![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
AMD RDNA5 rumors point to AT0 flagship GPU with 512-bit memory bus, 96 Compute Units
- Thread starter erek
- Start date
when the "investor" is openAI, otherwise no, Sony, me and you are way more an AMD customer than Vanguard or Blackrock (could be missing the point quite a bit here)
Their GAAP operating margin of the last 2 years were a very low 10% the last 2 years already, they stopped telling it for customer gaming because of how low it got.
As for the proposition make a 96GB ai inference super GPU, sell it to us for only $2500 while we have an easy way to make $2400 of annual revenue from them instead of making that revenu yourself, not sure if someone with options would bite, used 48GB RTX pro blacwell sell for over $3000 on ebay
THe word "pressure" here is being quite dramatic, a 7.5 millions gross orders is not pressuring anyone in that space, that around a single Helios rack worth of sales.
AMD has to focus on the Datacenters simply because of the software and development chain.
Datacenters can host and undercut costs, delivering ROI that any small to medium-scale developers can't hope to match. So if a developer has any hopes of a large product release, they need to look at Datacenter hosting, be it Azure, AWS, etc...
The short version is this:
If AMD isn't in the data centers, developers aren't as likely to build for AMD hardware. If developers aren't building for AMD hardware, they aren't buying AMD hardware to develop on; if they aren't developing on AMD hardware, they aren't developing for AMD hardware; if they aren't developing for AMD hardware, their customers aren't using AMD hardware.
That, respectively, leaves Nvidia hardware, and if you are developing on Nvidia, you use CUDA; if you aren't using CUDA, then why are you using Nvidia?
And if the product stack is built using CUDA, unless you are redeveloping your whole stack, you are staying with Nvidia, because why else would you replace your stack?
What I was getting at is the people who are buying AMD stock by the bucket have more say in release and product pricing than you, me, or Sony can ever dream to have.
Lisa was getting raked over the coals during the 2025 financial reports, with investors demanding higher profits and better margins. She was instructed to bring it above 54.5%, and she was more than happy to let them know that she got it to 57%
If AMD does offer a discount to Sony, they are offsetting that discount by charging more elsewhere.
many are extremelly passive, I doubt state street has more to say in release and ps6 soc cost than Sony or an OpenAI (that has yet to have stock), they want high profit/margin that look sustainable and with a good moat, how it is achieved they do not mind that much in details.
AMD did offer some high volume discount to Sony of course, they are not paying per mm2 of die what we do pay and Sony is actively involved in that soc design, size/power choice in a way investor have not a single idea before the fact.
so my understanding is that nvidia is now diverting all 5090 supplies to professionals
AMD has a pro card MI350P (but is HBM 144gb — so very costly)
https://www.phoronix.com/review/amd-instinct-mi350p
AMD is probably planning on releasing a prosumer version of gddr7 based RDNA 5 AT0 asap to partake the profits
in turn Nvidia might try to pre-empt AMD's schemes by launching a 6090 pro version
effectively no top end gaming card, until & unless Open AI goes bankrupt !!!!!
AMD has a pro card MI350P (but is HBM 144gb — so very costly)
https://www.phoronix.com/review/amd-instinct-mi350p
AMD is probably planning on releasing a prosumer version of gddr7 based RDNA 5 AT0 asap to partake the profits
in turn Nvidia might try to pre-empt AMD's schemes by launching a 6090 pro version
effectively no top end gaming card, until & unless Open AI goes bankrupt !!!!!
- Joined
- Dec 19, 2005
- Messages
- 18,432
RTX 7000 Pro (Rubin?)
yeah. MLID keeps insisting that Nvidia & AMD are all set to unleash next gen prosumer cards by Q2/Q3 next year
- Joined
- Dec 19, 2005
- Messages
- 18,432
MLID gives con artist vibes
Do not ask me the logic of it, but it will probably be RTX 6000 Rubin, not 7000 pro, they have been called RTX 6000 since Turing and each generation got one.
