AMD MI300 is the fastest ramping product in its history

M

Marees

2[H]4U
Joined
Sep 28, 2018
Messages
2,161
Lisa Su: Data Center GPU Sales To Hit $4B, Teases Future Chips

With AMD looking to challenge Nvidia’s dominance in the AI computing space, Su said expanded customer engagements with the company’s Instinct MI300 accelerator chips prompted it to upgrade its 2024 data center GPU revenue forecast to $4 billion.
This represents a $500 million bump from the data center GPU guidance the company offered in January. In the month prior, AMD marked the official launch of the MI300 chips as a challenge to Nvidia’s powerful and popular H100 GPU.

Su said more than 100 enterprise and AI customers are “actively developing or deploying” the MI300X, the GPU-only version of the processor. (The other version is the MI300A, which combines GPU cores and CPU cores onto a single package.)

What’s helping with demand are the multiple MI300X systems entering volume production this quarter with OEMs such as Dell Technologies, Hewlett Packard Enterprise, Lenovo, Supermicro and others, according to Su. Cloud service providers Microsoft and Oracle have also expanded their MI300X production environments, as did Facebook parent company Meta.

“What we see now is just greater visibility to both current customers as well as new customers committing to MI300,” she said.

This increased interest helped the MI300 become AMD’s “fastest-ramping product” in its history, generating more than $1 billion in “total sales in less than two quarters.”

Su suggested that AMD could sell even more MI300 products if it wasn’t for supply constraints that were holding the company back, which she considered an industry-wide issue. The company expects MI300 supply to “improve every quarter this year.”

“We are tight on supply, so there’s no questions in the near term that if we had more supply, we have demand for that product, and we're going to continue to work on those elements as we go through the year,” Su said, adding that she’s “very pleased with how the ramp is going.”

https://www.crn.com/news/components...g-product-teases-new-ai-chips-later-this-year
 
Marees said:
Lisa Su: Data Center GPU Sales To Hit $4B, Teases Future Chips

With AMD looking to challenge Nvidia’s dominance in the AI computing space, Su said expanded customer engagements with the company’s Instinct MI300 accelerator chips prompted it to upgrade its 2024 data center GPU revenue forecast to $4 billion.
This represents a $500 million bump from the data center GPU guidance the company offered in January. In the month prior, AMD marked the official launch of the MI300 chips as a challenge to Nvidia’s powerful and popular H100 GPU.

Su said more than 100 enterprise and AI customers are “actively developing or deploying” the MI300X, the GPU-only version of the processor. (The other version is the MI300A, which combines GPU cores and CPU cores onto a single package.)

What’s helping with demand are the multiple MI300X systems entering volume production this quarter with OEMs such as Dell Technologies, Hewlett Packard Enterprise, Lenovo, Supermicro and others, according to Su. Cloud service providers Microsoft and Oracle have also expanded their MI300X production environments, as did Facebook parent company Meta.

“What we see now is just greater visibility to both current customers as well as new customers committing to MI300,” she said.

This increased interest helped the MI300 become AMD’s “fastest-ramping product” in its history, generating more than $1 billion in “total sales in less than two quarters.”

Su suggested that AMD could sell even more MI300 products if it wasn’t for supply constraints that were holding the company back, which she considered an industry-wide issue. The company expects MI300 supply to “improve every quarter this year.”

“We are tight on supply, so there’s no questions in the near term that if we had more supply, we have demand for that product, and we're going to continue to work on those elements as we go through the year,” Su said, adding that she’s “very pleased with how the ramp is going.”

https://www.crn.com/news/components...g-product-teases-new-ai-chips-later-this-year
Click to expand...

Gaming revenue, on the other hand, declined sequentially by 32.6 percent and 47.5 percent year-over-year to $922 million, due to lower demand for PC GPUs. AMD’s CFO, Jean Hu, said the company doesn’t expect the situation to improve this year.

“Based on the visibility we have, the first half […] we guided down sequentially more than 30 percent. We actually think the second half will be lower than the first half,” she said.
 
