AMD RDNA5 rumors point to AT0 flagship GPU with 512-bit memory bus, 96 Compute Units

socK said:
I mean I'm not that surprised.

At that framerate you're flipping a frame in under 3ms. Low ping and encoding and decoding latency later and, yeah, you're probably there.

Even at 120hz, it's only 5.55~ milliseconds longer to render a frame. Maybe the bigger frame adds encoding/decoding latency on top, but it's fast at this point any way you slice it.



I played multiple games and all of Split Fiction on it when I was in between computers for a couple months. I played at 4k 120hz, at like 100 m/bits bitrate with 5 ping.

Reflex works, VRR works. Latency was imperceptible. I spent a lot of time trying to feel it. I couldn't. It is not laggy in the slightest.
Click to expand...
I would warn against using "sample size of one" combined with subjective data as several things in the past showed us that is a very bad metric.
Back in the multi-gpu days there were a lot of people claiming it was fine and called people perceiving something as "off" for delusional until test methods indeed showed that it was full of frame time fluctuations and the peopel claiming everything was fine was indeed claiming "horseshite": https://www.overclockers.com/micro-stutter-the-dark-secret-of-sli-and-crossfire/
Same thing happend with controller lantency and then people started testing and same thing happend I remember reading this years ago: https://displaylag.com/console-latency-exploring-video-game-input-lag/

You seem to confuse A<->B ping with input lantency those are 2 seperate measurments that cannot be interchanged.

That complicates a business case even more because now you have segment people into who perceives "lag" better than others as history shows us that homo sapines differ wildy in abilities and perceptions.

I would wager that there is a reason why "cloud gaming" has not killed of +$500 consumer GPU's and we all are all not just subcribing to online service.

Consumers move fast if you have identical services but one has a higher cost and what we see is ~$5B "cloud gaming" vs ~$110B "on-prem gaming"
 
littlegreenmenfromhome said:
You seem to confuse A<->B ping with input lantency those are 2 seperate measurments that cannot be interchanged.
Click to expand...

I am not, the math maths fine for their numbers.

You can see the info for the round trip network time, the framerate the game is running at, the framerate the service is running at, and details about the encoder being used and the encoding/decoding time.

There's only so much latency/overhead it can possibly add when it's giving you all the numbers before the frame is effectively delivered and the next cost is whatever it's taking your system to do something.

And when it's spinning at literally several hundred FPS it's obviously going to be low. The individual things that need to happen at that point are all happening in single digit milliseconds.
 
littlegreenmenfromhome said:
Your "busines plan" is already in trouble because you cannot deliver an accurate requirements specification. "10-15" != "30" != "60-100" as you would be cutting of portions of games by increasing pings and thus decrease your potential profitability.
Click to expand...
I think a possible business plan, is better image quality at higher framerate and lower latency than what console offer.

An Xbox gamer that purely play its gamepass offer, which are virtually all available on PC can still pay for gamepass and play them on this instead of an Xbox, not necessarily that easy of a sales (it is not polished enough as an offer right now), but once it is a couple of click away on a smartTV and work all the time (downtime would kill the enthusiast for it) with a solid GUI, the appeal for that audience could be obvious enough, no need to download the games that can be big, better visual, better frame rate, lower latency.

$20 a month is not cheap considering consoles low price and their life expectancy, $500 for 7 years is a really cheap way to enter the market, but right now an XboxX is $600 and we can expect the next generation for around 2027, that can make a $20 a month alternative to play on a $2000 computer instead attractive.

It does not need to compete with low latency 360hz monitor/pc gaming experience necessarily, just need to beat consoles, a significantly more reasonable bar, people on console played Cyberpunk at 160 ms of latency (even when removing display latency) still 2 years/5 major patch after launch... with those kinds of number gaining reflex tech /high FPS can easily make up for direct fiber in the home people the cost of encoding/decoding/round trip
 
The optimum way to do the cloud gaming stuff is to host the gaming services at the ISP Hub/HeadEnd/edge locations. With all of the functions (especially at cable/so sites) moving to virtualized things like vCMTS/vPacketCore/vBNG/vCDN and fixed hardware getting more converged, they all have tons of unused space and power at those locations. Which is actually why you see a lot of these larger ISPs trying to do that. The larger ones also all have their own roadms and fiber networks for Transport everywhere as well.

