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

[SteinFG]

Videocardz seems to think the 4060L is AD107 based actually.

AD107 is really tiny, don't think it will be used for the 4060.
Edit: If it will actually beat 3060 by a lot, then maybe it's true, but I doubt.

 

[Aapje]

If they really market the AD103 as a mobile 4090, that is a new level of shamelessness.

 

[MrTeal]

I don't know, that seems like a pretty typical level of mobile GPU shamelessness.

 

[gdansk]

Yeah, it's pretty typical. At least it isn't AD104.

 

[jpiniero]

AD107 is really tiny, don't think it will be used for the 4060.
Edit: If it will actually beat 3060 by a lot, then maybe it's true, but I doubt.

Looked into it and it's looking pretty likely that the 4060L is indeed AD107 based. That'd be a pretty sizable SM cut gen-to-gen.

 

[SteinFG]

Looked into it and it's looking pretty likely that the 4060L is indeed AD107 based. That'd be a pretty sizable SM cut gen-to-gen.

What makes you believe it? To me it just seems weird. 4050/Ti Laptop still has to exist.
My guess:
AD106 (36SM full die) -> 36SM version for 4070L, 30SM version for 4060L
AD107 (24SM full die) -> 24SM for 4050Ti L, 18SM version for 4050L
This is completely arbitrary, just for a look later when actual specs will be revealed.

Edit: maybe I'm wrong here. One of the leaks

https://twitter.com/x/status/1546136686660186112

pointed only to a single GPU on AD106 (probably 4070L) and two GPUs on AD107 (Probably 4060L and 4050L). So 4060L might get a very big cut. There will be no gen-on-gen improvement for 60 laptop series then. I'm less optimistic now :|

 

[jpiniero]

https://twitter.com/x/status/1602761028584247298

4060 Ti rumors. Yes that's less cores than the 3060 Ti and not even the full AD106.

Let me guess, 3070 Ti performance for $599?

 

[Saylick]

https://twitter.com/x/status/1602761028584247298

4060 Ti rumors. Yes that's less cores than the 3060 Ti and not even the full AD106.

Let me guess, 3070 Ti performance for $599?

Blehhh, if true that would be zero perf/$ increase. What's next? RTX 3060 levels of performance for $399 via an RTX 4050?

I can hear the marketing already.... "The RTX 4050. The perfect 1440p 144 Hz* gaming card."

*DLSS 3 games only.

 

[Thunder 57]

All on a 128 bit memory bus!

However, its memory bandwidth of 288 GB/s will be a whopping 41% lower compared to its predecessor.

https://www.tomshardware.com/news/preliminary-geforce-rtx-4060-ti-specs-leak

I really hope people give NVIDIA the finger this generation.

 

[Saylick]

All on a 128 bit memory bus!

https://www.tomshardware.com/news/preliminary-geforce-rtx-4060-ti-specs-leak

I really hope people give NVIDIA the finger this generation.

I hope so, too, but people are too busy pointing the finger at AMD for not bringing competition because they are a convenient scapegoat to the problem of rising GPU prices.

I don't know if people realize that the GPU space, like many markets, is a zero-sum game. Every dollar Nvidia makes is a dollar that AMD doesn't make, and it's especially true in a shrinking market. In general, it takes money to make money in an industry with high R&D and capital costs. Worse yet, when you have a shrinking market with rising capital costs, only those with a dominant enough position to capture the majority portion of the shrinking pie survive. I mean, just look at Intel vs. Samsung vs. TSMC for a textbook example.

TSMC has better node ---> customers flock to TSMC for fabbing cutting edge products ---> TSMC breaks even faster on the cost to develop the node ---> TSMC has more money ---> TSMC makes better node. Rinse and repeat. If you're Samsung, good luck trying to generate revenue when everyone else wants the competitor's product. To catch up when you're lagging behind in that market requires an even larger transfusion of money because you likely don't have as much revenue pouring in from customers to offset costs. Case in point, we have the Korean and American government pouring subsidy money to get Samsung and Intel back into a competitive position.

I've posed the question on these forums before, but with Nvidia being so dominant in the market, even if AMD did not flub N31, I have to seriously question whether or not it's competitiveness would have any bearing on how Nvidia prices their GPUs. When AMD finally had a leg-up over Intel in the server market, Intel didn't actually respond for years because AMD could not ship EPYC at a scale that would actually do any harm to Intel. To this day, AMD still only has sub-30% server market share and I doubt it will ever exceed 50%. There's just not enough runway before Intel gets their act together. Against Nvidia, a runway for AMD to gain market share likely never will exist. Nvidia is just too dominant, and people ultimately choose Nvidia over AMD even when AMD forces Nvidia to drop prices.

If AMD decided to call it quits in the GPU market, I wouldn't even be surprised if people blame AMD for Nvidia's monopoly at that point.

 

[tamz_msc]

I hope so, too, but people are too busy pointing the finger at AMD for not bringing competition because they are a convenient scapegoat to the problem of rising GPU prices.