Marees said:
Gaming revenue, on the other hand, declined sequentially by 32.6 percent and 47.5 percent year-over-year to $922 million, due to lower demand for PC GPUs. AMD’s CFO, Jean Hu, said the company doesn’t expect the situation to improve this year.

“Based on the visibility we have, the first half […] we guided down sequentially more than 30 percent. We actually think the second half will be lower than the first half,” she said.
Click to expand...
Also console sales are down (probably because new refreshes are due: 6nm discless xbox & PS5 pro)

Decision time for Radeon & Lisa Su
Do they hold tight to precious margins or price the GPUs to move 🤔
 
Last edited:
Mega6 said:
AMD is "chip production constrained". They will just make less GPUs and more Epyc, MI300's. Dropping the high end RDNA4 also shows their focus. Their margins went up 3% last qtr.
Click to expand...
To be fair (and I hate this) If someone offered you 10x the money for the same work, wouldn't you serve them first? Like Gamers seem like a publicity stunt or charity at the margins we're looking to pay.
 
Marees said:
Also console sales are down (probably because new refreshes are due: 6nm discless xbox & PS5 pro)

Decision time for Radeon & Lisa Su
Do they hold tight to precious margins or price the GPUs to move 🤔
Click to expand...
Or just cut back on GPU’s and only sell small ones. AMD can’t give graphics up because that would kill off their consoles and such which isn’t nothing. But this could snowball for them, they scale back hardware and reduce budgets for drivers while making more stuff open source, which leads to fewer sales. Rinse and repeat.
 
Marees said:
Do they hold tight to precious margins or price the GPUs to move 🤔
Click to expand...
Are these the "precious margins" you speak of?

Screenshot_20240501-140000.png
 
They need to make it so the "AI" stuff (and related, raytracing and other workloads) work on consumer GPUs as well. Trying to push only at the highest end is silly when you can do better , especially taking advantage of open source widespread development to help build a foundation of those with positive experiences with AMD to compete with Nvidia's massive integrated platform. There may be some edge cases, but if people know that all the software are used to tensorflow/CUDA etc.. and not the open alternatives they'll still lean to NV and the mistakes will compound. AMD has the potential to compete and offer better products across the line, consumer and otherwise, for good prices with top performance and technology but they have to get their shit together. I am hoping that the open Linux drivers will continue to be a step forward, but we also need things like fixing problems (ie HDMI 2.1 refusal to support FOSS drivers), expand planned projects like OpenSIL / Coreboot, ensuring that open spec/engine/standards for AI/RT etc..are viable (ie you can use an AMD GPU for AI work on selfhosted stuff like Stable Diffusion but it requires configuration if the default is NV focused backends etc).

Otherwise, they're just going to let NV continue to direct the market, always be looking like they're playing catch up, and even when they actually succeed (I'm thinking of the rasterization gains in the 6000 series allowing them to compete with even the halo project NV gaming cards, in a way that hadn't happened for generations but NV just basically said "All that matters now is raytracing and DLSS. Oh by the way, don't forget CRYPTO MINING! They changed their cards for top rasterized gaming performance and can't do that as well, either!" and everyone just went along with it even at a time when it was a far less usefu and mature tech that did little more than tech demos). I want to see AMD succeed when they're doing things actually good for the market, for users, for openness but they've got to get it together. If they pitch hard towards these AI professional cards and most of the industry shrugs and buys NV anyway, its not going to put them in a better situation especially if this bet means they don't have the money, time, or silicon to devote to other projects that could have come out better.
 
Last edited:
RanceJustice said:
They need to make it so the "AI" stuff (and related, raytracing and other workloads) work on consumer GPUs as well. Trying to push only at the highest end is silly when you can do better , especially taking advantage of open source widespread development to help build a foundation to compete with Nvidia's massive integrated platform. There may be some edge cases, but if people know that all the software are used to tensorflow/CUDA etc.. and not the open alternatives they'll still lean to NV and the mistakes will compound. AMD has the potential to compete and offer better products across the line, consumer and otherwise, for good prices with top performance and technology but they have to get their shit together. I am hoping that the open Linux drivers will continue to be a step forward, but we also need things like fixing problems (ie HDMI 2.1 refusal to support FOSS drivers), expand projects like OpenSIL / Coreboot, ensuring that open spec/engine/standards for AI/RT etc..are viable (ie you can use an AMD GPU for AI work on selfhosted stuff like Stable Diffusion but it requires configuration if the default is NV focused backends etc).

