Discussion Nvidia Blackwell in Q1-2025

[beginner99]

"AI" chips are limited by HBM and advanced packaging, not N4 process wafers - 50 series are totally incremental to that effort, fact selling 90 series for 2-3 times more requires it to be made in the first place and rebranded as pro card - where are those? Zero announcements.

These don't use HBM. they are for "professionals" that train / finetune their own models in a much smaller scale than new LLM training and don't need anything stronger or don't have the money for it. I have accesses to an A100 which back then was compared to now a steal. Anyway these use-cases exist and not everyone is training gigantic LLMs. you can also fine tune or training your own models completely unrelated to LLM / image generation.

 

[ToTTenTranz]

i wonder what the issue is? Yields?

There's plenty of redundancy in the 5090's GB202, both in the L2 and SMs. There is none in the PHYs, but I don't know if that's a problem.


Perhaps the only reason is that you can't really make that many GB202 in a 300mm waffer, and TSMC didn't start mass-producing those chips until recently (late December?).


der8auer mentions the fact that OEMs had almost no time to do Quality Control on their cards before, because Nvidia sent them the chips way too late.

Blackwell might have suffered from a bunch of internal delays (perhaps related to missing max frequency clock targets) but then Nvidia didn't want to delay GPUs' release, so they ended up in this situation where they started selling the cards without having accumulated any stock.

This last decision might have been pressured by the tariffs on China. Had the cards been announced in late March then perhaps Jensen couldn't have announced a $1000 MSRP for the 5080 and $2000 for the 5090. And unlike AMD, Nvidia knows full well the importance of release day reviews (which is probably what 90% of GPU buyers read/see).

 

[carrotmania]

And unlike AMD, Nvidia knows full well the importance of release day reviews (which is probably what 90% of GPU buyers read/see).

Haha, what? You think the abysmal reviews of the 5xxx paper launch were a good showing?

 

[Win2012R2]

These don't use HBM

Yes, those don't - but if they make those lower tier "AI" cards then they will have bin rejects that can be sold as 5090 and still make hefty profit.

Nvidia sent them the chips way too late.

Or they sent them too late because they did not want "partners" to know how poor Blackwell is.

 

[coercitiv]

This last decision might have been pressured by the tariffs on China. Had the cards been announced in late March then perhaps Jensen couldn't have announced a $1000 MSRP for the 5080 and $2000 for the 5090. And unlike AMD, Nvidia knows full well the importance of release day reviews (which is probably what 90% of GPU buyers read/see).

In hindsight this would make sense, especially as Jensen was a notable absence at the presidential inauguration.

 

[Heartbreaker]

This last decision might have been pressured by the tariffs on China.

10% China Tariff is pretty minor. It was 25% last time, and that probably moved a lot of final assembly out of China.

The real tariff of concern would be 25%-100% on Taiwan's chips. That's pretty much inescapable, but the Oligarchs know this one would hurt their bottom line, so this one may amount to nothing.

 

[SolidQ]

But only AMD have bad driver (c)

https://i.gyazo.com/8ee5e6cb18a1cd0776989d49fea8516b.mp4

 

[blckgrffn]

But only AMD have bad driver (c)

https://i.gyazo.com/8ee5e6cb18a1cd0776989d49fea8516b.mp4

She can never leave that building like I can never leave these forums

Even I would be irritated by the glitches in that video...

 

[Krteq]

But only AMD have bad driver (c)

https://i.gyazo.com/8ee5e6cb18a1cd0776989d49fea8516b.mp4

Take a look at launch driver known issues list... it's still growing

Open Issues:

