Discussion Ada/'Lovelace'? Next gen Nvidia gaming architecture speculation

[NTMBK]

Is there any way to calculate rough cost of the 3080 8nm 628mm dies vs 4-5nm 379mm of 4080?

Phone up JHH

 

[TESKATLIPOKA]

GeForce RTX 4060 Laptop GPU shows 20% higher 3DMark performance than RTX 3060



This is just a bit worse than what @Tigerick posted. It could be explained by having lower TGP. If that's 24SM Full ADA107, then that's a very good performance.
For comparison, the fastest 3050Ti st notebookcheck managed only 6122 points, so this Ada is 64% faster.
Let's not forget TGP is higher than what 3050Ti has.
It will be interesting to compare It at the lowest possible TGP for efficiency purposes.

 

[TESKATLIPOKA]

NVIDIA ‘accidentally’ confirms GeForce RTX 4070 Ti GPU specifications

It's with DLSS enabled on both + Frame generation enabled on Ada.

 

[Aapje]

@gdansk

Also keep in mind that the chip is merely the priciest component, but less than half of the cost of the card, so even if the chip got X% more expensive, the cost of the overal card should go up by less than X%.

@amenx

Unfortunately, we have very little information about what Nvidia paid to Samsung. I suspect that the rumors about half the cost were impacted by the low yields and that as yields went up, they had to pay more. So I have my doubts whether those rumors are true.

It certainly didn't make much sense that AMD could sell the 6900 XT for about $600, yet supposedly the 3080 couldn't be sold for that kind of price without making a loss, if Nvidia actually paid so little for Ampere chips.

@TESKATLIPOKA

If it's about twice as fast as a 3080 in MSFS with DLSS 3, that means that it is actually about the same speed, so...

But the true proof of the pudding will be the reviews, because Nvidia is as reliable as your local bridge salesman.

 

[Tigerick]

GeForce RTX 4060 Laptop GPU shows 20% higher 3DMark performance than RTX 3060

View attachment 73664

This is just a bit worse than what @Tigerick posted. It could be explained by having lower TGP. If that's 24SM Full ADA107, then that's a very good performance.
For comparison, the fastest 3050Ti st notebookcheck managed only 6122 points, so this Ada is 64% faster.
Let's not forget TGP is higher than what 3050Ti has.
It will be interesting to compare It at the lowest possible TGP for efficiency purposes.

Damn NV, I heard from China news that NV going to artificially limit RTX4050-70 TGP by around 20W. That explains low performance. The first batch of 13th gen notebook going to use RTX30 series. NV intends to ship RTX40 series one month later...

As for AMD new platform, they are going to ship in limited quantity in March, more on Q2

 

[TESKATLIPOKA]

Damn NV, I heard from China news that NV going to artificially limit RTX4050-70 TGP by around 20W. That explains low performance. The first batch of 13th gen notebook going to use RTX30 series. NV intends to ship RTX40 series one month later...

As for AMD new platform, they are going to ship in limited quantity in March, more on Q2

Don't you know If only N33 or also N32 will be announced at CES 2023?

 

[Tigerick]

Don't you know If only N33 or also N32 will be announced at CES 2023?

No news about N32 but we should be expecting it during CES, the issue is availability...

 

[Kaluan]

GeForce RTX 4060 Laptop GPU shows 20% higher 3DMark performance than RTX 3060

View attachment 73664

This is just a bit worse than what @Tigerick posted. It could be explained by having lower TGP. If that's 24SM Full ADA107, then that's a very good performance.
For comparison, the fastest 3050Ti st notebookcheck managed only 6122 points, so this Ada is 64% faster.
Let's not forget TGP is higher than what 3050Ti has.
It will be interesting to compare It at the lowest possible TGP for efficiency purposes.

If they're pricing it towards ODMs similarly (or even higher) than the prior GA106/3060L than it really doesn't matter to the market, you won't be paying 3050Ti laptop prices for a 4060 laptop.

Seeing as 3050Ti TGP config maxes out (</=80W) under where 4060 allegedly begins (>/=85W) it's hard to see these 2 fit in in similar chasis/designs anyway.

Laptop SKU performance is hard the frame too, you almost never know the TGP.

 

[TESKATLIPOKA]

If they're pricing it towards ODMs similarly (or even higher) than the prior GA106/3060L than it really doesn't matter to the market, you won't be paying 3050Ti laptop prices for a 4060 laptop.

Seeing as 3050Ti TGP config maxes out (</=80W) under where 4060 allegedly begins (>/=85W) it's hard to see these 2 fit in in similar chasis/designs anyway.

Laptop SKU performance is hard the frame too, you almost never know the TGP.

It could be that RTX 4500(18SM Ada107) 6GB will replace 3050Ti 4GB with comparable TGP and RTX 4060(24SM Ada107) 8GB will replace RTX 3060 6GB.

 

[jpiniero]

It could be that RTX 4500(18SM Ada107) 6GB will replace 3050Ti 4GB with comparable TGP and RTX 4060(24SM Ada107) 8GB will replace RTX 3060 6GB.

Videocardz is saying that the 4050L's TDP range is also 85-140. So it's looking like the 3050 Ti's replacement is... the 3050 6 GB?

 

[TESKATLIPOKA]

Videocardz is saying that the 4050L's TDP range is also 85-140. So it's looking like the 3050 Ti's replacement is... the 3050 6 GB?

