Question AMD 2Q25 Earnings

[adroc_thurston]

Why is Mi400 the thing that puts it all together for AMD?

training

 

[mikegg]

training

Stop with the low quality posts hyping AMD. Give some effort.

 

[adroc_thurston]

Stop with the low quality posts hyping AMD.

They kinda hype themselves.
They're the DC guys, after all.

 

[Joe NYC]

Why is Mi400 the thing that puts it all together for AMD?

NVidia has a networking system that lets as many as 72 GPUs share workload and share memory. This is important mainly in training but also may come more and more into play in complex inference tasks over time.

AMD can only share 8 GPUs in the same system, currently.

AMD is developing industry standard networking standard called UA Link to compete with NVidia proprietary NVLink - in order to similarly connect the GPUs. These UA Link products will be on the market at the same time frame as Mi400, which is when AMD will be able to also have 72 GPUs working in tandem and a year later, 144 GPUs.

 

[adroc_thurston]

NVidia has a networking system that lets as many as 72 GPUs share workload and share memory

There are a ton more caveats to that.

AMD is developing industry standard networking standard called UA Link

A stretch, UAL is pretty much IFIS with some parts chopped off.

and a year later, 144 GPUs.

A year later they go to 256 because that's real podracing.

 

[Joe NYC]

There are a ton more caveats to that.

A stretch, UAL is pretty much IFIS with some parts chopped off.

A year later they go to 256 because that's real podracing.

I think I read the limit for UA Link is 1024.

I don't know if this was just Dylan's hallucination, or if Mi400 will be only 4 GPUs per node, rather than 8. So, to get to 256, they will need 64 nodes.

Since they are going to be using double wide rack, maybe they are limiting themselves to 144 because there are only so many nodes they can fit in.

 

[adroc_thurston]

I think I read the limit for UA Link is 1024.

The limit is whatever your switching Si and topology allows for.

I don't know if this was just Dylan's hallucination, or if Mi400 will be only 4 GPUs per node

Helios is 1:4, it's basically a less dense Cray EX235a for hyperscaler proletariat.

Since they are going to be using double wide rack, maybe they are limiting themselves to 144 because there are only so many nodes they can fit in.

Double-wide racks are for heat density management.
These super-dense wunderwaffen have miserable MTBF that everyone is trying to fix right now.

 

[Joe NYC]

The limit is whatever your switching Si and topology allows for.

Helios is 1:4, it's basically a less dense Cray EX235a for hyperscaler proletariat.

Double-wide racks are for heat density management.
These super-dense wunderwaffen have miserable MTBF that everyone is trying to fix right now.

The goal seems to be to sell the entire rack as a self contained unit. So to go up to 2x GPUs, they came up with double rack, again as a unit. Maybe NVidia will announce something similar soon.

 

[adroc_thurston]

The goal seems to be to sell the entire rack as a self contained unit.

Oh no we're going bigger.

So to go up to 2x GPUs, they came up with double rack, again as a unit

Again, it's a double-wide rack because heat sucks.
It sucks so goddamn much.

Maybe NVidia will announce something similar soon

They already showcased their roadmap.
It looks to be an attempt at violating the laws of physics.

 

[Joe NYC]

Oh no we're going bigger.

Again, it's a double-wide rack because heat sucks.
It sucks so goddamn much.

They already showcased their roadmap.
It looks to be an attempt at violating the laws of physics.

I need to check.

But water cooling is kind of that - it can remove several X the amount of heat and at the same time, can keep the GPUs (and CPUs) cooler, reducing the stress on them, while also clocking them to the max.

Air cooling will be gone in not too distant future at the Hyperscalers...

 

[adroc_thurston]

But water cooling is kind of that - it can remove several X the amount of heat and at the same time, can keep the GPUs (and CPUs) cooler, reducing the stress on them, while also clocking them to the max.

You do understand that water-cooling there is a necessity because they're 1.5kW a pop GPUs (and higher soon) stacked like tunas in a can.

Air cooling will be gone in not too distant future at the Hyperscalers...

Oh no, CPUs will be air-cooled forever.

