RTX 5xxx / RX 8xxx speculation

They do have to get that out this year though to have any benefit probably a year before Blackwell.
Strong rumors that Blackwell has a Q4 2024 launch, could be top end only too and irrelevant to that RDNA 4 launch.

I too my feeling would be for $600 for a inbetween 7900xt-xtx for the very top end in they reach that performance and is launching soon, based on nothing, just a feeling that it would be low enough to sales well.
 
Strong rumors that Blackwell has a Q4 2024 launch, could be top end only too and irrelevant to that RDNA 4 launch.

I too my feeling would be for $600 for a inbetween 7900xt-xtx for the very top end in they reach that performance and is launching soon, based on nothing, just a feeling that it would be low enough to sales well.
Well latest rumors have been nvidia is going to prioritize Ai this year. Too much money. So wouldn’t be surprised if gaming Blackwell is mid 2025 in summer.
 
Well latest rumors have been nvidia is going to prioritize Ai this year. Too much money. So wouldn’t be surprised if gaming Blackwell is mid 2025 in summer.
NV will probably release 5090 this fall. It's a low volume/high margin product. Halo has always been fairly limited release. The rest of the stack is what is in question.
 
Well latest rumors have been nvidia is going to prioritize Ai this year. Too much money. So wouldn’t be surprised if gaming Blackwell is mid 2025 in summer.
Yes rumors are that Hopper replacement(B100-B40) release is planned to be first and AI can be vastly prioritized in volume, you can still have a small 200k unit of 5080-5090 that sell out the first 24 hours ready in late Q4 to steal all the thunder and mindshare and have the volume card (5060-5070) launch only in 2025 if they feel it is a good idea. "small" with how price goes it could be like 150 millions+ in sales

All just rumors for now I think, will see in 3 weeks if anything is announced during GTC. You could be right the H200 has yet to hit the ground and logically B100 could launch months after, leaving to time in 2025 for the 5090
 
Last edited:
you can still have a small 200k unit of 5080-5090 that sell out the first 24 hours ready in late Q4 to steal all the thunder and mindshare

I'm betting 30K at launch, maybe 50K total, plus a refresh that still won't bring them to 100K overall. I'm calling it, it will be slightly more than a paper launch with the rest going to compute.
 
I am thinking now that RDNA 4 may not launch until FSR with AI is ready.

Also if Sony implements the same in PS5 pro then that could take priority over a RDNA 4 launch


Interesting


Current-gen RDNA 3 has dedicated AI acceleration with Wave MMA (matrix multiply-accumulate) instructions, which can help improve AI-based performance and also benefits ray tracing instructions, similar to Nvidia's Tensor cores.


So AI supported FSR only for RDNA 3 & above 🤔

https://forums.tomshardware.com/threads/amd-confirms-it-is-working-on-an-ai-upscaler-for-gaming-–-cto-papermaster-says-its-part-of-ai-enabling-our-entire-portfolio.3838488/post-23215981
 
Recent rumours

5080 to be atleast as powerful as 4090 in Raster & more powerful in RT

5090? to be definitely released this year

Both 5090? & 5080 to have GDDR7
 
Recent rumours

5080 to be atleast as powerful as 4090 in Raster & more powerful in RT

5090? to be definitely released this year

Both 5090? & 5080 to have GDDR7

I want to know if this is with DLSS stuff on or not since everything seems to be geared towards that anymore over just plain old forced rendering.
 
I want to know if this is with DLSS stuff on or not since everything seems to be geared towards that anymore over just plain old forced rendering.
For a 5080 to be as powerful as a 4090 in raster would only need a extremely disappointing +20-25% improvement over the 4080, saying the 5080 will be at least as powerful as a 4090 in pure raster without DLSS on, it is virtually saying nothing (when they last time the xx80 upgrade gen-gen was not much better than that ?).

Even the 2080 was quite a bit better than the 1080Ti.
 
For a 5080 to be as powerful as a 4090 in raster would only need a extremely disappointing +20-25% improvement over the 4080, saying the 5080 will be at least as powerful as a 4090 in pure raster without DLSS on, it is virtually saying nothing (when they last time the xx80 upgrade gen-gen was not much better than that ?).

Even the 2080 was quite a bit better than the 1080Ti.