Otherwise, they're just going to let NV continue to direct the market, always be looking like they're playing catch up, and even when they actually succeed (I'm thinking of the rasterization gains in the 6000 series allowing them to compete with even the halo project NV gaming cards, in a way that hadn't happened for generations but NV just basically said "All that matters now is raytracing and DLSS. Oh by the way, don't forget CRYPTO MINING! They changed their cards for top rasterized gaming performance and can't do that as well, either!" and everyone just went along with it even at a time when it was a far less usefu and mature tech that did little more than tech demos). I want to see AMD succeed when they're doing things actually good for the market, for users, for openness but they've got to get it together. If they pitch hard towards these AI professional cards and most of the industry shrugs and buys NV anyway, its not going to put them in a better situation especially if this bet means they don't have the money, time, or silicon to devote to other projects that could have come out better.
Click to expand...
It is funny because over in another thread people argue that ROCm is useless to invest the resources for desktop cards. That said, I agree with a lot of what you say here
 
RanceJustice said:
AMD has the potential to compete and offer better products across the line, consumer and otherwise, for good prices with top performance and technology but they have to get their shit together.
Click to expand...
Could just be the wording but that undersell just how incredibly hard it is to compete with a company like Nvidia at this, AMD is a really well run company with a strong CEO that seem to have their shit together, Nvidia is just incredibly good and what they do with a lot of money to do it.

RanceJustice said:
(I'm thinking of the rasterization gains in the 6000 series allowing them to compete with even the halo project NV gaming cards, in a way that hadn't happened for generations but NV just basically said "All that matters now is raytracing and DLSS.
Click to expand...
I feel like they were pretty much selling everything they could make during a long time, supply chain capacity (when delivering console APUs and everything else) being more what limited how many 6700xt-6800xt and co. were sold in 2021 than people caring a lot of about raytracing.

Nvidia purchasing of supply power can be hard to overcome and AMD not using say GDDR6X or GDDR7, TSMC latest node is not that easy (a bit like Nvidia vs apple in that regard), not saying anything if you plan are good or bad, just that it feel to me extremelly hard to execute and not trying to do something like the MI300 and its successor with how much money at those margin are good into Nvidia pocket is such a giant bold move, for a co. like AMD that is arguably the second best placed company in the world to be a competitor to Nvidia in the open non in-house solution market. If the world most company cannot buy the latest Nvidia tech because there a giant waylist outside the special partner, AMD does not need to be as good or cheaper than Nvidia to be able to sell, a bit like they always sell and get closer to Nvidia in market share during crypto boom, when every cpu that work get sold the brand stop to matter, people cannot choose to buy the brand they prefer anyway.

The very important interconnect to scale those AI chips for example, apparently Nvidia bought pretty much everything and people are trying to make Ethernet based new alternative/solution.
 
Last edited:
RanceJustice said:
They need to make it so the "AI" stuff (and related, raytracing and other workloads) work on consumer GPUs as well. Trying to push only at the highest end is silly when you can do better , especially taking advantage of open source widespread development to help build a foundation of those with positive experiences with AMD to compete with Nvidia's massive integrated platform. There may be some edge cases, but if people know that all the software are used to tensorflow/CUDA etc.. and not the open alternatives they'll still lean to NV and the mistakes will compound. AMD has the potential to compete and offer better products across the line, consumer and otherwise, for good prices with top performance and technology but they have to get their shit together. I am hoping that the open Linux drivers will continue to be a step forward, but we also need things like fixing problems (ie HDMI 2.1 refusal to support FOSS drivers), expand planned projects like OpenSIL / Coreboot, ensuring that open spec/engine/standards for AI/RT etc..are viable (ie you can use an AMD GPU for AI work on selfhosted stuff like Stable Diffusion but it requires configuration if the default is NV focused backends etc).