Then as stated, when not in use, you use something like https://hydrahost.com/data-center/ or https://www.juicelabs.co/ (though I'm not fully sure juice labs makes sense outside of your own organization yet) tp rent them out.
 
Last edited:
LukeTbk said:
Which create an obvious business case where the 3 party can "win", like always when single user infracstructure become sharing by many infracstructure specially with 0 downtime, the math at least is quite clearly there, that it can cost less money to game on those with Nvidia making still good money (of couse company prefer people paying all of it upfront and not having to manage infracstructure themselve if people are ready to do so, but they can easily be forced into something more customer friendly by them/competition).
Click to expand...
As opposed to just buying a $300 graphics card and using it for 3-4 years?
socK said:
I played multiple games and all of Split Fiction on it when I was in between computers for a couple months. I played at 4k 120hz, at like 100 m/bits bitrate with 5 ping.

Reflex works, VRR works. Latency was imperceptible. I spent a lot of time trying to feel it. I couldn't. It is not laggy in the slightest.
Click to expand...
Every time we go through this cloud gaming situation, we always get someone that says that the input lag was imperceptible. When people actually try it, they say it's horrible. When reviewers actually review it, the overall input lag was over 100ms. An Android TV device or a Chromecast would get over 200ms because that's how bad those devices are. I'll wait for reviews instead of a testimony.
 
DukenukemX said:
As opposed to just buying a $300 graphics card and using it for 3-4 years?

Every time we go through this cloud gaming situation, we always get someone that says that the input lag was imperceptible. When people actually try it, they say it's horrible. When reviewers actually review it, the overall input lag was over 100ms. An Android TV device or a Chromecast would get over 200ms because that's how bad those devices are. I'll wait for reviews instead of a testimony.
Click to expand...
Home streaming to a steam deck works alright for chill games.
 
DukenukemX said:
As opposed to just buying a $300 graphics card and using it for 3-4 years?

Every time we go through this cloud gaming situation, we always get someone that says that the input lag was imperceptible. When people actually try it, they say it's horrible. When reviewers actually review it, the overall input lag was over 100ms. An Android TV device or a Chromecast would get over 200ms because that's how bad those devices are. I'll wait for reviews instead of a testimony.
Click to expand...

It's not over 100, it's not even close. It's a fraction.

That wouldn't even be possible unless your ping is shit or the game is lagging horribly.

It's not taking TWELVE times longer than the actual frame time to get one frame over to you.
 
DukenukemX said:
As opposed to just buying a $300 graphics card and using it for 3-4 years?
Click to expand...
in that space, less obvious (i.e. it work much better if you can rent the gaming GPU to do other task, because of the latency challenge that do not let you share gaming server over time zone), the buying a $1000 gpu type of high end alternative, but if vGPU can split an future RTX 6000 in 4, 5, 6, low end system it should work yes, you are 4 gamers buying a $300 GPU together and splitting the cost to a company that moneytise your downtime. (it work much better with casual gamer, if you are a 12+ hours a week type of gamers, then no, has the sharing will not happen often enough)
 
From what I understand, AT1 in its initial format has been canceled, but an updated card is still planned as a mid-cycle release. The targeted competitor to beat is the Nvidia 6080 Super, at a $1000 MSRP.
 
socK said:
I mean I'm not that surprised.

At that framerate you're flipping a frame in under 3ms. Low ping and encoding and decoding latency later and, yeah, you're probably there.

Even at 120hz, it's only 5.55~ milliseconds longer to render a frame. Maybe the bigger frame adds encoding/decoding latency on top, but it's fast at this point any way you slice it.



I played multiple games and all of Split Fiction on it when I was in between computers for a couple months. I played at 4k 120hz, at like 100 m/bits bitrate with 5 ping.
Click to expand...

socK said:
I am not, the math maths fine for their numbers.

