Discussion RDNA 5 / UDNA (CDNA Next) speculation

[1250]

How much do you expect AMD to buy each hbm4?
For reference, NVIDIA costs ₩1000000(approximately 700 dollars).

lower clock↓+ cbdie↑($30?) + sole vendor↓= 730 - × (🙏200?) = ₩800000×12😭

 

[Win2012R2]

How much do you expect AMD to buy each hbm4?

They don't need as high freq, so will get rejects = considerably lower, though reports surfaced that Nvidia will accept two tier HMB, no doubt to limit how many AMD can get, plus maybe Micron's stuff is good enough

 

[adroc_thurston]

They don't need as high freq, so will get rejects = considerably lower, plus maybe Micron's stuff is good enough

Micron stuff can't be "good enough" since AMD is doing kustom HBM4.

 

[Win2012R2]

Micron stuff can't be "good enough" since AMD is doing kustom HBM4.

In which case it ain't good enough, but will take off some demand from ASIC makers who might not need custom stuff

 

[adroc_thurston]

In which case it ain't good enough, but will take off some demand from ASIC makers who might not need custom stuff

ASIC slaves are getting HBM4 like late next year.
Irrelevant for MI455x ramp.

 

[1250]

Do I need to know more about cdna 5?
하산해도 되겠습니까?

 

[adroc_thurston]

Do I need to know more about cdna 5?

no.
Just wait until launch it's not even that far away.

 

[1250]

no.

not yet.
Next week is Helix(gdc), and the week after that, Nvidia(gtc).
I'll focus on 심상수련(Imagery Training).

 

[marees]

This AI Startup Demands AMD to Build a 96 GB RDNA 5 GPU for a Wild Venture, and Is Already Seeking Investors​

Muhammad Zuhair
Mar 8, 2026 at 03:53pm EDT

This AI Startup Demands AMD to Build a 96 GB RDNA 5 GPU for a Wild Venture, and Is Already Seeking Investors

The demand for compute power has gone wild out there, and in the midst of it, the AI startup TinyCorp has a rather interesting demand for AMD.

wccftech.com

TinyCorp has a rather interesting demand for AMD, involving a 'chonky' RDNA 5 GPU.

TinyCorp Hopes To Get an RDNA 5 GPU With 96 GB VRAM For $2,500 Each; Probably Ignoring the Memory Shortages​

(I asked grok & it said an AT0 with 96gb to 128gb ram will cost $4000 to $8000)

Interestingly, TinyCorp claims it will build its own board with RDNA 5 silicon onboard, provided AMD doesn't release a similar model.

https://twitter.com/x/status/2030311966066266340

 

[gdansk]

And I demand AMD sell me a RX 9070 XT for $200 but I don't get an article.

 

[coercitiv]

I think 96GB of VRAM for $2500 is doable, but then we'll have to see the price for the rest of the card.

 

[adroc_thurston]

I think 96GB of VRAM for $2500 is doable, but then we'll have to see the price for the rest of the card.

That's a proviz(-class) board with appropriate margins.
$5k at least.

 

[Makaveli]

That's a proviz(-class) board with appropriate margins.
$5k at least.

This no way they are selling these if they make them for 1/4 the price NV charges for a 96GB GPU.

will be in the 5-7k range.

 

[jpiniero]

I think 96GB of VRAM for $2500 is doable, but then we'll have to see the price for the rest of the card.

The RAM alone would be like $1000 or more currently, lol.

 

[soresu]

That's a proviz(-class) board with appropriate margins.
$5k at least.

Isn't that pointless when most modern renderers have some kind of "out of bounds" memory option or something that allows it to use the system memory instead when the VRAM fills up?

 

[rntongo]

Give me a $6-$8k 96GB GDDR7 RDNA 5/UDNA card with 4 way interconnect capability and I will buy those over any RTX Pro 6000. That would be a real game changer and also both economical and technically feasible

 

[ToTTenTranz]

The RAM alone would be like $1000 or more currently, lol.

I think @coercitiv was making a joke about $2500 being the cost of the 96GB VRAM alone.

This no way they are selling these if they make them for 1/4 the price NV charges for a 96GB GPU.

will be in the 5-7k range.

Pricing Pro cards for ML loads using the same "Nvidia -10%" method as they use in consumer GPUs would result in no one at all buying their cards. The consequences for inadequate GPUs in professional settings are completely different from a guy not being able to use the best upscaler in the latest videogame.

Give me a $6-$8k 96GB GDDR7 RDNA 5/UDNA card with 4 way interconnect capability and I will buy those over any RTX Pro 6000. That would be a real game changer and also both economical and technically feasible

As someone who's procured Pro GPUs for ML loads throughout the past 5 years, this is unfortunately way too risky.

Unless you know exactly what the loads you're going to be running and you're absolutely sure AMD GPUs can run them (not our case), odds are you'll run into a bunch of loads AMD GPUs just can't run properly if at all. This means tens / hundreds of euros/dollars sitting idle or worse, spending that or more in precious engineering resources to make things work.

ROCm may work fine for a select number of things, but every brand new ML state-of-the-art app or implementation comes out the window with full Nvidia support, and AMD hardware may or may not work some months/years afterwards.


For example, these tiny corp guys are saying they'll use AT0 only to run llm inference of the latest deepseek and qwen models like llama-cpp and then sell cloud services using that hardware. That's fine, but if a big new thing comes up that is built from the ground up on CUDA / CUDNN (which is very likely).

 

[dangerman1337]

That leaked MLID slide that 128GB AT0 SKU so they'll probably load pro ones with 4GB Modules.

 

[rntongo]

As someone who's procured Pro GPUs for ML loads throughout the past 5 years, this is unfortunately way too risky.

Unless you know exactly what the loads you're going to be running and you're absolutely sure AMD GPUs can run them (not our case), odds are you'll run into a bunch of loads AMD GPUs just can't run properly if at all. This means tens / hundreds of euros/dollars sitting idle or worse, spending that or more in precious engineering resources to make things work.

ROCm may work fine for a select number of things, but every brand new ML state-of-the-art app or implementation comes out the window with full Nvidia support, and AMD hardware may or may not work some months/years afterwards.


For example, these tiny corp guys are saying they'll use AT0 only to run llm inference of the latest deepseek and qwen models like llama-cpp and then sell cloud services using that hardware. That's fine, but if a big new thing comes up that is built from the ground up on CUDA / CUDNN (which is very likely).

You're right that Nvidia has better software support and every new ML implementation comes with full Nvidia support... and thats exactly why I said AMD needs to create workstation competitors to Nvidia like an RTX Pro 6000 competitor. Nvidia is being used by most professionals for their prototyping systems and its a really decent deal. But the market is waiting for such products from AMD to support but they havent had a proper productline. CUDA moat is on shaky ground. At my workplace we use mainly Nvidia GPUs for prototyping and self hosting part of our services(I would say 90%). Tested AMD R9700 and its actually really good especially with Vulkan(ROCm has good potential but again AMD just needs to give us the kind of cards we need and we'll sort out the software),but it just doesnt meet our minspec requirements by some small margins because AMD didnt make any Halo cards this time otherwise an AMD "R9800" would have been a huge game changer if it existed.

With AI driven development even being used for kernel level optimizations, and the growing move towards optimizing GPUs and other ASICs for ML workloads, AMD could seriously increase its market share in that core workstation and entrylevel datacenter cards. which is key to getting the same level of developer support that Nvidia gets with CUDA that you mentioned. The issue now is what kind of product lineup AMD makes with RDNA 5/UDNA. They are doing a good job with their rackscale systems like Helios and working with META but its when they get more market share in the workstation area that they will actually seriously dent the CUDA moat. Make that RTX pro 6000 competitor even if its just 30% more powerful and give me 96GB of VRAM with 4 way interconnect and I will buy at least 8 of them even more over Nvidia. Unless Nvidia puts Nvlink back in its workstation cards, otherwise AMD will have a serious opportunity.

 

[1250]

Nah mabye fake ㅠㅠ

 

[marees]

https://twitter.com/x/status/2030895608262013134

4gb os?

more likely dev kit has 32gb
but real console will have only 24gb. that doesn’t require clamshell memory

 

[1250]

Or 4×3*2 + 2×2*2. 8gb(4×1*2) cpu

 

[marees]

Or 3x8 + 2x4, 8gb of which could be dedicated to CPU
No 192bit?

with 192bit the max you can do is 24gb

that 8gb must be a separate piece of hardware outside of that 192

 

[basix]

Obviusly you could do more with 192bit. Just use clamshell mode. PS6 is rumored to feature 30 GByte on 160bit, which is 2*5*3 GByte. 36 GByte would be the result if Microsoft opts for such a solution.

 

[1250]

What about tdp of mi455x
Is the image far from reality? It looks lower than 2000.