Everything wrong with RTX 50 series launch [complete list]

[appletree]

A quick summary of everything that went wrong. I thought I'll summarize it so we do not forget since there is so much of it and most of it is completely outrageous.

1. RTX 5070 being presented as similar performance to RTX 4090 thanks to "AI". While this is obviously false in so many ways (FG and MFG introduces latency over native, also MFG 4x has much more artifacts over 2x not to mention over native, 5070 has only half the RAM of 4090)
2. RTX 50 cards being artificially VRAM limited - i.e. compared to relative compute performance / VRAM vs. competition. This will no doubt become similarly crippling in the future as it is currently crippling to i.e. 3070 in certain games.
3. 50 series presented as having 3x AI, 2x RT and 1.5x shader speedup relative to 40 series. But somehow none of this shows in the benchmarks. For example there seems to be no RT speedup relative to raster speedup in 5090 and 5080. While the consensus nowadays seems to be that Blackwell is basically the same architecture as Ada with just minor tweaks. And yet competitors were able to i.e. provide meaningful improvement in RT relative to raster despite facing similar node limitations as Nvidia.
4. Also the naming scheme of what 5080, 5070Ti etc. are supposed to mean seems to be quite inflated relative to what it used to mean. 5080 being only small improvement over 4080Super. 5070Ti is basically 1:1 performance replacement for 4080.
5. We were promised "great availability" yet the cards are nowhere to be found and it is by far the most paper launch of launches by Nvidia in history.
6. MSRP prices are nowhere to be found either (as expected). But the 5070Ti launch was cause of yet another controversy as there is no FE version of the card and thus Nvidia sent Asus Prime cards to some reviewers as the "MSRP" variant. Yet Asus had this card listed with price 20% over MSRP. Thus depending on who you asked (Nvidia or Asus) the same card had completely different value proposition.
7. RTX 40 series cards are also nowhere to be found. Creating artificial scarcity and driving prices of all Nvidia GPUs up. A scenario orchestrated by Nvidia either on purpose or by sheer incompetence (but judging by lessons from history, the former is more likely)
8. Despite problems with melting 12VHPWR connectors & RTX 4090s, Nvidia didn't make any design improvements and is now using it for 5090s which for certain AIB variants can exceed 600W. While 4090 is a design downgrade from 3090Ti which had certain protections built-in so that you couldn't cut 5 out of 6 cables and keep the card still running. But this is completely possible on 40 and 50 series - i.e. delivering all the current via single cable until meltdown. It is now consensus of most electrical engineers that Nvidia's design is inexplicable as it leaves no safety margin thus even a slight deviation from ideal scenario can create meltdown or potential fire hazard scenario. Also the tactic of blaming customer for "improper" setup etc. is a completely inappropriate excuse. Even if the meltdowns were caused by some suboptimal handling by customer - it is quite certain it wouldn't have happened if the cards were using the standard PCIe cable connectors. While at the same time this could easily be caused by imperfections in connectors / cables and their respective manufacturing processes instead of just customer handling error.
9. Nvidia enforces use of 12HPWR connectors even for their AIB partners for completely unknown reasons. While the competition keeps using PCIe connectors and doesn't seem to have any problems powering GPUs with comparable TDP to Nvidia cards.
10. Blackwell has stability issues when PCIE 5.0 is enabled on certain (or maybe most?) motherboards. The problem could be farther exacerbated by early driver versions. Supposedly some cards have been bricked because of these problems, but usually it is enough to downgrade PCIE to 4.0 in motherboard BIOS. However this is a pretty big oversight and no explanation from Nvidia if this is hardware related or could be fixed by BIOS update etc. Also some PCIE extenders don't work even if you downgrade to 4.0 speed.
11. In their flashy presentation Nvidia forgot to mention that support for 32bit PhysX on Blackwell is discontinued. Among affected games are titles like most of the games from Batman Arkham series, most from older Metro series etc. Now if you want to play these games on Blackwell GPUs you can either emulate it on CPU (while for 8core CPU you can expect around 40ish fps on your 5090) or you can buy separate older Nvidia GPU just for "PhysX". Isn't that nice and convenient? Yet no word from Nvidia of some emulation layer that could translate it to 64bit calls.
12. Update 1: Missing ROPS from 5090, 5090D and 5070Ti cards. I wonder when was the last time something like this happened. At least there is an official statement from Nvidia: "We have identified a rare issue affecting less than 0.5% (half a percent) of GeForce RTX 5090 / 5090D and 5070 Ti GPUs which have one fewer ROP than specified. The average graphical performance impact is 4%, with no impact on AI and Compute workloads. Affected consumers can contact the board manufacturer for a replacement. The production anomaly has been corrected."
    Will be interesting to see if truly only < 0.5% GPUs are effected since at this point I am a bit skeptical of anything Nvidia is saying. Also note that by ROP they mean one ROP unit which consists of 8 ROPS. (just to elaborate on reported 8 ROPS missing vs. 1 ROP missing in Nvidia statement)

