnt nearly as optimized as a console and is gonna brute force its way to the finish line.

if a friend of mine was building a new pc to get into it or just upgrade and they were looking at an 8gb card I'd say no, either go amd, or bite the bullet and spend more if you really want nvidia.

 

[MasterMace]

A $250 card priced at $400 and it only has 8gb vram. Dead on arrival, just like the 4070.

 

[Hodor]

A $250 card priced at $400 and it only has 8gb vram. Dead on arrival, just like the 4070.

You can bet that AMD sponsored games will need 12 GB to run properly. That might shake NVidia a bit.

 

[MasterMace]

You can bet that AMD sponsored games will need 12 GB to run properly. That might shake NVidia a bit.

Jedi: Survivor is an AMD sponsored game and requires less than 12gb RAM. It just got released.

 

[Hodor]

Jedi: Survivor is an AMD sponsored game and requires less than 12gb RAM. It just got released.

10.5 GB?