Otherwise, they're just going to let NV continue to direct the market, always be looking like they're playing catch up, and even when they actually succeed (I'm thinking of the rasterization gains in the 6000 series allowing them to compete with even the halo project NV gaming cards, in a way that hadn't happened for generations but NV just basically said "All that matters now is raytracing and DLSS. Oh by the way, don't forget CRYPTO MINING! They changed their cards for top rasterized gaming performance and can't do that as well, either!" and everyone just went along with it even at a time when it was a far less usefu and mature tech that did little more than tech demos). I want to see AMD succeed when they're doing things actually good for the market, for users, for openness but they've got to get it together. If they pitch hard towards these AI professional cards and most of the industry shrugs and buys NV anyway, its not going to put them in a better situation especially if this bet means they don't have the money, time, or silicon to devote to other projects that could have come out better.
Click to expand...
CUDA is good and it works across the entirety of the Nvidia stack going back to Maxwell, if AMD wants ROCm to be taken seriously they need to do the same. ROCm can do a lot but what's the point in leveraging it for a consumer project if nothing on the consumer side actually supports it. Maybe AMD is worried about what supporting a project of that size looks like but I think it is needed, ROCm could be great but with the current hardware and software support it's OK at best, and as much as AMD wants to make it an Opensource Community problem to fix, this sort of thing needs to happen in house, make it open source sure not going to say it shouldn't be but it needs a lot of first hand AMD eyeballs on there to get it working on existing and future consumer product stacks.
 
FrgMstr said:
Are these the "precious margins" you speak of?

View attachment 651249
Click to expand...
I am glad the margins are the same, given that there was a brutal 50% year on year reduction in gaming revenue

I understand microsoft has stopped xbox chip purchases temporarily (until whenever the 6nm refresh launches I guess)
Sony also could have slowed until PS5 pro launches

That said I hope RDNA 4 launches this year. It should help them get some laptop wins (if that is possible)
Also with 7900 GRE discounting for as low as $510, the rdna 4 (navi 48) card which is size of a 6600 xt ahould give them better margins at $550, I guess

(I sincerely hope they learnt something from RDNA 2 oversupply & produced less of RDNA 3 chips !!!)


A 48% YoY decline for the first quarter of 2024
AMD's gaming business earned $922 million in Q1 2024 — down 48% year-over-year (from $1.757 billion) and down 33% quarter-over-quarter (from $1.368 billion). AMD's gaming segment still posted an operating income of $151 million, but that's a significant decline from $314 million a year ago. As a percentage of revenue, the gaming unit's operating income totaled 16% in Q1 2024 and 18% in Q1 2023.

Gaming revenue may not recover until 2025
While the Radeon RX 7900 GRE is one of the best graphics cards around, this addition is arguably not enough to make the whole lineup significantly more competitive. This is why AMD expects sales of its gaming solutions to decline by around 30% year-over-year in the first half of 2024, and expects even worse results in the second half of the year.

"[The] gaming segment, based on current demand signals, revenue to decline by significant double-digit percentage," said Hu. "[…] Based on the visibility we have, the first half, both Q1 [and] Q2, we guided down sequentially more than 30%. We actually think the second half will be lower than first half. That is basically how we are looking at this year for the gaming business."
Click to expand...

https://www.tomshardware.com/pc-com...sedives-48-not-expected-to-recover-until-2025
 
Well gaming div. margins are down 2% (18->16%). Overall company margins increased 2-3% depending gaap vs non-gaap.. An illustration of the overall company divisions may be more helpful. Embedded (mostly former Xilinx) was overstocked with inventory and gaming (gpu /consoles) consoles are old and are in desperate need of a refresh and we all know no one buys AMD GPUs so gaming crashed ~50% y/y. No surprise there. Client (Ryzen CPUs) had a nice pop as that whole CPU market is recovering (as Kyle said in an earlier post). Data Center was actually very good +80% y/y with the MI300 selling very well.