You can see the info for the round trip network time, the framerate the game is running at, the framerate the service is running at, and details about the encoder being used and the encoding/decoding time.

There's only so much latency/overhead it can possibly add when it's giving you all the numbers before the frame is effectively delivered and the next cost is whatever it's taking your system to do something.

And when it's spinning at literally several hundred FPS it's obviously going to be low. The individual things that need to happen at that point are all happening in single digit milliseconds.
Click to expand...

You math is off:
Pure render time at 120 fps: ~8,333333333333334 ms
5 ms pure render time would give you 200 fps
And that is without display scan-out, driver overhead, frame queuing, encoding/decoding, input polling, OS scheduling jitter, GPU command submission, VRR sync or frame pacing, transmission and routing

Your math is bad or someone is feeding you bogus numbers or you are confusing ping numbers with input latency and like I mentioned before those are interchangeable.
To use you own number of "5 ms at 120 fps" would render you have a negative time of -3,33333334 ms per frame and we are moving into "FPS Inflation":
1759642485293.png
 
littlegreenmenfromhome said:
You math is off:
Pure render time at 120 fps: ~8,333333333333334 ms
5 ms pure render time would give you 200 fps
And that is without display scan-out, driver overhead, frame queuing, encoding/decoding, input polling, OS scheduling jitter, GPU command submission, VRR sync or frame pacing, transmission and routing

Your math is bad or someone is feeding you bogus numbers or you are confusing ping numbers with input latency and like I mentioned before those are interchangeable.
To use you own number of "5 ms at 120 fps" would render you have a negative time of -3,33333334 ms per frame and we are moving into "FPS Inflation":
View attachment 758090
Click to expand...

My math is fine, you just ignored the post I was replying to and lost the context.

"Even at 120hz, it's only 5.55~ milliseconds longer to render a frame."

The context being I'm replying to a post where it's running at 360 fps. Frame time at 120 fps is 5.55ms longer than 360 fps

(1000/120)-(1000/360)=5.55~.
 
Mackintire said:
From what I understand, AT1 in its initial format has been canceled, but an updated card is still planned as a mid-cycle release. The targeted competitor to beat is the Nvidia 6080 Super, at a $1000 MSRP.
Click to expand...
would be interesting if that happens.

as of now rumours point only to
  • AT0 (xcloud),
  • AT2 (first party series X2)
  • AT3 (strix halo successor)
  • AT4 (third party series S2 / handhelds Z3 extreme)
 
Last edited:
Intel's Pantherlake embedded B390 GPU has likely lit a small fire under AMD. I'm betting the next round of planned but not yet announced RDNA chips for the laptops will finally be RDNA 4 class or higher.
AMD didn't see the need to innovate as they already held the crown. They were partially depending on Intel to fail again. Only it looks like Intel managed to hit the target this time.
 
Marees said:
Microsoft has contracted AMD for xcloud gaming hardware chip which can virtualize & run 8 series s2 instance at one go

series s2 (AT4) = ps5 pro
series x2 (AT2) = triple series s2

AMD is expected to cut the xcloud chip (AT0) by 25% for top xx90 card

so basically 10090xt should be equal to 6x 9060xt or 6x PS5 pro

https://x.com/MrMPFR/status/2017277227168076016?s=20
Click to expand...
My math isn't mathing...
The PS5 Pro is roughly equivalent to the RX 7700 XT, which is roughly equivalent to the 9060xt, the RTX 5090 is roughly 2.8x the performance of either of those cards..

So that would make this mystical 10090xt roughly double the performance of the RTX 5090, which I find a somewhat bold claim.... That would be an incredible, almost unbelievable leap in performance per watt.
 
Lakados said:
My math isn't mathing...
The PS5 Pro is roughly equivalent to the RX 7700 XT, which is roughly equivalent to the 9060xt, the RTX 5090 is roughly 2.8x the performance of either of those cards..