I don't know if people realize that the GPU space, like many markets, is a zero-sum game. Every dollar Nvidia makes is a dollar that AMD doesn't make, and it's especially true in a shrinking market. In general, it takes money to make money in an industry with high R&D and capital costs. Worse yet, when you have a shrinking market with rising capital costs, only those with a dominant enough position to capture the majority portion of the shrinking pie survive. I mean, just look at Intel vs. Samsung vs. TSMC for a textbook example.

TSMC has better node ---> customers flock to TSMC for fabbing cutting edge products ---> TSMC breaks even faster on the cost to develop the node ---> TSMC has more money ---> TSMC makes better node. Rinse and repeat. If you're Samsung, good luck trying to generate revenue when everyone else wants the competitor's product. To catch up when you're lagging behind in that market requires an even larger transfusion of money because you likely don't have as much revenue pouring in from customers to offset costs. Case in point, we have the Korean and American government pouring subsidy money to get Samsung and Intel back into a competitive position.

I've posed the question on these forums before, but with Nvidia being so dominant in the market, even if AMD did not flub N31, I have to seriously question whether or not it's competitiveness would have any bearing on how Nvidia prices their GPUs. When AMD finally had a leg-up over Intel in the server market, Intel didn't actually respond for years because AMD could not ship EPYC at a scale that would actually do any harm to Intel. To this day, AMD still only has sub-30% server market share and I doubt it will ever exceed 50%. There's just not enough runway before Intel gets their act together. Against Nvidia, a runway for AMD to gain market share likely never will exist. Nvidia is just too dominant, and people ultimately choose Nvidia over AMD even when AMD forces Nvidia to drop prices.

If AMD decided to call it quits in the GPU market, I wouldn't even be surprised if people blame AMD for Nvidia's monopoly at that point.

Jensen has said before that what AMD does has no bearing on their own business. You can't blame NVIDIA for the biggest reason why they're successful - they produce more GPUs, an order of magnitude more than AMD.

 

[JujuFish]

Jensen has said before that what AMD does has no bearing on their own business.

He can say all he wants, but actions speak louder than words.

 

[amenx]

Pricing has all to do with supply. Their initial stocks were feeble (of the 4080 at least) so they could easily price it at $1200. If they're locked into a binding multi-billion $ contract with TSMC with very little wiggle-room (can only delay shipments by a quarter or so), eventually they will be swamped with inventory that cannot be moved at silly prices. I can easily see the 4080 at $799 some 6 months down the road.

 

[MrTeal]

All on a 128 bit memory bus!

https://www.tomshardware.com/news/preliminary-geforce-rtx-4060-ti-specs-leak

I really hope people give NVIDIA the finger this generation.

The 4080 has a slight SP deficit to the 3080 Ti and is still 30% faster at 4k thanks to arch improvements and clock speed increases even with 2/3rds the bus width. I wouldn't be surprised if this proposed 4060 Ti performed about the same over the 3060 Ti.

30% gen over gen increase for the same card sucks even if they keep it at $400, but this is the company and generation that released a 4080 that's 50% faster at 4k than the launch 10GB 3080 for only a 70% higher price.

 

[Thunder 57]

The 4080 has a slight SP deficit to the 3080 Ti and is still 30% faster at 4k thanks to arch improvements and clock speed increases even with 2/3rds the bus width. I wouldn't be surprised if this proposed 4060 Ti performed about the same over the 3060 Ti.

30% gen over gen increase for the same card sucks even if they keep it at $400, but this is the company and generation that released a 4080 that's 50% faster at 4k than the launch 10GB 3080 for only a 70% higher price.

128 bit bus width is anemic. Faster memory won't be able to make up for it. Also it should be performing far better than the 3060 Ti, not on par with it. "Only 70% higher price", haha! 8GB is also too little for a card like that.

Whatever, let the chumps by this crap. Hopefully AMD does well in the mid range and NVIDIA can't sell crap like that without huge price cuts.

 

[MrTeal]

Oh I agree. It's part of the reason I'm having a hard time putting faith in these most recent Kopite7kimi numbers; they just seem way too little gen over gen to be worth it. That 4060 Ti spec should be the 4060, but given the prices and spec of the rest of the lineup it actually seems plausible.

 

[Mopetar]

128 bit bus width is anemic. Faster memory won't be able to make up for it. Also it should be performing far better than the 3060 Ti, not on par with it. "Only 70% higher price", haha! 8GB is also too little for a card like that.

The much larger L2 cache will compensate for the lower bus size more than the faster memory will. Surprisingly, Nvidia didn't have some flashy name for it.

The memory is far more concerning though. Not all games need 8 GB of VRAM, but we've already seen several that do, and not all of them are the latest titles either.

I'm not sure Id want to pay more than $300 (maybe $400 if it's a premium build) for an 8 GB card. Polaris was offering 8 GB for as low as $200 5 years ago and inflation isn't enough to explain the price disparity today.

 

[scineram]

Bad news. I wouldn't expect over 8GB form the new generations until above $400. Maybe at $399 if they feel generous.
Unless Intel wants to give away free candy I guess.

