FeelasThis is however fascinating, since I guess I just hit an ASRock which doesn't want to go over 2400MHz, which would go really closed to the reported instability issues around 2500MHz for some parts.

Which would mean, that some of the worldwide stock is repurposed 7900GRE for China which didn't sell (like mine). It looks like a very confusing case, because I feel that should be refundable... It's not even silicon lottery, it's a hidden-revision-that-is-not-in-specs at all. Either purposefully locked based on BW or binned based on max stable memory controller speed.

alwaysstsLots of really fascinating info/takes is this thread, thanks all for your contributions (and for helping inform people/make the whole ecosystem better).

I, for one, had forgotten the limits to specs (other than flops) wrt exports to China (which may impact some 'stock' product decisions beyond 4090D, perhaps even how certain chips are designed with that limitation in mind in the future) in ways I haven't personally looked into on a deep-enough level to comment. Thanks for the info/reminder.

That said, GDDR would be volatile memory, correct? I imagine (but am not well-versed on the current definition of I/O bw) this is in relation to external bandwidth (links/protocol) and if anything may perhaps only relate to their (capabilities of) external cache structure? While I don't know the actual bw (or limitations of the link), using it's observational impact on performance I've generally equated AMD's L3 structure to something like adding 3mhz(/gbps) over the bus in relation to core speed (as the cache is likely clocked in relation to core; haven't looked into it). So, for instance, if something like a 7800xt/7900gre were running at 2800mhz core clock, the L3 would be contributing what equates to 268.8Gbps for DRAM bw. Obviously it's faster as the size is smaller, but in terms of perf does appear to shake out if you look at it wrt bw limitations. The same is true of nVIDIA's L2 cache and roughly double at same size (similar to core clock*6mhz over the bus), which nVIDIA all-but-confirmed with one product release not so long ago (something in the 4060/4070 series iirc) to excuse the more narrow bus wrt the competition.

In the case of 4090 (with 72MB L2), for instance, that translates to something like ~2730*4.5*384/8 = ~589.68Gbps, which is indeed oddly close to the number you stated. That's internal though, so I don't know if that was part of that decision-making process of the chip and/or effected by those export rules (or rather the rules were built around them after it was completed and/or released)? Now you've got me curious, but I always had a feeling these 'rules' were constructed as a compromise between the government and nVIDIA as for them to still be able use that chip in China, but for them to be conscious of using restraint moving forward.

I'm sure someone more studied could read those last couple paragraphs and declare me a simpleton, and that's fair-enough! I respect there are people that know much more in the intricate areas of these things (and their definitions/classifications) than I do, have have kept up on it better, and/or have information most don't have. I generally focus on the real-world impact to most users, and where limitations are/how they can be improved. I don't always know the correct specific language/terminology regarding how each area is classified. I'm sure there are people around here that can shed much more light on it than myself. I don't *think* what you are implying would apply, but I absolutely could not state that as a fact as I just don't have the inherent knowledge or haven't done the reseach to declare that as fact.

alwaysstsKeep banging that drum, buddy. I'd love for you and any other people that have similar issues to have them completely resolved. Compatability with old titles can sometimes be a tricky thing (depending on how something was programmed), and sometimes not (it's just a matter of getting ahold of the right person and them having the time to fix it/implement a change). All I can hope is that if there is an easily-enough implementable fix on AMD's side that you are able to make that connection to them and it can/will be sorted to the best of their capability and your appeasement, even if it takes them a few tries to address your (or others) particular issue(s). I wish you much luck, and appreciate you for trying to get that sorted for your community of players. That's what it's all about, right? Just trying to make things the best they can be if it's possible; calling out issues/perceptions and hoping it reaches the right ears so that both they know about it and getting someone capable to help with a solution that is applicable/agreeable to the most people, if able/possible...and doing one's best at going about those things without being offensive/negative as possible.

(The last part is a personal reminder [I mean well, but know I can sometimes come across unintendly harsh towards people when trying to make a point], your attitude towards the issue appears good-natured and on-point, and once-again I applaud your persistance. :))

FordGT90ConceptThe GDDR modules installed on the card are up to the AIB partner's discretion so long as they meet or exceed reference spec. In this case, PowerColor (see OP) has much more capable Samsung chips installed compared your ASRock model. In short, ASRock skimped on the GDDR and PowerColor did not (likely used the same modules the 7900 XT and XTX have). All AMD did was enable overclocking of the VRAM so you can unlock whatever capability the GDDR modules have from the factory that were previously limited by the VBIOS.

FeelasSo you are suggesting that I have received a different memory PN or Samsung is selling overspecced ICs? Given that TPU BIOS dumpfor the card is correct, the BIOS supports either Hynix H56G42AS8DX014 or Samsung K4ZAF325BC. Judging by Google, both chip models are possible to find on GRE/XT/XTX all alike and the Samsung part doesn't come in a variant specced for 18Gbp: it is available as SC16 for 16Gbps or SC20 speecced for 20Gbps. Looks like an issue with the memory controller instead, which of course points that it is "in the reference spec", given they underclock for 2250MHz. It is completely baffling they would put 20Gbps VRAMs into a card capable of only 18Gbps...

FordGT90ConceptI did a lot more digging into what happened. Short version is this:
1) November 2022, the performance export ban was issued that mentions the 600 GB/s rule on non-volatile memory: hackaday.com/2022/11/09/chinese-chips-are-being-artificially-slowed-to-dodge-us-export-regulations/
2) July 2023, 7900 GRE launches: www.tomshardware.com/news/amd-radeon-rx-7900-gre-launch
3) October 2023, Department of Commerce announces change in the export ban that strongly targets tensor cores (TFLOP calculation): www.tomshardware.com/news/no-nvidia-isnt-breaking-gpu-sanctions-analyst
4) November 2023, the rule goes into effect, 4090 disappears because it's not compliant, and 7900 series sales surge in China to fill the demand: hothardware.com/news/amds-flagship-radeon-rx-7900-xtx-and-xt-gpus-flourish-in-china

I believe GRE was created in case the rule was modified to apply the 600 GB/s rule to volatile memory which would limit Chinese sales of 7900 XT and 7900 XTX. Dell apparently misinterpreted the rule and proactively applied it in November 2023: www.techpowerup.com/316044/dell-allegedly-prohibits-sales-of-high-end-radeon-and-instinct-mi-gpus-in-china

AMD wasn't the only company to release a 600 GB/s product preemptively; Birin Technology (a Chinese company) did the same (see hackaday article above).

Because the 2023 replaced the rule invalidating the 2022 rule which mentioned 600 GB/s, AMD felt they were no longer threatened by it so they enabled overclocking in the driver for the card.

So, all good now, except that Chinese can't get 4090 anymore (hence 4090D www.tomshardware.com/pc-components/gpus/nvidia-launches-china-specific-rtx-4090d-dragon-gpu-sanctions-compliant-model-has-fewer-cores-and-lower-power-draw).