So that would make this mystical 10090xt roughly double the performance of the RTX 5090, which I find a somewhat bold claim.... That would be an incredible, almost unbelievable leap in performance per watt.
Click to expand...
5090 is 4nm
10090xt would be 3nm

so double 5090 should be doable. but the price is a big question mark
 
Lakados said:
My math isn't mathing...
The PS5 Pro is roughly equivalent to the RX 7700 XT, which is roughly equivalent to the 9060xt, the RTX 5090 is roughly 2.8x the performance of either of those cards..

So that would make this mystical 10090xt roughly double the performance of the RTX 5090, which I find a somewhat bold claim.... That would be an incredible, almost unbelievable leap in performance per watt.
Click to expand...

Possible if the card is working like 2 video cards in one. With Chiplet design it's possible it could work like crossfire, just hopefully more refined. Were still way too early I think to really know, but it sounds promising so far.
 
Gideon said:
Possible if the card is working like 2 video cards in one. With Chiplet design it's possible it could work like crossfire, just hopefully more refined. Were still way too early I think to really know, but it sounds promising so far.
Click to expand...
that usually purely to be cheaper at the cost of some performance going chiplet (when you do not go over reticle maximum size), the AT0 is not rumored to be bigger or more watt than a 5090, doubling its raster performance while loaded in ML capacity at the same time, would be a giant leap per watt/mm, even with tsmc 3N helping.
 
LukeTbk said:
that usually purely to be cheaper at the cost of some performance going chiplet (when you do not go over reticle maximum size), the AT0 is not rumored to be bigger or more watt than a 5090, doubling its raster performance while loaded in ML capacity at the same time, would be a giant leap per watt/mm, even with tsmc 3N helping.
Click to expand...
i think nvidia pushed the 5090 to the inefficient part of the power curve

a redesign on 3nm that operates purely in the sweet efficient zone should give much better performance per watt
 
Marees said:
i think nvidia pushed the 5090 to the inefficient part of the power curve

a redesign on 3nm that operates purely in the sweet efficient zone should give much better performance per watt
Click to expand...
way more performance per watt would not be surprising (the 5080 for example, or a downclocked 5090 are already quite better than a pushed 5090), per mm of die too... would be impressive for TSMC 3 vs 4.

Spark can give some idea for tsmc 4->3, the 5070 6,144 cuda core chiplet can fit on a ~30% smaller die I think, if they can make per CU perf +33% and fit all of them in there, could be special.
 
Marees said:
5090 is 4nm
10090xt would be 3nm

so double 5090 should be doable. but the price is a big question mark
Click to expand...
The performance increase from TSMCs N4 processes to their N3 processes varies depending on the flavor chosen, but given the density of GPU's you are getting at best 10%.

An RTX 5090 pulls 800w, so lets assume moving to 3nm yielded the best case scenario, which yields a 10% drop in power consumption with a full 5% performance increase per watt.
You are still looking at a GPU trying to draw upwards of 1400 watts. That is an absolute monster.
 
Lakados said:
The performance increase from TSMCs N4 processes to their N3 processes varies depending on the flavor chosen, but given the density of GPU's you are getting at best 10%.
Click to expand...
how we can see something like that ? for example spark seem to be rumored to be 30% smaller than a 5070 for the gpu tile, but being a chiplet/tile make that comp a bit hard, but that would be around a 5060 sized/powered die, doing 5070 stuff, I think ?, m3 gpu was about 30% faster than M2 when they shifted to N3, but apple was evolving fast.

Lakados said:
An RTX 5090 pulls 800w
Click to expand...
https://www.techpowerup.com/review/msi-geforce-rtx-5070-ti-ventus-3x-pz/38.html

a bit over 520w, 600 max I would think, without mods
 
LukeTbk said:
how we can see something like that ? for example spark seem to be rumored to be 30% smaller than a 5070 for the gpu tile, but being a chiplet/tile make that comp a bit hard, but that would be around a 5060 sized/powered die, doing 5070 stuff, I think ?, m3 gpu was about 30% faster than M2 when they shifted to N3, but apple was evolving fast.


https://www.techpowerup.com/review/msi-geforce-rtx-5070-ti-ventus-3x-pz/38.html

a bit over 520w, 600 max I would think, without mods
Click to expand...
Yeah, but the claim was basically saying the AMD flagship would be providing double the GPU performance of an RTX 5090.... I just don't see how any of the math adds up.