Results fell exactly in line exactly as expected, wall st was looking for the MI300s to add another $3-4B to datacenter. Well the AI bubble and unrealistic expectations are not reality for most companies (yet) so the stock got punished.

Imagine yourself as CEO,. It's pretty obvious where to pivot your resources and how to play to the company's strengths. I would not worry too much about AMD GPUs though. AI accelerators and GPUs are very much related so I can't see AMD quiting the GPU business anytime soon. It's just too bad they gave up the high end without a fight this cycle, That 8900XTX would have been a nice piece to see.

1714633122990.png
 
Last edited:
I think Nvidia's 12th biggest customer places single orders that are bigger than AMD's entire data center business.

“Based on the visibility we have, the first half […] we guided down sequentially more than 30 percent. We actually think the second half will be lower than the first half,” she said.
Click to expand...

So no RDNA 4 this year. Got it.

Marees said:
Su suggested that AMD could sell even more MI300 products if it wasn’t for supply constraints that were holding the company back, which she considered an industry-wide issue. The company expects MI300 supply to “improve every quarter this year.”
Click to expand...

This is the situation AMD is in. The same supply constraints that are holding back MI300 sales are also the only reason the MI300 sells in the first place. If not for industry-wide supply constraints, everyone would be buying the in-all-ways-superior Nvidia product.
 
Thunderdolt said:
I think Nvidia's 12th biggest customer places single orders that are bigger than AMD's entire data center business.



So no RDNA 4 this year. Got it.



This is the situation AMD is in. The same supply constraints that are holding back MI300 sales are also the only reason the MI300 sells in the first place. If not for industry-wide supply constraints, everyone would be buying the in-all-ways-superior Nvidia product.
Click to expand...

Actually, hardware wise the MI300 is superior to the NV offering currently. So I don't really believe your quota based argument here. AMD is working hand in hand with large customers such as Oracle and MS with the MI300.

I truly believe industry wants an alternative or secondary source for AI hardware other than Nvidia. Just as much as gamers want a second choice for GPUs.
 
Last edited:
Marees said:
I am glad the margins are the same, given that there was a brutal 50% year on year reduction in gaming revenue

I understand microsoft has stopped xbox chip purchases temporarily (until whenever the 6nm refresh launches I guess)
Sony also could have slowed until PS5 pro launches

That said I hope RDNA 4 launches this year. It should help them get some laptop wins (if that is possible)
Also with 7900 GRE discounting for as low as $510, the rdna 4 (navi 48) card which is size of a 6600 xt ahould give them better margins at $550, I guess

(I sincerely hope they learnt something from RDNA 2 oversupply & produced less of RDNA 3 chips !!!)




https://www.tomshardware.com/pc-com...sedives-48-not-expected-to-recover-until-2025
Click to expand...
Does not address my post.
 
FrgMstr said:
Does not address my post.
Click to expand...
A bit hard for me to follow for the lay person, are those margin particularly high for AMD (company wide they were at negative to 2.5% recently), making that segment one they can play with a bit ?
Or already low as they can comfortably be, so no room and rhey should-will hold tight to those precious margins ?
 
Last edited:
Mega6 said:
Actually, hardware wise the MI300 is superior to the NV offering currently.
Click to expand...
That seem incredibly hard to judge something like that are we talking about the 50,000 parts of the hardware solution:

nvidia-dgx-superpod-dgx-h100-systems_mid.png


Or just a specific part ?

I think it is something someone could say, but I would be curious how that was determined. Googling, I cannot find a single instance of someone using a cluster of MI300 and saying it is better hardware wise than a cluster of H100, there is much more in a full stack supercomputer solution than the die and ram over it. Is the hardware connection to share compute task among a vast numbers of nodes scale has well ?
 
Last edited:
Mega6 said:
Actually, hardware wise the MI300 is superior to the NV offering currently. So I don't really believe your quota based argument here. AMD is working hand in hand with large customers such as Oracle and MS with the MI300.