Yeah sure...if you count 10% as being quite a bit better 😂
 

Attachments

  • IMG_1006.png
    IMG_1006.png
    379.8 KB · Views: 2
Navi 48/44 speculation:



Navi 33Navi 44Navi 32Navi 48
productionTSMC N6TSMC N4PTSMC N5 (GCD) + N6 (MCDs)TSMC N4P
chip area204mm²~130mm²348mm²~240mm²
architectureRDNA3RDNA4RDNA3RDNA4
Raster Engines (SE)2234
Shader cluster (CU)32326064
Chip clock (real)2.7GHz (7600XT)?2.4GHz (7800XT)adj. 3.0GHz
FP32 computing power22 TFlops?37 TFlopsadj. 50 TFlops
Infinity Cache32MB32MB64MB64MB
Storage interface128bit GDDR6128bit GDDR6256bit GDDR6256bit GDDR6
PCI ExpressPCIe 4.0 x8during PCIe 4.0 x8PCIe 4.0 x16PCIe 4.0 x16


Navi 48Navi 44probable explanation
3216Number of WGPs (number of CUs = 2x WGP)
6432Infinity Cache in MB (alternative explanation: number of CUs)
256128Width of the memory interface in bits
693288nominal bandwidth in GB/sec (results in storage with 21.7 or 18 Gbps)
2770515effective bandwidth in GB/sec (including the effect of the infinity cache)
~240mm²~130mm²Chip area under TSMC's N4P manufacturing


https://m-3dcenter-org.translate.go...vi-4?_x_tr_sl=auto&_x_tr_tl=en&_x_tr_hl=en-GB


There are also historical role models directly from AMD: The Polaris generation entered the race in 2017 with only 232mm² chip area for Polaris 10 (Radeon RX 470/480/570/580) and 123mm² for Polaris 11 (Radeon RX 460/560). . At that time there was already a 256-bit memory interface in the small chip area of 232mm², but at that time the hardware difference between Polaris 10 & 11 with regard to the FP32 units was even slightly larger (2304 vs 1024) than just twice. At the time, the Polaris series was quite popular as powerful accelerators with good amounts of memory at, above all, reasonable prices. It's quite possible that AMD will try to repeat this approach with the RDNA4 generation (of course out of necessity, since the faster RDNA4 chips have been canceled) .
 
that very exact doubling of everything from the 5080 to 5090 (if true, really big if) good way to have rumors of two die connected a la B200 going on
 
https://videocardz.com/newz/nvidia-...080-to-launch-in-fourth-quarter#disqus_thread

Something to look forward to I'm about a week behind on my Tech news.
that very exact doubling of everything from the 5080 to 5090 (if true, really big if) good way to have rumors of two die connected a la B200 going on
I'll be happy if the price isn't completely insane and the 5080 is close to half of the price of the 5090. My top of the head guesss is going to be a 16 gb 5080 for 1499 that beats the 4090 in gaming by a decent margin, and a 32 gb 5090 at 2499. I love my 7900 XTX but i'd move on if the 5080 is 999 like the 4080 super. But my guess is they intend to have these cards on the top of the stack along with the the 40 super cards.
 
7900 gre now down to $510

If navi 48 raster = 7900 xt (which seems likely based on IPC & freq increase) then its price should be between $550 & $600 ??

AMD Radeon RX 7000 GPU Deals:
7900 XTX For $799,
7900 GRE For $510,
7800 XT For $457,
7700 XT For $351,
7600 XT For $299



Newegg has new ZIP discounts live across several GPUs, most of which are part of the AMD Radeon RX 7000 GPUs. These GPUs include all the way from the entry-level Radeon RX 7600 XT to the top of the spectrum, the Radeon RX 7900 XTX. The ZIP code "ZIPGAME" can be applied to select Radeon RX 7000 GPUs with discounts of up to $100 (max @ 12%). This drops the prices on several GPUs below their MSRP.

https://wccftech.com/amd-radeon-rx-...-gre-510-7800-xt-457-7700-xt-351-7600-xt-299/

  • MSI RX 7900 XTX 24 GB Gaming Trio Classic - $899 US / $799.99 US (w/code ZIPGAME)
  • ASRock RX 7900 GRE 16 GB Steel Legend OC - $579 US / $510.39 US (w/code ZIPGAME)
  • Sapphire RX 7800 XT Nitro+ OC Triple-Fan - $539 US / $475.19 US (w/code ZIPGAME)
  • ASRock RX 7800 XT 16 GB Steel Legend OC - $519 US / $457.59 US (w/code ZIPGAME)
  • XFX RX 7800 XT 16 GB Speedster MERC319 - $519 US / $457.59 US (w/code ZIPGAME)
  • Sapphire RX 7700 XT 12 GB PULSE Dual-Fan - $399 US / $351.99 US (w/code ZIPGAME)
  • ACER RX 7600 XT Nitro OC - $339.99 US / $299.19 US (w/code ZIPGAME)
 
Yeah sure...if you count 10% as being quite a bit better 😂
I'll grant you Luke said 2080>1080ti.. probably misremembering.
The graph shows 27% difference for 1080->2080 and 29% difference for 2080->2080ti.
For a few generations I think this was all we got, about 30% uplift.

