RTX 5xxx / RX 8xxx speculation

[Gideon]

Not every card is going to cost over $1000 though. The 5080 might get the price raised back up to $1200 to match the 4080's initial launch price but everything else under it is going to be less than $1,000. If Nvidia starts resting on their laurels then that only allows AMD an opportunity to take the performance crown and I'm sure Mr. Leather Jacket would never allow that to happen.

I feel a 5080 priced over 1,000 bucks will be a dust collector for retailers, sort of like the current 4080. Honestly I was a bit surprised how well the 4090 sold, so that may still do well at a higher price. Profit is far more important to the leather jacket right now, otherwise stock value might drop.

 

[LukeTbk]

I feel a 5080 priced over 1,000 bucks will be a dust collector for retailers,

A $1,199 5080 that goes +15% the 4090 if the 5090 is $2000 and has 24 GB of VRAM could still sell..... specially if it has over 1600 TFLOPS at FP8 / 3000 peak INT TOPS

There is no 4080 super at msrp on newegg outside the PNY-Zotac brands I think

Everything at msrp is sold out at Best Buy:
https://www.bestbuy.com/site/search...UPER&sp=+currentprice skuidsaas&st=4080+super

 

[Decko87]

A $1,199 5080 that goes +15% the 4090 if the 5090 is $2000 and has 24 GB of VRAM could still sell..... specially if it has over 1600 TFLOPS at FP8 / 3000 peak INT TOPS

There is no 4080 super at msrp on newegg outside the PNY-Zotac brands I think

Everything at msrp is sold out at Best Buy:
https://www.bestbuy.com/site/searchpage.jsp?id=pcat17071&qp=gpusv_facet=Graphics Processing Unit (GPU)~NVIDIA GeForce RTX 4080 SUPER&sp=+currentprice skuidsaas&st=4080+super

I'd be shocked if it were more than 16 gigs. It'll sell out regardless, if it doesn't Nvidia is more than happy to sit on them until they sell. I don't expect them to lower prices until demand drops in the AI sector.

 

[Gideon]

A $1,199 5080 that goes +15% the 4090 if the 5090 is $2000 and has 24 GB of VRAM could still sell..... specially if it has over 1600 TFLOPS at FP8 / 3000 peak INT TOPS

There is no 4080 super at msrp on newegg outside the PNY-Zotac brands I think

Everything at msrp is sold out at Best Buy:
https://www.bestbuy.com/site/searchpage.jsp?id=pcat17071&qp=gpusv_facet=Graphics Processing Unit (GPU)~NVIDIA GeForce RTX 4080 SUPER&sp=+currentprice skuidsaas&st=4080+super

The super 4080's definitely sold, buy the regular 4080's never really sold well. But I just think the upcoming 5000 series will likely lack the performance to really push anyone to upgrade except the few that have to have the best, especially at 1,000 plus for it. Amazon shows stock for the 4080 supers from different brands. Will see though, were all just speculating based on the little information out there.

 

[GoodBoy]

You do know that the 980 Ti and 780 Ti are the same node process (28nm) yet the 980 Ti delivered pretty sizeable gains right?

From the GTC 2024 keynote (at ~57 minutes), he shows the Inference performance increase of Blackwell over Hooper/Ada, it's 30x faster:

I believe the Blackwell is the 2 GPU's combined into 1 unit. Plus he attributes a portion of this increase due to the crazy fast interconnectivity.

So if we also suppose that the 30X increase is also a best case scenario or cherry picked workload, lets say a more realistic increase is 24X, to compare as a single die lets cut that in half to 12x. If it's getting a 50% boost from the interconnects, lets drop it to 8x. That's still huge. This is Hopper, which is the datacenter part that is parallel to Ada gaming GPU's.

No way to know how this will translate to gaming performance improvements. It is the same TSMC 4N process. Hell if the performance of the gaming GPU is double the 4xxx, that will be nuts.

Historic performance increases over the old gen's, and the process nodes they were built on:
780Ti to 980Ti 41% - As pointed out both of these were the same 28nm process node. Pretty damn impressive.
980Ti to 1080Ti 75% 16nm
1080Ti to 2080Ti 41% 12nm
2080Ti to 3090Ti 56% 8nm Samsung process derived from their 10nm process
3090 to 4090 was 35% TSMC 4N
3090Ti to 4090 was 27% TSMC 4N

So here's to being optimistic that the increase will exceed the historic increases.

 

[LukeTbk]

it's 30x faster:

This is comparing FP4 operation per second with FP8 for the other gen, I think ( the other chips like the H200 are almost multi GPU) and vast part of the boost is from the complete ecosystem like you say.

It is not a big indicative value because its comparing different operations and obviously specialized into doing 4 bits hardware will run them much faster than 8bits operation.


https://www.semianalysis.com/p/nvidia-blackwell-perf-tco-analysis

32-16-8 bits is more
H100->B100: +77% is more representative than +3000%, now is it lower than if they did not put a lot of hardware on 4 bits... could not tell you

 

[GoodBoy]

I figured it was cherry picked in some way, all presentations and PR material you gotta expect that.

77% is much more believable. Still an amazing uplift as well. Sounds like it could match the 9xx to 10xx generational performance increases.

 

[LukeTbk]

77% is much more believable. Still an amazing uplift as well.

H100 was 2.5-3 time the A100 on those metric for a recent reference, will have to see but if the B100 is a bigger chips and using more power... it shows that the stagnant node could mean a low raw upgrade and for Blackwell to be a very short lived stop gap and for the next gen to be in 2025 has rumoured.

 

[GoodBoy]

We can hope that by the 6xxx generation, that the will be on a 3nm process.

 

[Nebell]

I might have to sell my kidney for a 5090 laptop.

 

[Armenius]

Umm, no it's not. Link for graph https://fourweekmba.com/nvidia-revenue-by-segment/

View attachment 649621

I said 77% of their graphics business, not overall business. That accounts for $10 billion of that $13.5 billion in your chart.

 

[Gideon]

I said 77% of their graphics business, not overall business. That accounts for $10 billion of that $13.5 billion in your chart.

That is a ton smaller then the 47.5 billion they made on other products, like AI chips. What would you do when you are facing a capacity strained TSMC and need to pick a priority on what to produce?

 

[Binar]

Blackwell will be faster (duh) but top card 5090 will have less Nvram..with the nvram shortage and the premium pricing it can get for professional cards, gaming cards will have less Nvram. The only ones looking for lots nvram on consumer cards are ML/AI part-time reasearchers or small startups that try to cut cost by not buying professional lvl similar cards.
Remember the whole crypto debacle , where the newer the card the less efficient it was in mining

 

[LukeTbk]

That is a ton smaller then the 47.5 billion they made on other products, like AI chips. What would you do when you are facing a capacity strained TSMC and need to pick a priority on what to produce?

There 2 main option, they can do like they did with Ampere Samsung 8nm for the gaming product, TSMC 7NM for the commercial product.

They can do like what all the rumors seem to point out, launch the AI chips before the gaming product having a 100% allocation for the AI chips for some months.

A mix of both could become common, has I imagine there is not just purely the TSMC part, but the substrate and other element of the supply chain, some high end chips on a different node than the rest, the H100 type having the very best and the L40 and the gaming affair for the missed one on a different node.

The rumours for 2025 would be an extreme example of that, the rumors are that Blackwell replacement would be as early as 2025 and on NVIDIA edition of TSMC 3N (or better), so the AI cards could have more of a full year all alone on the latest node before the 2026 gaming card new gen.