And the founders edition has been shown to peak over 850w draw when it spikes.
 
Lakados said:
The performance increase from TSMCs N4 processes to their N3 processes varies depending on the flavor chosen, but given the density of GPU's you are getting at best 10%.

An RTX 5090 pulls 800w, so lets assume moving to 3nm yielded the best case scenario, which yields a 10% drop in power consumption with a full 5% performance increase per watt.
You are still looking at a GPU trying to draw upwards of 1400 watts. That is an absolute monster.
Click to expand...
Will this be on N2 or N3? Is there any info on that yet?
 
Decko said:
That's a disappointment, considering they likely won't release until the end of next year imo.
Click to expand...
the original road map for both was end of this year

but you are right. there is likely a delay of 1 year now
 
Marees said:
RDNA 5 on N3

likely same for nvidia 6xxx series also
Click to expand...
N2 is way too expensive for anything consumer right now, TSMC is cranking their prices across the board.
Between Nvidia and Apple, the two of them have N2 locked down, but Nvidia's N2 offerings will be for their Datacenter AI parts, not likely their consumer ones.
 
Decko said:
That's a disappointment, considering they likely won't release until the end of next year imo.
Click to expand...
2027 era N3P would be quite denser than the original N3 of 2023 too, a bit like original N5 is quite different from the latest version (N4P/N4X used by nvidia-amd)

Lakados said:
N2 is way too expensive for anything consumer right now, TSMC is cranking their prices across the board.
Click to expand...
The TSMC 2 Iphone 18 pro if they launch in september are being made right now no ?

Consumer pay giant price for small chips that make it work. Not sure Nviida and apple has locked N2, AMD Zen6 CCD have talk to be on N2p ?
https://www.techpowerup.com/344375/...v-cache-could-pack-up-to-288-mb-of-cache?cp=3

which would show up in production for early 2027 consumer product soon enough (assuming zen6 in 2026 is only Epyc side of things)
 
LukeTbk said:
2027 era N3P would be quite denser than the original N3 of 2023 too, a bit like original N5 is quite different from the latest version (N4P/N4X used by nvidia-amd)


The TSMC 2 Iphone 18 pro if they launch in september are being made right now no ?

Consumer pay giant price for small chips that make it work. Not sure Nviida and apple has locked N2, AMD Zen6 CCD have talk to be on N2p ?
https://www.techpowerup.com/344375/...v-cache-could-pack-up-to-288-mb-of-cache?cp=3

which would show up in production for early 2027 consumer product soon enough (assuming zen6 in 2026 is only Epyc side of things)
Click to expand...
That's good to know, if yields are solid and we get a traditional generational uplift for once, that'd be nice. 9070 XT considering it's 64 CUs is a big upgrade from the 7800 XT. I doubt we'll see that kind of upgrade again, but hope springs eternal.
 
LukeTbk said:
2027 era N3P would be quite denser than the original N3 of 2023 too, a bit like original N5 is quite different from the latest version (N4P/N4X used by nvidia-amd)


The TSMC 2 Iphone 18 pro if they launch in september are being made right now no ?

Consumer pay giant price for small chips that make it work. Not sure Nviida and apple has locked N2, AMD Zen6 CCD have talk to be on N2p ?
https://www.techpowerup.com/344375/...v-cache-could-pack-up-to-288-mb-of-cache?cp=3

which would show up in production for early 2027 consumer product soon enough (assuming zen6 in 2026 is only Epyc side of things)
Click to expand...
TSMC had Nvidia scheduled for high-volume manufacturing on N2 for January, for the "Angstrom Era" chips.
Which will start coming out in the second half of 2026 and into 2027, where they will be transitioning to a split TSMC / Intel manufacturing process for whatever supposedly comes next.

Apple had its production run on N2 in Q4, 2025. So they likely already have all their chips for the phones.
 