Since Blackwell is still on the (as far as we know) same TSMC 4N node as 4xxx uses, 30 to 60% is probably all we will get going from 4080 to 5080, or 4090 to 5090, and likely trend towards that lower % guess. The faster Ram will help some but not another 30%. So my prediction would be 35 to 45%.
60% if we get it will be great, but if its only 30%, then demand will be tempered. If it's 100% watch out, that will sell out the first day.

Someone find a graph with all of the generation to generation uplifts we got in the previous Nvidia releases.
 
Last edited:
I'll grant you Luke said 2080>1080ti.. probably misremembering.
The graph shows 27% difference for 1080->2080 and 29% difference for 2080->2080ti.
For a few generations I think this was all we got, about 30% uplift.

Since Blackwell is still on the (as far as we know) same TSMC 4N node as 4xxx uses, 30 to 60% is probably all we will get going from 4080 to 5080, or 4090 to 5090, and likely trend towards that lower % guess. The faster Ram will help some but not another 30%. So my prediction would be 35 to 45%.
60% if we get it will be great, but if its only 30%, then demand will be tempered. If it's 100% watch out, that will sell out the first day.

Someone find a graph with all of the generation to generation uplifts we got in the previous Nvidia releases.

1080 Ti to 2080 Ti was the only time we got a 30% uplift. Going from a 780 Ti to 980 Ti was more than 30% and so was 980 Ti to 1080 Ti, 2080 Ti to 3090, and 3090 to 4090. I am comparing the top dog dies for each generation so GK110 vs GM200 vs GP102 vs TU102 vs GA102 vs AD102. Out of those top end dies only going from GP102 to TU102 was the only time the uplift was 30%.

1713809928217.png
1713809964056.png
1713810221916.png
 

Attachments

  • 1713810120652.png
    1713810120652.png
    169.7 KB · Views: 0
Last edited:
All tests were 2560x1440, without DLSS, RT, or frame gen.
Looks like 780Ti to 980Ti was 29%
980Ti to 10080Ti was 43%
1080Ti to 2080Ti was 29%
2080Ti to 3090Ti was 36% ? I could swear that one was higher. Maybe it was the DLSS improvements.
3090 to 4090 was 26%
3090Ti to 4090 was 21%

I believe all of these were accompanied by process node improvements. So with 5090 being on the same node, the only increases will be in the processor design, and the vRam speed.
It might turn out to be the smallest increase yet as far as the above charts/testing method is concerned. DLSS keeps getting better and frame gen already works. efficiency improvements to those can be expected, just not sure it will mean double digit performance jumps. Guess we will see.

Edit: The above is relative performance new gen to old. Performance increase over the old gen is:
780Ti to 980Ti 41%
980Ti to 10080Ti 75%
1080Ti to 2080Ti 41%
2080Ti to 3090Ti 56%
3090 to 4090 was 35%
3090Ti to 4090 was 27%
 
Last edited:
2080Ti to 3090Ti was 36% ? I could swear that one was higher. Maybe it was the DLSS improvements.
3090 to 4090 was 26%
3090Ti to 4090 was 21%
USed to look at 4k ? Even at 1440p on TPU it is higher than those
relative-performance_2560-1440.png


+41% for the 2080TI->3090TI

the 4090 was +35%:
relative-performance_2560-1440.png


(64% at 4k when it can use its "force")
 
Last edited:
I was just subtracting the percentages. You are right the perf gain over the old gen is a bit higher. I'll edit the post
 
Yes the performance uplifts ever since Kepler have all been well over 30% when comparing top die vs top die with again the jump from 1080 Ti to 2080 Ti being the exception when tested at 1440p, at 4K the uplift is just over 40% going from a 1080 Ti to 2080 Ti. So going from the 4090 to 5090 I am once again expecting the performance gains to be at minimum 40% and would not be surprised to see 50% or higher. "Double" performance gain is extremely unlikely except for the heaviest of ray traced/path traced workloads.
 