• Changing state of "Display GPU Activity Icon in Notification Area" does not take effect until PC is rebooted [4995658]
• [VRay 6] Unexpected Low Performance on CUDA Vpath Tests for Blackwell GPUs [4915763]
• [GeForce RTX 5080] Valorant may crash when starting game [4951583]
• 3rd party utility NVIDIA Profile Inspector may crash when used with NVIDIA driver 572.16 [5073734]
• [No Mans Sky] Game may crash when enabling frame generation [5083718]
• [Windows 11 23H2] When exiting Final Fantasy XVI, PC may freeze [5083532]
• [Delta Force] Some PC configurations may experience performance regression when Resizable BAR is enabled [5083758]
• [GODOT/Red Dead Redemption 2/ImGui] Some Vulkan apps may display minor rendering artifacts with GRD 572.16 [5086373]
• [GeForce RTX 5080/5090] ARK Survival Evoled crashes shortly after launching game [5086449]
• [Shadow of the Tomb Raider] Game may crash when Ray Traced Shadow quality is set to High or Ultra [5086423]

 

[AdamK47]

I really wanted to like the FE card. The problem with them is the internal PCIe riser card not playing well with PCIe 5.0. My Windforce 5090 has no problems with it.

I wonder what it's like for those with the FE 5090 plugged into a riser. It's essentially a riser plugged into another riser.

 

[CP5670]

The FE is also the hardest one to actually find. In the past ebay/fb prices for the FE were comparable to the higher end AIB models.

I know you often trade up to a better card model in the future, let me know if you are ever looking to sell that one.

 

[Heartbreaker]

While the FE has a unique cooler, that keeps the GPU die cool, it looks like the memory gets over 90C. So probably better off with a partner cards in the long run.

 

[In2Photos]

I also saw that some cables were not able to plug into the FE card. Specifically the Lian Li strimer cable and possibly some Lian Li PSU cables. They are too wide to fit in the recessed hole of the FE.

 

[Sable]

Nvidia RTX 5090 seemingly has a spanner thrown in the works by new graphics driver, but we should be very careful around reports of ‘bricked’ GPUs

Some RTX 5090 owners are reporting that their GPUs are no longer being recognized, but there’s a possible fix

www.techradar.com

Bloody typical! I decide to upgrade to a new system after 11 years, EVEN manage to order a 5090 (today) and there's potential bricking of cards.

 

[igor_kavinski]

Nvidia's Grand GPU Beta Test 2.0 for willing gamer peasants?

Jensen: Welcome to the 5090. You select few have been chosen. Let us make history together!

 

[igor_kavinski]

 

[Elfear]

I watched that yesterday. He wasn't a big fan of the added input latency from MFG. Same input we've seen from other reviews where the latency is pretty atrocious if the actual FPS is low. He did seem to like the new DLSS though.

 

[CastleBravo]

There's plenty of redundancy in the 5090's GB202, both in the L2 and SMs. There is none in the PHYs, but I don't know if that's a problem.


Perhaps the only reason is that you can't really make that many GB202 in a 300mm waffer, and TSMC didn't start mass-producing those chips until recently (late December?).


der8auer mentions the fact that OEMs had almost no time to do Quality Control on their cards before, because Nvidia sent them the chips way too late.

Blackwell might have suffered from a bunch of internal delays (perhaps related to missing max frequency clock targets) but then Nvidia didn't want to delay GPUs' release, so they ended up in this situation where they started selling the cards without having accumulated any stock.

This last decision might have been pressured by the tariffs on China. Had the cards been announced in late March then perhaps Jensen couldn't have announced a $1000 MSRP for the 5080 and $2000 for the 5090. And unlike AMD, Nvidia knows full well the importance of release day reviews (which is probably what 90% of GPU buyers read/see).

IMO, the smart move for Nvidia would have been to prioritize production of the GPU dies that are going to compete against RDNA4 so they can have excellent supply for those GPUs when they are launched. No one shopping for a 5090 is going to settle for a 9070 XT , they will wait until they can get the 5090. Someone interested in the 5070 or 5070 Ti however are far more likely to cross shop AMD if they can't find the Nvidia product in stock.

 

[Dezii90]

IMO, the smart move for Nvidia would have been to prioritize production of the GPU dies that are going to compete against RDNA4 so they can have excellent supply for those GPUs when they are launched. No one shopping for a 5090 is going to settle for a 9070 XT , they will wait until they can get the 5090. Someone interested in the 5070 or 5070 Ti however are far more likely to cross shop AMD if they can't find the Nvidia product in stock.

Jensen wants to keep his 70% margin, plus amd is even shifting wafers to ai.