Last edited:
Lakados said:
TSMC had Nvidia scheduled for high-volume manufacturing on N2 for January, for the "Angstrom Era" chips.
Which will start coming out in the second half of 2026 and into 2027, where they will be transitioning to a split TSMC / Intel manufacturing process for whatever supposedly comes next.

Apple had its production run on N2 in Q4, 2025. So they likely already have all their chips for the phones.
Click to expand...
LukeTbk said:
Not sure Nviida and apple has locked N2, AMD Zen6 CCD have talk to be on N2p ?
Click to expand...

AMD & Apple have N2 locked

nvidia is skipping to the next node A16 because of that, I believe

(AMD using N2 for CPUs but N3 for GPUs — except MI500 )
 
Based on the specifications, the RDNA 5 series is likely to encompass a full range:
  1. a flagship with 96 CUs,
  2. a mid-range model with 40 CUs,
  3. a mainstream one with 24 CUs, and
  4. an entry version with 12 CUs.
Bus widths decrease alongside memory capacities, creating a typical tiered approach.


The pivotal question remains whether the flagship tier chip will actually be released in a gaming configuration.

Should the AT0 debut as a gaming card, its significance would primarily lie in testing the architectural limits, not in direct sales. The core's ability to sustain an acceptable power density, frequency range, and cooling requirements will determine whether it stands as a noteworthy symbol or integral part of a sustainable product lineup.

The pricing of AT0, if released as a gaming chip, will be crucial. A previously mentioned price of over $2000 is not consistent with AMD's recent pricing strategies. Venturing into this range directly challenges the RTX 5090 tier, as well as involves higher wafer costs and reduced shipping flexibility. Large chips are more susceptible to yield variations in advanced manufacturing processes, where minor defects can significantly affect the number of sellable units.

The chip's design echoes that of the Radeon VII, the Vega 7nm chip initially aimed at the HPC market but later adapted as a gaming card released in limited numbers. Typically, large chips enter the consumer sector not because of gaming demand but due to pre-existing wafer inventory, yield windows, or product timing needing a channel. If AT0 follows a similar trajectory, its justification is more rooted in the chip's development course rather than sheer market demand.

https://www.topcpu.net/en/news/amd-rdna-5-new-graphics-card-with-up-to-96-cus

Marees said:

AMD’s RDNA 5 “AT0”: Enthusiast GPU with the handbrake on ???​

27. February 2026 06:00
Samir Bashir

Realistic assessment​

  • Rumor status: Plausible, but unconfirmed
  • Market strategy: Image cultivation instead of mass market
  • Competitive pressure: High, especially in the >$1,500 segment
  • Timeframe: Realistic in 2027 at the earliest
If AMD launches AT0, it will likely be as a prestige project with limited quantities. If not, RDNA 5 will remain positioned in the high-end performance segment, with prices between $700 and $1,000 as the likely focus.

In short: AT0 is currently more of a strategic option than a finished product. And as long as AMD does not officially confirm anything, it remains just that, a rumor with medium plausibility.

https://www.igorslab.de/en/amds-rdna-5-at0-enthusiast-gpu-with-the-handbrake-on/
Click to expand...
 
there should definitely be a professional version of this chip as explicit demand is already there

https://x.com/__tinygrad__/status/2...its-with-memory-on-a-512-bit-bus.18926/page-3


Everything is lining up for next years RDNA5 launch. We need @AMD to make a big chip with a 512-bit wide GDDR7 bus and no crazy markup on the 96 GB edition. Then we should be able to offer a reasonably priced 200 tok/s inference machine for the biggest models.​
 

TinyCorp pressures AMD to build a 96GB GPU that doesn’t exist yet​

TinyCorp Is Challenging AMD to Build a 96GB RDNA 5 GPU, But the Memory Crisis Makes It Nearly Impossible
March 9, 2026

The company wants to raise $20 million at a $200 million valuation, buy an $11.5 million building in Oregon with 5 megawatts of power capacity, pre-order 3,000 units of an AMD RDNA 5 GPU with 96GB of VRAM at around $2,500 each,
Hotz tagged AMD CEO Lisa Su directly and wrote:
“RDNA 5 should be GDDR7 with a 512-bit bus. Get the 3GB modules and double side it like the Blackwell. We’ll build the board if they don’t.”​


A 96GB GPU that only NVIDIA makes, for $8,000 to $10,000​

Right now, the only GPU on the market with 96GB of VRAM is NVIDIA’s RTX PRO 6000 Ada Blackwell, which retails between $8,000 and $10,000 per unit. That is a workstation card aimed at professionals and large enterprises, not something a small AI lab or independent developer can easily budget for. For context, TinyCorp is asking for that same memory capacity at less than a third of the price.