Yes the performance uplifts ever since Kepler have all been well over 30% when comparing top die vs top die with again the jump from 1080 Ti to 2080 Ti being the exception when tested at 1440p, at 4K the uplift is just over 40% going from a 1080 Ti to 2080 Ti. So going from the 4090 to 5090 I am once again expecting the performance gains to be at minimum 40% and would not be surprised to see 50% or higher. "Double" performance gain is extremely unlikely except for the heaviest of ray traced/path traced workloads.

I find that unlikely unless they do more then a small node shrink. You just are not getting the same gains out of the shrinking nodes. I would be surprised if they eek out much more then a 25% uplift, unless they are doing more then just relying on a node shrink for the performance. Ray tracing is a whole different ball game and should be measured separately.
 
I find that unlikely unless they do more then a small node shrink. You just are not getting the same gains out of the shrinking nodes. I would be surprised if they eek out much more then a 25% uplift, unless they are doing more then just relying on a node shrink for the performance. Ray tracing is a whole different ball game and should be measured separately.

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?
 
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?

You do know they didn't rely on a node shrink for that. That took a significant core design change.
 
Bingo. So what makes you think the same thing cannot happen for Blackwell?

Seen no rumors that suggest that is happening. Could it, sure. But with no info stating major core changes then, I don't expect much from the next chip. If more info come to light then my thoughts may change.
 
Seen no rumors that suggest that is happening. Could it, sure. But with no info stating major core changes then, I don't expect much from the next chip. If more info come to light then my thoughts may change.

Fair enough. But in my opinion Nvidia has to deliver on next gen, otherwise they can't justify such high prices. Can't speak for others but I would not be willing to pay $1600+ to upgrade from my 4090 if all I got was a measely 25% uplift. I happily upgraded from a 3080 Ti to a 4090 because the uplift was huge, so if they can't deliver then I would imagine many people are going to pass on upgrading and that's going to hurt their sales numbers which they don't wanna see. Nvidia also has way more $$$ for RnD now than ever before to make it happen.
 
Fair enough. But in my opinion Nvidia has to deliver on next gen, otherwise they can't justify such high prices. Can't speak for others but I would not be willing to pay $1600+ to upgrade from my 4090 if all I got was a measely 25% uplift. I happily upgraded from a 3080 Ti to a 4090 because the uplift was huge, so if they can't deliver then I would imagine many people are going to pass on upgrading and that's going to hurt their sales numbers which they don't wanna see. Nvidia also has way more $$$ for RnD now than ever before to make it happen.

I am a little less convinced they care about the consumer market as much right now, considering what their AI chips sell for at the moment. Will they give up on it, no. But I feel their focus is elsewhere and I think generational performance increases will start to suffer as they focus elsewhere in the business. With the state of the economy I have a feeling video card prices are going to have go down, 1000+ for a card is just not sustainable.
 
I am a little less convinced they care about the consumer market as much right now, considering what their AI chips sell for at the moment. Will they give up on it, no. But I feel their focus is elsewhere and I think generational performance increases will start to suffer as they focus elsewhere in the business. With the state of the economy I have a feeling video card prices are going to have go down, 1000+ for a card is just not sustainable.
Gaming makes up 77% of NVIDIA's graphics business. You don't throw away a $10 billion market segment, especially when you hold close to 80% of the total consumer market for discrete video cards. Their market share will slip if they start to slack.
 
I am a little less convinced they care about the consumer market as much right now, considering what their AI chips sell for at the moment. Will they give up on it, no. But I feel their focus is elsewhere and I think generational performance increases will start to suffer as they focus elsewhere in the business. With the state of the economy I have a feeling video card prices are going to have go down, 1000+ for a card is just not sustainable.

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.
 
Umm, no it's not
How does the graph tell which percentage gaming is of that 13 billion graphic segment ;) ?

But I feel their focus is elsewhere and I think generational performance increases will start to suffer as they focus elsewhere in the business.
Outside the 5090, it is a lot just a business choice how much the performance increase the 5070 will have over the 4070, we could have one the biggest jump ever with very little technical improvement, just go back to a 400mm die like the 3070 reusing the current Lovelace die on the revised TSMC node, GDDR7 and you are pretty much already there.
 
Back
Top