CES 2025 AMD & Nvidia Keynote Live Chat [06/01/25]

[TNA]

It may be "dominated" by the numbers in the table but many of these numbers have little to no impact on performance.

Memory clock, bit rate and power draw don't matter, what does matter is bandwidth, and yes it's hugely higher but I'm not convinced it's needed unless you are a 8k gamer. Memory capacity also doesn't matter

Then for core count, it's a 30% increase and a 5% decrease in clocks, for a net 25% improvement. As we saw with the 4090 each extra 1 core you add on Nvidia cards doesn't add 1 unit more performance, so this net 25% won't be 25% more gaming rasterisation

Weren't you crying not long ago wanting more than 24gb on next gen cards?

Guy changes his mind like the wind. Almost as bad as mrk

 

[AMDPower]

this is a very intelligent design, can repair damage to pci connector very easily now without removing and reballing the gpu and the VRAM's.

 

[humbug]

I don’t think anybody knows for sure, it’s just extrapolation from the spec sheet:

GPU	RTX 4090	RTX 5090
CUDA Cores	16,384	21,760
Core Frequency	2.23 GHz	2.01 GHz
Boost Core Frequency	2.52 GHZ	2.41 GHZ
Memory Capacity / Type	24 GB GDDR6X	32 GB GDDR7
Memory Bandwidth	1008 GB/s	1792 GB/s
Memory Interface	384-bit	512-bit
Memory Clock	21 Gbps	28 Gbps
TGP	450W	575 W

Ignoring all the DLSS stuff the 4090 is still absolutely dominated by the 5090, other than the 5% drop in boost clocks. I’d be very surprised if it’s only a 20% gain, but you never know!

By the time we’re seeing above 24g vram usage as standard and the 5090 pulls ahead significantly, it’ll probably be 4 years time and these cards may both be outdated compared to the new top end.

33% more cores and 70% higher memory bandwidth, i can see it being 30 to 50% faster depending on the game.

 

[jonneymendoza]

I don’t think anybody knows for sure, it’s just extrapolation from the spec sheet:

GPU	RTX 4090	RTX 5090
CUDA Cores	16,384	21,760
Core Frequency	2.23 GHz	2.01 GHz
Boost Core Frequency	2.52 GHZ	2.41 GHZ
Memory Capacity / Type	24 GB GDDR6X	32 GB GDDR7
Memory Bandwidth	1008 GB/s	1792 GB/s
Memory Interface	384-bit	512-bit
Memory Clock	21 Gbps	28 Gbps
TGP	450W	575 W

Ignoring all the DLSS stuff the 4090 is still absolutely dominated by the 5090, other than the 5% drop in boost clocks. I’d be very surprised if it’s only a 20% gain, but you never know!

By the time we’re seeing above 24g vram usage as standard and the 5090 pulls ahead significantly, it’ll probably be 4 years time and these cards may both be outdated compared to the new top end.

How much was the restar increase from 30 series to 40?

 

[Sedgey123]

Can't really compare the 3090 to 4090 in terms of the differences to 4090 to 5090 as they went from an 8nm node to a 4Nm node

 

[Nitefly]

How much was the restar increase from 30 series to 40?

Found this table on Reddit re: specs:

Card	3090	4090	5090
Boost Clock	1695MHz	2520MHz	2407MHZ
FP16 Cores	10496	16384	21760
FP16 TFLOPS	142.33	330.4	419.01
Memory Clock	2437.5MHz	2625MHz	4375MHz
Bus Width	384-bit	384-bit	512-bit
Memory Bandwidth 936GB/s	936GB/s	1008GB/s	1792GB/s
TDP	350W	450W	575W

I think the raster performance of 4090 over 3090 was huge, generally 60-90% increase native at 4k.

I don’t think anyone is expecting that sort of uplift this time, estimates at the 5090 increase seem to start from 20% 4k native… my wholly uninformed guess is around 30% but nobody really knows.

 

[mrk]

Finally proper engineering, that is very impressive tbh.

 

[jonneymendoza]

Found this table on Reddit re: specs:

Card	3090	4090	5090
Boost Clock	1695MHz	2520MHz	2407MHZ
FP16 Cores	10496	16384	21760
FP16 TFLOPS	142.33	330.4	419.01
Memory Clock	2437.5MHz	2625MHz	4375MHz
Bus Width	384-bit	384-bit	512-bit
Memory Bandwidth 936GB/s	936GB/s	1008GB/s	1792GB/s
TDP	350W	450W	575W

I think the raster performance of 4090 over 3090 was huge, generally 60-90% increase native at 4k.

I don’t think anyone is expecting that sort of uplift this time, estimates at the 5090 increase seem to start from 20% 4k native… my wholly uninformed guess is around 30% but nobody really knows.