(and we all know AMD will sell for nv - $500 dollars. tiny corp is asking for 1/3rd of that price)


TinyCorp and AMD have a long and bumpy history​

TinyCorp’s frustration with AMD isn’t new, and neither is their commitment to working with Team Red anyway. The original TinyBox Red ships with six Radeon RX 7900 XTX cards, built on the belief that AMD’s price-to-performance ratio makes them the smarter pick for accessible AI computing compared to NVIDIA’s more expensive hardware.

But the relationship has had real friction. AMD’s ROCm software stack caused enough problems that TinyCorp eventually decided to build their own solution from scratch. They developed a completely independent software stack for AMD GPUs, their own driver, runtime, libraries, and emulator, all in roughly 12,000 lines of code, bypassing ROCm entirely. That is a serious engineering investment, and it underlines just how committed they are to making AMD hardware work, even when AMD doesn’t make that easy.


https://geekrealmhub.com/tinycorp-amd-96gb-rdna5-gpu/
 
Marees said:

TinyCorp pressures AMD to build a 96GB GPU that doesn’t exist yet​

TinyCorp Is Challenging AMD to Build a 96GB RDNA 5 GPU, But the Memory Crisis Makes It Nearly Impossible
March 9, 2026

The company wants to raise $20 million at a $200 million valuation, buy an $11.5 million building in Oregon with 5 megawatts of power capacity, pre-order 3,000 units of an AMD RDNA 5 GPU with 96GB of VRAM at around $2,500 each,
Hotz tagged AMD CEO Lisa Su directly and wrote:
“RDNA 5 should be GDDR7 with a 512-bit bus. Get the 3GB modules and double side it like the Blackwell. We’ll build the board if they don’t.”​


A 96GB GPU that only NVIDIA makes, for $8,000 to $10,000​

Right now, the only GPU on the market with 96GB of VRAM is NVIDIA’s RTX PRO 6000 Ada Blackwell, which retails between $8,000 and $10,000 per unit. That is a workstation card aimed at professionals and large enterprises, not something a small AI lab or independent developer can easily budget for. For context, TinyCorp is asking for that same memory capacity at less than a third of the price.

(and we all know AMD will sell for nv - $500 dollars. tiny corp is asking for 1/3rd of that price)


TinyCorp and AMD have a long and bumpy history​

TinyCorp’s frustration with AMD isn’t new, and neither is their commitment to working with Team Red anyway. The original TinyBox Red ships with six Radeon RX 7900 XTX cards, built on the belief that AMD’s price-to-performance ratio makes them the smarter pick for accessible AI computing compared to NVIDIA’s more expensive hardware.

But the relationship has had real friction. AMD’s ROCm software stack caused enough problems that TinyCorp eventually decided to build their own solution from scratch. They developed a completely independent software stack for AMD GPUs, their own driver, runtime, libraries, and emulator, all in roughly 12,000 lines of code, bypassing ROCm entirely. That is a serious engineering investment, and it underlines just how committed they are to making AMD hardware work, even when AMD doesn’t make that easy.


https://geekrealmhub.com/tinycorp-amd-96gb-rdna5-gpu/
Click to expand...
This also underlines that much of AMD’s user base is simply a “Not Nvidia” user base.
But sadly AMD’s ability to price their hardware is directly dependent on costs, and costs are skyrocketing.
They could take a smaller margin, but their investors would slaughter them over it because even at high prices products are flying off the shelves.

AMD’s investors are their customers, the rest of us are just their users. Sadly a very big difference.
 
You must log in or register to reply here.
Back
Top