I wonder If 3050 6GB is based on GA107 or GA106.
If It's GA107, then It has only 96bit bus + 3x 2GB modules. If It's based on GA106, then it can have even 192bit bus + 6x 1GB modules.
I am not sure what to think about this segmentation.
It could look like this.
GA107 - RTX 3050 6GB 96bit for <80W TGP with 2560SP and <=1700MHz.
ADA107 - RTX 4050 6GB 96bit for 80-120W with 2304SP and >=2200MHz.
ADA107 - RTX 4060 8GB 128bit for >=120W with 3072SP and >=2200MHz.

 

[jpiniero]

I wonder If 3050 6GB is based on GA107 or GA106.

Definitely GA107. Memory is probably faster to compensate somewhat for the bus drop.

 

[TESKATLIPOKA]

Definitely GA107. Memory is probably faster to compensate somewhat for the bus drop.

That's very likely. 16gbps memory would be 33% faster than what RTX 3050(Ti) uses.
I have to wonder If they have so much GA107 in stock, then what about GA106?

Just a few days left until announcement.

 

[poke01]

I am interested to see the RTX 4070 Ti. It's a full die too.

 

[KompuKare]

Is there any way to calculate rough cost of the 3080 8nm 628mm dies vs 4-5nm 379mm of 4080?

Roughly isn't that hard. Image they are both square (they are not but sqrt of 628 / 379 is easier), then we get yield per wafer of:

Of course, we can only guess at the cost per wafer and the defect density.
TSMC have at least quoted 0.09 for 7nm. Samsung I have seen no quotes for, but they were meant to have been poor especially for those huge dies. On the other hand, their wafer prices should have started lower hence I started my chart at $5K.

 

[jpiniero]

See I think nVidia could have say done respins of Ampere instead of doing Ada, and done a GA102 with maybe more shaders but less memory controllers, faster memory, etc and charged about what the 4070 Ti costs. But would that really be a better product.

 

[amenx]

Roughly isn't that hard. Image they are both square (they are not but sqrt of 628 / 379 is easier), then we get yield per wafer of:

Of course, we can only guess at the cost per wafer and the defect density.
TSMC have at least quoted 0.09 for 7nm. Samsung I have seen no quotes for, but they were meant to have been poor especially for those huge dies. On the other hand, their wafer prices should have started lower hence I started my chart at $5K.

Thanks. I had $16,000 as a rough figure for 5nm:

Taiwan’s TSMC could reportedly charge US$20,000 per 3nm wafer | Taiwan News | 2022-11-23 15:44:00

High cost of production and long manufacturing times for cutting edge chips pushing up prices | 2022-11-23 15:44:00

www.taiwannews.com.tw

 

[Kaluan]

Calling it now (and technically again):
Vaporware 4070Ti "FE" - $799 MSRP
AIB 4070Ti - $899 MSRP (and up)

Just like the fake $999 2080Ti MSRP of yesterday

There are no situations were nVidia doesn't have it's cake and eat it too. 😂

 

[KompuKare]

Thanks. I had $16,000 as a rough figure for 5nm:

Taiwan’s TSMC could reportedly charge US$20,000 per 3nm wafer | Taiwan News | 2022-11-23 15:44:00

High cost of production and long manufacturing times for cutting edge chips pushing up prices | 2022-11-23 15:44:00

www.taiwannews.com.tw

Okay, so IF Samsung's 8nm was ~$7,000 then the costs are pretty similar.
And even if the wafer costs is a bit more, AD104 should work with cheaper PCB, a smaller memory bus, simpler cooling etc. Ergo, costs could potentially be quite very similar. But hey, what do we know? Nvidia wants to increase their ASP during a recession and so do AMD, Intel etc.

 

[jpiniero]

Ooh, the 4050/4060/4070 laptop has a TDP range of 35-115 W. That's a very big range. Also nVidia said (4050) laptops start at $999 which is a tad high.

 

[TESKATLIPOKA]

Ooh, the 4050/4060/4070 laptop has a TDP range of 35-115 W. That's a very big range. Also nVidia said (4050) laptops start at $999 which is a tad high.

Here you have It.
Specs:

Price:

They mentioned performance, but most likely with DLSS 3 on. So fake frames on.
I will wait for real tests.

 

[biostud]

Hopefully the $799 4070Ti will translate to $599 7800XT, or whatever the full N32 will be named.

 

[TESKATLIPOKA]

Here are specs for RTX 30 series:


Here is Nvidia RTX 40 series webpage.
RTX 4050 has the same number of SM as RTX 3050 Ti and 6GB Vram.
RTX 4060 has only 24SM, but more Vram.
Boost frequency is ~35-42% higher.
Both of them will end up faster than the previous generation even without DLSS 3, RTX 4050 by a large margin.
This all looks good.
Price is not low at $999, but we don't know the rest of the system.
Which CPU, how much RAM, SSD and so on.

What I find especially interesting is this:
RTX 3050TI 20SM 1695MHz 4GB 80W
RTX 4050 20SM 1605MHz 6GB 35W
Performance comparison will be very interesting.

I don't like that RTX 4070 and 4090 didn't increase memory.
RTX 4080 did increase and also decrease the Vram, because you can find RTX 3080 Laptop with both 8GB and 16GB Vram.

I think Nvidia did a pretty good job with Ada laptop.

 

[jpiniero]

The specs of the 3050 laptop is new/changed. That's pretty wild they have all those different combinations.

 

[Ranulf]

I saw that the bald guy still calls it a xx70 "TIE" card. At least Nvidia is consistent in their delusions on pricing, naming and fps numbers. The gaming trailers with DLSS 3.0 numbers being emphasized really helped bring home how silly it all is anymore.