Discussion Nvidia Blackwell in Q1-2025

[maddie]

If this is what future is going to look like from now on, not speaking strictly about Nvidia and GPUs, but broadly about industries/society, it pretty much means technological progress available only to "rich" people going forward.

For all the margin chasing crowd, they should know they would not exist without the mass market to drive volumes, making those very technologies possible. Arrogant people are dumb about their successes and why it happens, thinking only of self.

 

[Gideon]

I'm not sure if this was posted here:

https://twitter.com/x/status/1879563653298655355

Cyberpunk 5090 mfg 4x dlss quality RT+PT 201 "fps"72ms latency ?

and in the comments

72 ms of PC latency at 200 FPS would feel horrible and jarring. Hell I was forced to disable path tracing in games such as CP 2077, AW2 & BMW because even with DLSS 3 FG on my 4090 they'd drop to ~120 FPS with about 40-50 ms of PC latency, which felt quite input laggy to me.

IMO 30-40ns is fine, particularily in a 3rd-person game like Alan Wake 2, but yeah 50+ get's really bad

 

[Sunaiac]

72ms == 14FPS gameplay

 

[MrTeal]

Zen is way different. Zen does not use a huge complex monolith for their consumer line and is cheap to build. AMD just been delivering at a consistence cadence for their CPU division while keeping power consumption relatively low which is what help it gain consumer confidence. It was not the threadrippers that did it.

Going Halo for Nvidia makes more sense since they have a professional visualization market which provides guaranteed income at high margins.

AMD doesn't have much for this market. As we can see with the possible performance flop of the RTX 5090, going big carries more risk as more things can go wrong. The GTX 480 and Fury X are examples of this. The RTX 5090 is clocked awfully low for the amount of power it uses. This might be a fermi like chip which misses targets in the gaming segments but recovers sales in the professional markets.

I don't disagree on GPUs, but I would disagree about halo in the GPU space not making a big difference. It's just that consumer sockets have gotten to the point that HEDT is effectively dead, so Threadripper really isn't halo anymore for gamers. Back with Sandy-E and Haswell-E you'd get to move past 4 cores with HEDT, get enough PCIe lanes to do crazy SLI systems, and wouldn't have to give up gaming performance to get it.

AMD definitely has the halo now with the X3D chips though. Even with Zen3, they were right there at the top basically neck and neck with Intel an offering great value to gamers. Even then just from experience building/selling gaming PCs and talking to average customers, the large majority would prefer a 10600k to a 5600X even at a higher price point. Intel's stumbles and the excellence of the X3D chips has gotten enough mindshare now that people are starting to shift to "I don't know much about PCs, but I know I want an AMD CPU and Nvidia GPU." That's a big swing I don't think would have happened without the X3Ds as the gaming halo.

Now, all it took was AMD beating Intel in gaming performance in most cases by double digit percentages while consuming 1/3rd the power and not destroying themselves in a year. The chance of AMD doing that to Nvidia at the high end is effectively 0.

 

[GTracing]

72ms == 14FPS gameplay

No, there are other factors adding latency. If you were lock the framerate at 14fps (without frame-gen) and measure the latency as they do in the tweet, it would be higher than 72ms.

 

[coercitiv]

I'm not sure if this was posted here:

https://twitter.com/x/status/1879563653298655355

It's complicated, the screenshot represents a latency spike from the average that was under 50ms while moving the mouse. The author of the video also mentions that specific CP2077 build had some isolated stutter issues, that they later "fixed" by running the game in borderless window mode.

This is why I don't like these controlled "previews", they can only be used to evaluate the best case scenarios and they come with multiple asterisks for the worst case problems.

 

[Heartbreaker]

No, there are other factors adding latency. If you were lock the framerate at 14fps (without frame-gen) and measure the latency as they do in the tweet, it would be higher than 72ms.

Yeah, this is not 14 FPS.

But at 72ms (maybe a spike), they are definitely running at some very low frame rate, definitely below 60 FPS.

Lets say this was ~50 FPS = 20 ms of frame latency, and say another 10ms from the game engine , so you are at 30ms, then you engage MFG, and that adds another 20 ms (this is the big problem using FG to escape low frame rates), and now you are 50ms, and then you get a slight spike...

When you see reasonable reviewers talk about turning on FG, they suggest you need over 60 FPS real frame rate to consider using it. 80-100 FPS is the suggestion I see most often.

So FG is essentially useless when you need it most (low frame rates).

FG is mostly just window dressing on already good frame rates. This makes it all the more misleading for NVidia to be selling it as a real frame rate.

 

[lixlax]

---
So FG is essentially useless when you need it most (low frame rates).
---

Also fast paced shooters.