 

[TESKATLIPOKA]

RTX 4060(Ti) with 34 SM even at ~2.6GHz is likely comparable to RTX 3070Ti in performance based on TFlops. For some It may not be that impressive.
What is pretty impressive is that 32MB L2 halved the required memory bandwidth. The only real downside is 8GB Vram, but of course we need reviews to see the real performance.
220W TBP is pretty high.

 

[MrTeal]

RTX 4060(Ti) with 34 SM even at ~2.6GHz is likely comparable to RTX 3070Ti in performance based on TFlops. For some It may not be that impressive.
What is pretty impressive is that 32MB L2 halved the required memory bandwidth. The only real downside is 8GB Vram, but of course we need reviews to see the real performance.
220W TBP is pretty high.

The 3060 Ti (along with the OG 3080 10G) were the real values of the last generation for team green so maybe anything will seem worse, but getting 3070 Ti performance (1.2x 3060 Ti @ 1440p) in the 4060 Ti just seems like a crappy uplift, even if it comes in at the same price and power. AD106 is probably going to come in at close to half the size of GA104, it really feels more that a part that should top out at the 60 tier.

 

[TESKATLIPOKA]

The 3060 Ti (along with the OG 3080 10G) were the real values of the last generation for team green so maybe anything will seem worse, but getting 3070 Ti performance (1.2x 3060 Ti @ 1440p) in the 4060 Ti just seems like a crappy uplift, even if it comes in at the same price and power. AD106 is probably going to come in at close to half the size of GA104, it really feels more that a part that should top out at the 60 tier.

Based on specs, I would like AD104 or even AD107 in my next laptop If It's cheap enough.
Nvidia did a great job with Ada, the only real problem is the price.

Personally, I would change the specs of ADA chips to this:

	SM (Cuda)	TMU	ROP	L2 Cache	Memory subsystem	Bandwidth	TBP
ADA102	160 (20,480) [+25%]	640	160 [+25%]	80 MB [+25%]	48 GB 384-bit 24gbps	1152 GB/s [+20%]	450 W [+25%]
ADA 102 cut	128 (15,360) [+33.3%]	480	128 [+33.3%]	64 MB [+33.3%]	40 GB 320-bit 24gbps	960 GB/s [+25%]	360 W [+30.9%]
ADA103	96 (12,288) [+20%]	284	96 [+20%]	48 MB [+20%]	32 GB 256-bit 24gbps	768 GB/s [+14.3%]	275 W [+19.6%]
ADA103 cut	80 (10240) [+25%]	336	80 [+25%]	40 MB [+25%]	28 GB 224-bit 24gbps	672 GB/s [+22%]	230 W [+24.3%]
ADA104	64 (8192) [+23%]	256	64 [+23%]	32 MB [+23%]	24 GB 192-bit 23gbps	552 GB/s [+20%]	185 W [+19.4%]
ADA104 cut	52 (6656) [+30%]	208	52 [+30%]	26 MB [+30%]	20 GB 160-bit 23gbps	460 GB/s [+25%]	155 W [+29%]
ADA106	40 (5120) [+25%]	160	40 [+25%]	20 MB [+25%]	16 GB 128-bit 23gbps	368 GB/s [+24.3%]	120 W [+26%]
ADA106 cut	32 (4096) [+33.3%]	128	32 [+33.3%]	16 MB [+33.3%]	16 GB 128-bit 18.5gbps	296 GB/s [+30%]	95 W [+32%]
ADA107	24 (3072) [+20%]	96	24 [+20%]	12 MB [+20%]	12 GB 96-bit 19gbps	228 GB/s [+19%]	72 W [+20%]
ADA107 cut	20 (2560) [100%]	80	20 [100%]	10 MB [100%]	12 GB 96-bit 16gbps	192 GB/s [100%]	60 W [100%]

I got rid of GDDR6x and tried to make the relative increase in specs comparable between chips along with TBP. The main problem with these specs is that I had to use clamshell.

AMD pretty much botched RDNA3, so I don't have high expectations from them, maybe N32 will be a decent chip, N31 is a FLOP and N33 also doesn't look very promising.
I am eagerly waiting for CES 2023 for some new info.

Edit: I expanded my table to also include the cutdown versions.

 

[jpiniero]

AD106 is probably going to come in at close to half the size of GA104, it really feels more that a part that should top out at the 60 tier.

Clearly there was an attempt to keep the costs down because of how much more expensive TSMC is.

 

[amenx]

GN showing history of Nvidia xx80 series cards launch prices vs die sizes.

 

[TESKATLIPOKA]

Lenovo has a Phoenix or Raptor Lake based Ideapad Pro prepared for CES 2023 paired with RTX 4050 6GB. Link

 

[jpiniero]

Lenovo has a Phoenix or Raptor Lake based Ideapad Pro prepared for CES 2023 paired with RTX 4050 6GB. Link

I actually think it's legitimately a new 3050 SKU and not a typo. But we will see.