 

[Kepler_L2]

You do understand that water-cooling there is a necessity because they're 1.5kW a pop GPUs (and higher soon) stacked like tunas in a can.

Oh no, CPUs will be air-cooled forever.

I thought MI400 was like 3.4KW or something.

 

[Abwx]

Their net revenue is higher than what used to be their whole quarterly revenue back in Q1 2016, wich was the lowest since 2000 or so.

Quartalszahlen von AMD: Ryzen und Radeon werden zum stärksten Pferd im Stall

Im globalen KI-Wahn sollten doch Epyc-Prozessoren und Instinct-GPUs am meisten Umsatz bei AMD generieren? Im zweiten Quartal ist das anders.

www.computerbase.de

 

[BorisTheBlade82]

Why is Mi400 the thing that puts it all together for AMD?

First of all, because you can scale it up much better than MI300, which gives you much bigger "world sizes".
As always, SemiAnalysis has a nice Article on this.

AMD Advancing AI: MI350X and MI400 UALoE72, MI500 UAL256

Software Improvement, Marketing RDFs, AMD Fostering Neocloud, MI355 is not Rack Scale, MI400 is UALoE, Not UALink

semianalysis.com

 

[jpiniero]

No, not even close.
But NV will do anything to not let their DC biz turn Xeon.

I would think AMD should be concerned more about Cloud Providers using their own, and not nVidia.

 

[inquiss]

I would think AMD should be concerned more about Cloud Providers using their own, and not nVidia.

Why? AMD can eat into Nvidia's lunch since they have most of the market. If ASICs take some of that pie so be it, not good for AMD but it's better to go after the lions share of the market than waste their time trying to convince cloud providers to use them instead of ASICs..

 

[jpiniero]

Why? AMD can eat into Nvidia's lunch since they have most of the market. If ASICs take some of that pie so be it, not good for AMD but it's better to go after the lions share of the market than waste their time trying to convince cloud providers to use them instead of ASICs..

The Cloud Providers almost have to be most if not almost all of AMD's sales right now.

The rest of the market are those "Startups" that will buy nVidia because of CUDA and/or because buying nVidia makes it looks better for the "Investors" they are trying to scam.

 

[Josh128]

Zen 5 launched hard. Ryzen 9000 clearly moving well, especially with the mid/high-end DIY crowd. Also wild to see Radeon making a comeback, probably fueled by the 7900/7800XT price drops and 9700/9800 leaks.

No doubt helped by insanely bad press on Arrow Lake. People just always expecting a giant breakout in AI like what happened with NV.

Hint: A giant breakout is never going to happen for AMD in AI. It didnt happen in the CPU sector, even though AMD is eating Intels lunch for 5 years now, technically, in enterprise CPU tech and in desktop tech for at least 3 years now, so expecting it against an "firing on all cylinders" Nvidia is a fools errand. If they do gain marketshare, its going to be a slow, steady, incremental gain. Huang will make it difficult, as he really does have the best AI products and is constantly working to improve them.

AMD seems to be doing the right things, but even so, its going to be a slog to gain market and mindshare. By all accounts ROCm is still nowhere near CUDA for AI development, thats what AMD needs to improve the most.

 

[Win2012R2]

It seems that this time, the immediate reaction is misinterpreting the results

It's (AMD stock drops big after beat in earning) is the classic reaction, happens all the time, people now sell because they expect everyone else will, so it drops regardless, anybody who has been watching AMD in recent 5 years knows it.

Why is Mi400 the thing that puts it all together for AMD?

Finally rack based solution - breaking through their current 8 GPUs wall.

Yeah Rockstar delayed it but Sony might have scheduled a big order a long time ago in anticipation.

Sony got marketing deal with Rockstar re GTA6, they get all the info they need.

 

[desrever]

Are we residing in the same universe?

This was an excellent quarter, and old Mi300/Mi325, which were at lower profit margins, did not sell very well, are finally exiting the picture.