 

[Saylick]

Welp, looks like no cache changes at the SM level for Blackwell either:

https://twitter.com/x/status/1879900220001263658

 

[Win2012R2]

Budget was reallocated to the RT core. Might even call it a level 4 RT core with the features.

Can't be - FP32/INT+FP32 (dual issue) is surely cheaper than FP32/INT+FP32/INT - if this is still dual issue like but supporting INT on all pipes, or just all FP32/INT as they imply but it can't be - IPC would have increased then considerably.

The oversll INT throughput is doubled but the capability to concurrently issue INT that Turing brought is seemingly lost.

This sucks, if true - Turin did improve normal perf with FP32/INT concurrency

 

[adroc_thurston]

Welp, looks like no cache changes at the SM level for Blackwell either:

SNOOZE
i'm gonna grill someone elsewhere

Can't be - FP32/INT+FP32 (dual issue) is surely cheaper than FP32/INT+FP32/INT - if this is still dual issue like but supporting INT on all pipes, or just all FP32/INT as they imply but it can't be - IPC would have increased then considerably.

GPUs don't quite work like that.

 

[Win2012R2]

GPUs don't quite work like that.

Whatever they did change don't seem to make it all proper full non-fake FP32/INT core, otherwise perf in old stuff would have gone up higher, maybe they just did not finish the job because it needed N3 budget

 

[adroc_thurston]

Whatever they did change don't seem to make it all proper full non-fake FP32/INT core

It seems like an inherently incremental SM optimization.
Looks like they expected higher fmax but it did not pan out.

 

[Win2012R2]

It seems like an inherently incremental SM optimization.

It did not seem to have panned out either - clocks about same and perf seem to be also, so more or less no IPC improvement

 

[adroc_thurston]

It did not seem to have panned out either - clocks about same and perf seem to be also, so more or less no IPC improvement

Yeah. It looks like SM streamlining aimed at higher clocks and higher raw PPA, but something went wrong. Mildly wrong.
Wasn't client blackwell rumored to be up to 2.9G fmax?

 

[SolidQ]

Switch 2 Techs Specs.
Tensor Core seems from ADA, but 2 RT cores? 🙄

 

[Win2012R2]

Wasn't client blackwell rumored to be up to 2.9G fmax?

Pretty sure they had to switch to N4 vs expected N3 for it, so perhaps that's who dunnit

 

[ToTTenTranz]

no this just smells like an fmax miss

So AMD decided not to compete during Nvidia's RDNA3 moment?

Wow, they're really the best at wasting opportunities.

 

[adroc_thurston]

So AMD decided not to compete during Nvidia's RDNA3 moment?

Yea.

Wow, they're really the best at wasting opportunities.

Turns out sometimes your infallible comp screws up and you roll them over. Too bad!

 

[Mopetar]

So AMD decided not to compete during Nvidia's RDNA3 moment?

Wow, they're really the best at wasting opportunities.

I'm sure AMD has the ability to know years in advance when Nvidia will flounder so that they can decide not to release a full range of products to capitalize.

It's more likely the other way around where Nvidia realized that AMD wasn't going to compete against their full range so they didn't need to try very hard.

In the past when ATI/NV were more competitive you still had rebadges and generations that were essentially product refreshes, but that usually meant lower prices to compensate. If AMD had more/better products the. NVidia would be charging a lot less for these cards. Look at their Ampere pricing when they had to go up against a far more competitive RDNA2 lineup.

Nvidia also didn't have a massive AI boom that makes their consumer GPU sales look like a rounding error on the balance sheet by comparison either, so they have even less reason to care if this generation is a sidegrade at best for most consumers. They'll just wait it out and cash in on their next generation when a 30% uplift seems impressive again.

 

[Heartbreaker]

So AMD decided not to compete during Nvidia's RDNA3 moment?

Wow, they're really the best at wasting opportunities.

More like the economics of GPU progress are unfavorable this generation and it's impacting both.

 

[adroc_thurston]

More like the economics of GPU progress are unfavorable this generation and it's impacting both.

Is it?

 

[adroc_thurston]

It's more likely the other way around where Nvidia realized that AMD wasn't going to compete against their full range so they didn't need to try very hard.

Wrong, GB202 is the chungusest chungus ever, the first 512b NV part since GT200 too.

 

[DisEnchantment]

Switch 2 Techs Specs.
Tensor Core seems from ADA, but 2 RT cores? 🙄

Looks surprisingly similar to Jetson Orin NX/AGX

 

[Heartbreaker]

Looks surprisingly similar to Jetson Orin NX/AGX

Why would similarity be surprising? It's kind of the expected performance envelope, and NVidia has SoCs in that range, so similarity would be expected.

Though note one claims to have A78AE cores and the other A78C. I'm not sure if that designation involves actual hw changes or software plus validation changes...