There was talk that the 2GB gddr7 supply could be drying out (Official for Micron now), which could make them hard to make, they could go 54GB edition but that starting to mess up the RTX pro line quite a bit and way too expensive for gaming, forcing them into a super series with 9 (96 bits version 5050)-12-18-24GB cards with no 5090 that time around.
Last edited:
- Joined
- Aug 27, 2025
- Messages
- 550
He is in reards to NVIDIA:
When evaluating Moore's Law Is Dead (MLID) specifically on NVIDIA leaks, a clear pattern emerges.
Mostly False or Overhyped
MLID's Nvidia leaks have historically suffered from a lack of reliable, primary deep-node sources, leading to a tendency to mistake Nvidia's internal contingency testing for final production specifications. [1]
- The "600W–800W RTX 4090" Panic: Before the Ada Lovelace launch, he repeatedly claimed Nvidia was terrified of AMD's chiplet design and was pushing its flagship GPUs to melt-inducing 600W–800W power limits to compete. The RTX 4090 launched with a highly efficient 450W TDP and completely dominated the market.
- Samsung vs. TSMC Nodes: During the Ampere (RTX 30-series) era, he released wildly conflicting videos within short windows, claiming Nvidia would use TSMC 7nm for top cards, then changing it to Samsung 10nm marketed as 8nm. Throwing multiple claims at the wall allowed him to retroactively say he was "right" on one.
[1] - The "Anti-DLSS" Nvidia Conspiracy: He claimed through anonymous sources that Nvidia was secretly orchestrating rumors that AMD was anti-competitively blocking DLSS implementations in sponsored games (like Starfield). The community and other investigators widely dismissed this as baseless speculation. [1]
- Joined
- Jul 5, 2012
- Messages
- 4,743
lol that is what I notice about MLID, he literally says anything and everything so he can never be wrong. He will say Nvidia will do something then also say "OR they may not!" That way no matter what the outcome is, he can claim to be right.
- Joined
- Aug 27, 2025
- Messages
- 550
He might have a source at AMD, but his NVIDIA claims are like you say, scatter al sort of possible claims and hope some lands true and then claim he was right.
But his failures on NVIIDA means I 100% consider him noise on the topic and a waste of time.
Samsungs 8 nm was basically their 10++ node, same production line, just tighter tolerances.
The best we can hope for without an ai crash is a 9070 successor probably using a version of what’s going into the helix minus the cpu.
They might have some big rdna5 silicon ready to go for us, but it will all go to enterprise unless there’s a crash.
They might have some big rdna5 silicon ready to go for us, but it will all go to enterprise unless there’s a crash.
agree
Considering Ampere did use bothTSMC 7N for the A100 type cards and Samsung for others (on the 10nm enhanced node) and the strong negotiation before that, not a terrible level of leak, youtuber only have access to low level quite indirect source that speculate from limited internal access/memo (TSMC-Nvidia internal source will maybe comments of the type; that not really true or on already happened/announced explication, but not add anything new information wise yet to be public). Getting that Ampere would be a dual node like that generation was not bad at all.
Cache decision that seem more for AI than gaming, API matching RDNA-CDNA under UDNA, large monolithic expensive die, they look like they want a bit of that Nvidia playbook and spending that much on a niche gaming gpu to be paid for in part via enterprise sales.
It is a large die, they will have a lot of die not near 100% enabled core / 16 memory controller working (if they go to 512 bits), they could like Nvidia do, make a binned down version they sales to public/regular customer facing channel.
We can buy the W7900 for $4,400 on newegg right now: https://www.newegg.com/amd-100-300000074-radeon-pro-w7900-48gb-graphics-card/p/N82E16814105114 that about exactly to MSRP + inflation since launch, currently a rebate on the W7800 https://www.newegg.com/amd-100-300000075-radeon-pro-w7800-32gb-graphics-card/p/N82E16814105115, sold under MSRP before inflation.
Things could radically change on RDNA 5, but AMD right now is really far of a we sell all we can make to enterprise so we have to make a choice between the 2, it is more retailer sell them under MSRP to try to past their stock for the binned down version, an AT0 consumer card they could very well plan to used bin down die for non-enterprise channel.
Last edited: