Discussion RDNA 5 / UDNA (CDNA Next) speculation

[adroc_thurston]

With Palatnir, and who knows what else, income is guaranteed even if LLMs for general public stagnate.

Technogulag stuff is a tiny income stream.

 

[soresu]

Good deal for AMD, that should put billions into improving stupidly named ROCm framework, perhaps CUDA's moat will finally be bridged...

More stupidly named than CUDA which is just Barracuda shortened and thus ripping off Seagate's branding..... 🤦‍♂️

 

[basix]

They will know what MI500 will look like.

That is what I were implying. OpenAI does it like this because they expect market capitalization to increase. And this only works if future roadmap products like MI500 are looking good as well.

 

[marees]

A little OT, but what is stopping them from lending compute power, similarly to AWS?
With Palatnir, and who knows what else, income is guaranteed even if LLMs for general public stagnate.
Line go up ...

This is a valid Q

What is happening is this:

1. Hyperscalers (& Meta) are in an arms race
2. Google, Microsoft, Amazon, & Meta will mostly fab custom hardware (like they did with ARM CPUs)
3. This leaves open AI as the sole remaining "Hyperscaler" for whose order nv, Broadcom, AMD, (+ MS & google) are competing
4. Stakes very high for nv & amd as they need open AI to tailor software to CUDA / ROCm (or create a new standard like deepseek is doing with Chinese hardware companies)

 

[adroc_thurston]

Google, Microsoft, Amazon, & Meta will mostly fab custom hardware (like they did with ARM CPUs)

Only TPUs are worth mentioning.

 

[Joe NYC]

I dunno, AMD's stock is up 25% premarket. It's like they do this because they know AMD stock will skyrocket. OpenAI buys a billion dollars worth of AMD GPUs, gets stock worth 25 billion for free.

I think, more like $10s of billions worth of AMD products (CPU, GPU, Networking) up to $100 billion.

 

[Joe NYC]

They will know what MI500 will look like.

Good deal for AMD, that should put billions into improving stupidly named ROCm framework, perhaps CUDA's moat will finally be bridged...

It is interesting that AMD is attacking the CUDA moat with AI.

On a recent appearance, Lisa said AMD is using AI to translate code that was targeted for CUDA to work with AMD software stack.

 

[adroc_thurston]

It is interesting that AMD is attacking the CUDA moat with AI.

Nvidia's moat (horrible no good word) is in the hardware stack.

 

[Joe NYC]

Nvidia's moat (horrible no good word) is in the hardware stack.

That's the new moat, which AMD is also attacking (with hardware and industry consortiums).

 

[adroc_thurston]

That's the new moat, which AMD is also attacking (with hardware and industry consortiums).

It's the only one.

 

[Joe NYC]

It's the only one.

I think another moat was perception (true in the past) that NVidia was the only game in town.

From that POV, the OpenAI deal with AMD is a game changer, elevating AMD from being a bottom feeder, with low single digit market penetration in datacenter GPU to well into double digits.

Maybe not too much unlike client GPU, where AMD is underrepresented, but the opportunity to increase market share is very feasible

OpenAI validated AMD Mi450, and now other can also follow.

 

[inquiss]

I think another moat was perception (true in the past) that NVidia was the only game in town.

From that POV, the OpenAI deal with AMD is a game changer, elevating AMD from being a bottom feeder, with low single digit market penetration in datacenter GPU to well into double digits.

Maybe not too much unlike client GPU, where AMD is underrepresented, but the opportunity to increase market share is very feasible

OpenAI validated AMD Mi450, and now other can also follow.

They validated much more than MI450. They validated the things that come after MI450..

 

[Joe NYC]

They validated much more than MI450. They validated the things that come after MI450..

Yup, it is said to be multi-generational deal.

 

[fastandfurious6]

before the AI bubble bursts.

Won't happen... AI is here to stay

What will burst is meaningless investment into meaningless "startups"

Then the only bubble burst re hardware will be when performance requirements plateau

 

[RnR_au]

1. Stakes very high for nv & amd as they need open AI to tailor software to CUDA / ROCm (or create a new standard like deepseek is doing with Chinese hardware companies)

FWIW OpenAI uses the Triton language for their AI efforts. They don't use CUDA.

 

[RnR_au]

Won't happen... AI is here to stay

What will burst is meaningless investment into meaningless "startups"

Then the only bubble burst re hardware will be when performance requirements plateau

Yeah... its like the dot.com bubble... just because it burst didn't make websites and online stores go away.

 

[adroc_thurston]

FWIW OpenAI uses the Triton language for their AI efforts. They don't use CUDA.

Triton has to run on top of stuff you know

 

[RnR_au]

Triton has to run on top of stuff you know

yeah, but it can be any stuff.

 

[adroc_thurston]

but it can be any stuff.

Nope, the IHV still has to do a ton of s/w enablement to make Triton viable on their hardware.
Stuff ain't free.

 

[RnR_au]

Nope, the IHV still has to do a ton of s/w enablement to make Triton viable on their hardware.
Stuff ain't free.

Not sure what you are arguing here. Its clear given the OpenAI/AMD announcements that Triton will compile down to AMD hardware quite well.

 

[adroc_thurston]

Its clear given the OpenAI/AMD announcements that Triton will compile down to AMD hardware quite well.

Yeah but AMD still has to do a lot of work to make it seamless.

 

[RnR_au]

Yeah but AMD still has to do a lot of work to make it seamless.

I'm sure that AMD will be happy to provide engineers for such an effort. I'm not entirely sure what Triton compiles down to for Nvidia hardware. It could be CUDA. It could be PTX. Not sure what AMD provides in terms of language intrinsics that maps similarly to PTX.

 

[adroc_thurston]

Not sure what AMD provides in terms of language intrinsics that maps similarly to PTX.

amdgcn-flavored SPIR-V (amdgcnspirv), or goto amdgcn directly.