Discussion Nvidia Blackwell in Q1-2025

[CP5670]

I said it before here, but it's odd to have a AMD CPU that uses 50-70W in games on AIO and a 575W monster on an air cooler. It should be the other way around.

 

[CastleBravo]

More along the lines of at 100% CPU utilization, if a case or CPU fan fails (at least 2 of each type are installed), the CPU might run slightly hotter or even thermal throttle (I set the motherboard to throttle at 90°C in both of my signature's systems), but I'll get a warning and maybe lower performance until I can correct the hardware failure. In the case of my home fileserver, it has an extremely large heatsink with no fan on the CPU (5700X @65w) and the CPU can run at 100% with zero case airflow; at an elevated temperature of course.

Edit: I also use solid graphite based thermal pads for the CPU TIM, so cooling is always consistent and there are no worries about thermal pastes migrating, drying out, etc.

The only way you would get away with that in a standard case configuration using a 575W air cooled GPU would be a 7 slot blower fan design, and it would be loud. You are in the extreme minority if you want such a GPU. If you really want that level of air cooled redundancy with a 5090, I would do a deshroud and mount the GPU heatsink on case exhaust fans using a PCIE riser cable.

 

[Hitman928]

wait, the RTRT bump is less?

The RT tests were done at 1440p, often with upscaling so even lower render res. At actual 4K resolution, the RT improvement was in line with the non-RT results.

 

[adroc_thurston]

The RT tests were done at 1440p, often with upscaling so even lower render res. At actual 4K resolution, the RT improvement was in line with the non-RT results.

Ah.
Well that's also kinda ass. A whole uarch wasted on...
... nothing.

 

[cusideabelincoln]

View attachment 115482


This is bad design, I changed by mind, this is poor design.

I think the real problem is the up-to 600W of power the card is pumping into the case. AIB cooling solutions won't really solve that either since they still dump the majority of heat inside the case.

Further on you can see they mitigated the CPU temps by increasing all the fans speeds. You simply need to have a case with good airflow with this card, and their case only has 1 exhaust outlet which is pretty bad.

They would probably get similar results using an overclocked 3090 Ti card.

 

[CastleBravo]

I said it before here, but it's odd to have a AMD CPU that uses 50-70W in games on AIO and a 575W monster on an air cooler. It should be the other way around.

The thing I want to see is a cooler like the FE dual blow through, but designed to easily reverse the fan flow, and also an mATX MB with the bottom PCIE slot being the one wired for x16. Then you could have the GPU blow straight out the bottom of the case, or out the top if it is an inverted case.

 

[maddie]

View attachment 115482


This is bad design, I changed by mind, this is poor design.

Whoever wrote that has no clue about thermodynamics.

More power to dissipate is the reason. ALL cards except water-cooled and blower models transfer ALL their heat into the case. For a given closed case setup, the additional power consumed by the card (176W re TPU) and the additional power consumed by the CPU, caused by a higher performance GPU, will result in a significant temperature rise.

This has nothing to do with the cooler design. As total wattage increases, the case itself must be more ventilated. Water cooling of both CPU and GPU will fix part of the problem, with this (additional power consumed by the CPU, caused by a higher performance GPU) remaining to raise CPU temps.

 

[Heartbreaker]

So when do we get DLSS 4 for older RTX cards?

 

[Abwx]

A 2 x 6 pins power connector is not serious for currents ranging in the 40 - 50A zone,
that s 6 - 8A per pin with an average of 7A or so, 3 x 8 pins should be a minimum, this would reduce the current/pin by 50%.

 

[Win2012R2]

wait, the RTRT bump is less?

They really dropped the ball on this one

 

[SolidQ]

More RT % drops? in some games. There more games like Silent Hill 2 also

 

[adroc_thurston]

They really dropped the ball on this one

I was told by Very Green nvidia people that Blackwell is an epic mean new SM and if this is the end result, I'm worried how less massive NV efforts look like.

 

[CastleBravo]

View attachment 115482


This is bad design, I changed by mind, this is poor design.

Reading the computer base review, that test was done with NHD15 fans at only 800RPM, and the case fans at only 400RPM. When they bumped the case fans to 650RPM and the CPU to 1,000RPM (still quite low on both IMO), the 9800X3D maxed out at 84°C. With case fans at full speed (no reported change in CPU fan speed, so likely still 1k), CPU temp maxed out at 64°C.