There is so much of it that it's still possible I forgot about something. Anyways, since I need Nvidia for work I don't have a choice. And yes Nvidia is still the best, but I nevertheless hope the incredible arrogance displayed by them will one day lead them to similar path to Intel. As this seems highly reminiscent of Intel couple of years ago (except maybe even worse). 4 cores as desktop standard forever, miniscule 5% improvements between generations, shady tactics with other HW manufacturers, total market domination, especially server market etc. All reasons why I am not sorry for what happened to them (again despite most of my CPUs were intel)

PS: I may update this list if new facts emerge

 

[appletree]

Yes thanks all for the news of missing ROPs. Wow I definitely didn't expect there can be such a major development so soon after I finished my initial list. This launch is truly incredible. I updated the list together with official Nvidia statement.

 

[appletree]

1 - Agree, but not a huge deal because nobody buying one was born yesterday and if you know enough to watch the presentation then you know enough to not believe it's as powerful as a 4090

3 - Really? It was always assumed to be just a refresh because it's on the same node and the design is pretty much the same. It was never going to be a huge performance uplift over 4000 series

5 - Where was 'great availability' promised? I don't recall seeing that anywhere.

7 - They stopped making 40 series so they could make 50 series, they're on the same node. What do you expect?

10 - PCIe 5.0 is working fine here. It's also a motherboard issue not a GPU issue. The cards are designed to 5.0 spec, but if the motherboard is doing it wrong how can you blame the GPU?

1, 3 and 5: all all from Jensen's CES presentation, i.e. from this video

Here I have compiled the relevant timestamps:
TOPS, RT & shader speedup vs. Ada 39:07
the infamous "5070 == 4090 thanks to AI" 42:00
"producing at very large scale availability starting January" 42:50
Now we can probably agree that that all of that looked very suspicious - maybe with some hints of truth but then considering Jensen's reputation
of basically overselling the improvements with every generation, yeah clearly I was very skeptical of anything he said.
And as it turned out later on all of them turned out to be 100% lies. Not just overselling it but 100% lies.
Even if we were to somehow agree that Jensen saying lies shouldn't be a "big deal" let me ask from different point of view.
Why from the manufacturer's perspective should it be such a big deal to tell the truth for once?

7: I would expect that well timed switch to new gen products shouldn't cause extreme shortages of old AND new products at the same time. If they switched production to new products then where are they? At the very least they could have prepared a buffer of older cards if the were expected shortages of new cards. Also when they introduced 40 series they had the opposite problem that there were too many 30 series card on the market which was not good for their profits. Therefore I would speculate that the current market development is more of a consequence of them intentionally going for the opposite extreme (to avoid profit loss) instead of some production continuity on similar node problems.

10: Glad to hear PCIe 5.0 is working OK on at least some combinations of cards and motherboards. But I don't think at this time it's so clearcut that the blame lies with motherboard manufacturers. I.e. there were some problems with PCIe raisers which didn't work on 50 series even if in MB was set to PCIe 4.0 That could point to some signaling problems on the cards themselves. And since the FE cards have their PCIe connector on a separate PCB connected to the main PCB only via thin cables it's not difficult to imagine this could potentially be cause of signaling issues. I think we need to wait for official Nvidia statement for this one. The one we got so far says they are investigating it, not that it's clearly supposed to be MB manufacturers problem.