There is nothing AMD could do about the (temporary) ban of Mi308 GPUs to China, and all of the consequences were absorbed in Q2, and even despite that, AMD had the highest revenue quarter in history

I mean its good in the sense that there is growth but comparatively to other semis (excluding intel), the growth is just fine. It looks like AMD is still bleeding market share to Nvidia in AI space even tho they only hold single digit market share and there isn't any major guide up even in Q3 which means ramp of MI350/355 is just business as usual and not gaining any real momentum. With how hungry everyone is for AI servers, AMD's slow and steady approach is losing them significant potential business. Waiting game is now shifting to MI400 but who knows how that will really play out.

Sure AMD is doing well in a vacuum or compared to Intel but broad competition landscape is still kind of under whelming. The MI308 stuff is a bummer but considering how little expectations Lisa had for it in the conference and even if granted an license, it would only really start materializing in Q4 means its very much not the only issue.

 

[yuri69]

Why is Mi400 the thing that puts it all together for AMD?

Simply because it is the AMD's +1 gen "GPU". The big thing is always the one on the horizon.

 

[adroc_thurston]

It didnt happen in the CPU sector

It happened right past these two quarters with AMD finally going to town in commercial.
And obviously they coup de grace'd whatever remains of Intel DIY share.

Simply because it is the AMD's +1 gen "GPU".

No, because it's a rack that actually works.

 

[jpiniero]

Sony said in their last financial call that the forecast was for 15 million PS5's for their entire FY, but also "wasn't married" to it (because of tariffs). Last year was 18.5 million. That's not surprising for a decline, just that AMD getting a huge increase in revenue is.

Sony got marketing deal with Rockstar re GTA6, they get all the info they need.

They made the order before the delay was announced?

 

[Joe NYC]

I mean its good in the sense that there is growth but comparatively to other semis (excluding intel), the growth is just fine. It looks like AMD is still bleeding market share to Nvidia in AI space even tho they only hold single digit market share

AMD started at 0% market share. Hitting a ceiling in low single digits market share - I would not call bleeding, just being stuck at that level for last ~6 months.

and there isn't any major guide up even in Q3 which means ramp of MI350/355 is just business as usual and not gaining any real momentum.

+ $1 billion increase in guidance, majority of which will be Mi355 is a major guide up.

Wells Fargo analyst estimated ~ $900m range shipments in Q2. Som increase of $550m+, QoQ is quite major

With how hungry everyone is for AI servers, AMD's slow and steady approach is losing them significant potential business. Waiting game is now shifting to MI400 but who knows how that will really play out.

NVidia Blackwell has been shipping (arguable) since Q4 2024. It left Mi300 behind even in markets where AMD was competitive.

Mi355 is very competitive with Blackwell now, in 1-8 GPU deployments (mainly for inference)

So, the ceiling to AMD can grow was raised quite a bit. To say 10%. The actual market share can resume some growth to mid single digits (from low single digits).

Sure AMD is doing well in a vacuum or compared to Intel but broad competition landscape is still kind of under whelming. The MI308 stuff is a bummer but considering how little expectations Lisa had for it in the conference and even if granted an license, it would only really start materializing in Q4 means its very much not the only issue.

The licenses were not granted yet, as of the conference call. So, no point to raise expectations. But the $800m writedown of part in process can be as much as $2 billion in revenue when completed and sold. So, not irrelevant revenue compared to current datacenter GPU revenue. Even split over 3 quarters, it can be a significant upside.

Also, the Mi308 platform also has AMD EPYC CPUs to go along with the GPUs, so some increment from that too.

 

[Josh128]

It happened right past these two quarters with AMD finally going to town in commercial.
And obviously they coup de grace'd whatever remains of Intel DIY share.

I mean crazy, parabolic stock breakout NV style. It looked like it might finally do so from Oct '23 to Feb '24, but gave back 100% of those gains by a week or so after Trumps "Liberation Day" 2025. Now its just got its bubble popped again. NV OTOH, has 2.5x'ed its valuation in the same time period and is still touching all time highs right now.

Its just not going to happen unless AMD can not just merely compete well with NV, but chew them up, swallow them, and sh!t them out from a technical standpoint. I just dont see that happening. Slow and steady gains are still better than 3rd place...