 

[Tup3x]

Fermi 2.0 (or 3.0 depending on how you look).

 

[adroc_thurston]

Reading the computer base review, that test was done with NHD15 fans at only 800RPM, and the case fans at only 400RPM. When they bumped the case fans to 650RPM and the CPU to 1,000RPM (still quite low on both IMO), the 9800X3D maxed out at 84°C. With case fans at full speed (no reported change in CPU fan speed, so likely still 1k), CPU temp maxed out at 64°C.

drone music for free huh

 

[gdansk]

Fermi 2.0.

It's not that bad. It's like the 2080 Ti repeated but with an even smaller efficiency improvement.

 

[Golgatha]

The only way you would get away with that in a standard case configuration using a 575W air cooled GPU would be a 7 slot blower fan design, and it would be loud. You are in the extreme minority if you want such a GPU. If you really want that level of air cooled redundancy with a 5090, I would do a deshroud and mount the GPU heatsink on case exhaust fans using a PCIE riser cable.

I'm not worried about the fans on the video card failing. Typically they have 2-3 fans anyway, so you can at least use them for desktop and probably gaming at elevated temps. Worst case scenario is I use the integrated graphics on the CPU until I'm able to replace or RMA the GPU.

 

[Heartbreaker]

More RT % drops? in some games. There more games like Silent Hill 2 also

Definitely NOT a step forward for RT.

 

[adroc_thurston]

It's not that bad. It's like the 2080 Ti repeated but with an even smaller efficiency improvement.

Turing had an excuse of "we've built a modern SM". This has none.

 

[PJVol]

how are laws of physics related to logic or microarch design

They aren't directly connected to design, but the power of operating ICs is subject to the laws of physics, where dynamic power consumed during the charge/discharge of the xtor capacitance.

It's a simpler SM.

Simpler not necessarily means less switching capacitance, and based on specified boost clocks of gb202 my rough math still stands.

 

[adroc_thurston]

They aren't directly connected to design, but the power of operating ICs is subject to the laws of physics, where dynamic power consumed during the charge/discharge of the xtor capacitance.

Yea.

Simpler not necessarily means less switching capacitance

Yeah it does, especially when you gut the shader core pain points (SMSP schedulers and opforwarding networks).
Like they did away with gimmicky ALU subsets and went back to Maxwell.
Where is le fmax.

 

[Meteor Late]

-2080 ti was 39% higher performance at 4K on average compared to 1080 ti, so a bit better uplift compared to the 30% or so that the 5090 gets over the 4090, it's 27% on HUB.
-2080 ti had 27% more bandwidth than 1080 ti, whereas 5090 has almost 80% more bandwidth!
-Turing introduced Raytracing and tensor cores, so it is understandable that a lot of effort went there.

All in all, Turing was MUCH better than Blackwell as an architecture.

One can say that Turing increased the die size much more, but again, there is hardware dedicated to Raytracing and tensor cores that were not there in 1080 ti.

This was the Zen moment for AMD, but Lisa Su wasted this oportunity with her obsession with PPA and margins.

 

[gdansk]

All in all, Turing was MUCH better than Blackwell as an architecture.

It's 1.39x vs 1.34x if you're using TPU summaries.
Pretty much the same idea. It's only slightly worse. And Turing had a 70% price increase whereas this has 25% price increase.
So overall it is the closest NVidia release to Blackwell.

 

[Win2012R2]

I was told by Very Green nvidia people that Blackwell is an epic mean new SM and if this is the end result, I'm worried how less massive NV efforts look like.

Well, allegedly Blackwell was meant to be on N3, just like Zen 5, so then they had to port it to N4 and probably had to cut some stuff and freq regressed.

FP32/INT cores certainly don't seem to work as proper fully independent ones - Turin had good boost with that, and going away from crappy dual issue should have produced meaningful results even with same clock, but we are not seeing that.

And Turing had a 70% price increase whereas this has 25% price increase.

Turin was big increase on low base, and this has got "small" increase on already high base, this crock should have been -XX% across the board.

 

[Meteor Late]

It's 1.39x vs 1.34x if you're using TPU summaries.
Pretty much the same idea. It's only slightly worse. And Turing had a 70% price increase whereas this has 25% price increase.
So overall it is the closest NVidia release to Blackwell.

You are right, I though the difference was a bit bigger.
Still, I think it is a worse architecture for sure, given the massive difference in bandwidth.