Well I'm coming from a 3090 so I should expect a 90 to 120 percent increase? Aka double the performance I'm getting for a 3090?

 

[Nitefly]

Well I'm coming from a 3090 so I should expect a 90 to 120 percent increase? Aka double the performance I'm getting for a 3090?

I’m on a 3090 myself and that’s around what I’m hoping for tbh. But again, nobody will really know until the proper reviews and benchmarks come out in a couple of weeks.

 

[jonneymendoza]

I’m on a 3090 myself and that’s around what I’m hoping for tbh. But again, nobody will really know until the proper reviews and benchmarks come out in a couple of weeks.

Well the fact that 4090 is almost double the rasta performance from our old cards, the 5090 should at least match that?

 

[Nitefly]

Well the fact that 4090 is almost double the rasta performance from our old cards, the 5090 should at least match that?

Yes, it would be a bit absurd and wholly unexpected for the 4090 to ‘beat’ the 5090!

 

[t2na]

Yes, it would be a bit absurd and wholly unexpected for the 4090 to ‘beat’ the 5090!

Would be pretty funny though.

 

[Purgatory]

 

[Curtis vincent]

Think I saw a 5090 from Inno3D that came preinstalled with an alphacool block

 

[Boomstick777]

Found this table on Reddit re: specs:

Card	3090	4090	5090
Boost Clock	1695MHz	2520MHz	2407MHZ
FP16 Cores	10496	16384	21760
FP16 TFLOPS	142.33	330.4	419.01
Memory Clock	2437.5MHz	2625MHz	4375MHz
Bus Width	384-bit	384-bit	512-bit
Memory Bandwidth 936GB/s	936GB/s	1008GB/s	1792GB/s
TDP	350W	450W	575W

I think the raster performance of 4090 over 3090 was huge, generally 60-90% increase native at 4k.

I don’t think anyone is expecting that sort of uplift this time, estimates at the 5090 increase seem to start from 20% 4k native… my wholly uninformed guess is around 30% but nobody really knows.

Wow the 5090 really is a monster! The cores and memory bandwidth are impressive.

 

[stooeh]

Hopefully it is actually 30%+ raster increase as it kinda looks like an overclocked 4090

 

[KompuKare]

Found this table on Reddit re: specs:

Card	3090	4090	5090
Boost Clock	1695MHz	2520MHz	2407MHZ
FP16 Cores	10496	16384	21760
FP16 TFLOPS	142.33	330.4	419.01
Memory Clock	2437.5MHz	2625MHz	4375MHz
Bus Width	384-bit	384-bit	512-bit
Memory Bandwidth 936GB/s	936GB/s	1008GB/s	1792GB/s
TDP	350W	450W	575W

I think the raster performance of 4090 over 3090 was huge, generally 60-90% increase native at 4k.

I don’t think anyone is expecting that sort of uplift this time, estimates at the 5090 increase seem to start from 20% 4k native… my wholly uninformed guess is around 30% but nobody really knows.

350W >> 450W >> 575W. That's pretty unsustainable.

If GPU vendors don't do something, then I really expect governments to eventually do something. Is also expect most computer forums to be very vocal about any energy regulations.

Cooler and board design really has advanced to cope with these monsters but I'd rather see innovation elsewhere.

 

[Purgatory]

350W >> 450W >> 575W. That's pretty unsustainable.

If GPU vendors don't do something, then I really expect governments to eventually do something. Is also expect most computer forums to be very vocal about any energy regulations.

Cooler and board design really has advanced to cope with these monsters but I'd rather see innovation elsewhere.

If we want to go RTX top cards power increase we really need to start at 2080ti.

2080Ti is 250W

3090 is 350W (+100W more than previous generation)

3090Ti is 450W (+100W more than previous card same generation)

4090 is 450W (+100W more than previous generation 90 none 90Ti )

5090 is 575W (+125W more than previous generation 90 none 90Ti)

so 6090 600W? unless they do a new spec connector again..


Power has got silly and they need to get back down to 350W max for the 90 class and 250W for 80 class.

 

[Curtis vincent]

Perhaps slightly off topic although kinda related. The lossless scaling app has a version 3 out today that reportedly has much improved frame generation. Also now includes a custom setting for frame gen multiplier. Not expecting results to be as good as hardware frame gen but interested to see the results.

 

[Grim5]

350W >> 450W >> 575W. That's pretty unsustainable.

If GPU vendors don't do something, then I really expect governments to eventually do something. Is also expect most computer forums to be very vocal about any energy regulations.

Cooler and board design really has advanced to cope with these monsters but I'd rather see innovation elsewhere.

High performance GPUs and PCs are excluded from power restrictions in the US states that have them


It's not just GPUs, power requirements for running data centres is exploding as well