I truly believe industry wants an alternative or secondary source for AI hardware other than Nvidia. Just as much as gamers want a second choice for GPUs.
Click to expand...
The MI300 offers more ram than the H100 and slightly edges it out in some use cases, but the MI300 is up against also the H200 and the BK100.
 
LukeTbk said:
A bit hard for me to follow for the lay person, are those margin particularly high for AMD (company wide they were at negative to 2.5% recently), making that segment one they can play with a bit ?
Or already low as they can comfortably be, so no room and rhey should-will hold tight to those precious margins ?
Click to expand...
Marees stated this, "Do they hold tight to precious margins or price the GPUs to move 🤔"

Radeon GPU is making 16% margin. How much lower margin do you go? 5%? 0%? This is simply a lack of understanding how to run a profitable business IMO.
 
LukeTbk said:
Or already low as they can comfortably be, so no room and rhey should-will hold tight to those precious margins ?
Click to expand...
Yes, this is what I was saying. Margins are still positive but dangerously low.
Clearly AMD has unsold inventory.

Can Lisa Su convince share holders of the need to reduce margins further to clear the inventory ?
 
FrgMstr said:
Marees stated this, "Do they hold tight to precious margins or price the GPUs to move 🤔"

Radeon GPU is making 16% margin. How much lower margin do you go? 5%? 0%? This is simply a lack of understanding how to run a profitable business IMO.
Click to expand...
Honestly in retail if your margin is less than 20% you’re loosing money on it unless it’s something that flys out the door. 16% is actually quite low.
I’m guessing software support and driver/feature development are what’s killing their margins. AMD can’t quite reach Nvidia’s economy to scale there, so proportionally Nvidia gets much better returns.
 
Lakados said:
Honestly in retail if your margin is less than 20% you’re loosing money on it unless it’s something that flys out the door
Click to expand...
If it is operating and not raw margin (and not purchase price - selling price type) are >0% you are not loosing money. Hardware company getting close to 30% that entering Apple territory of all time good, 14-15% that Tesla on a good quarter.

Operating margin usually: operating margin is calculated by dividing a company's earnings by its revenue.
Which would include all expense.

I could be all wrong, but for AMD to be negative in a lot of sector I would imagine they are calculating it like that.
 
KazeoHin said:
To be fair (and I hate this) If someone offered you 10x the money for the same work, wouldn't you serve them first? Like Gamers seem like a publicity stunt or charity at the margins we're looking to pay.
Click to expand...

Gamers want AMD to lower their prices so Nvidia lowers theirs, and then they buy Nvidia. AMD knows this and will likely just continue with their "10% less than Nvidia" pricing strategy.
 
FrgMstr said:
Radeon GPU is making 16% margin. How much lower margin do you go? 5%? 0%? This is simply a lack of understanding how to run a profitable business IMO.
Click to expand...
Not necessarily. NV ernjoys economy of scale and charges a premium for its "brand name." As second banana, AMD has to price below NV.
 
Lakados said:
AMD has lots of cards still on shelves.
Click to expand...
How strange. I distinctly remember members here crowing about how AMD sold every single CPU and GPU it made. According to them, AMD shouldn't have cards sitting on shelves or any unsold inventory at all.
 
ZeroBarrier said:
How strange. I distinctly remember members here crowing about how AMD sold every single CPU and GPU it made. According to them, AMD shouldn't have cards sitting on shelves or any unsold inventory at all.
Click to expand...
Well, one must realize that supply chain is had from static.
 
ZeroBarrier said:
How strange. I distinctly remember members here crowing about how AMD sold every single CPU and GPU it made. According to them, AMD shouldn't have cards sitting on shelves or any unsold inventory at all.
Click to expand...
would that have been when they could release a RX 6600 for $330 and sell it out ? There an reddit thread called :

Is an RX 6600 for 500 dollars worth it atm?​


As anyone suggested such things since Etherum shifting Proof-of-stake looked like it would be real ?
 
You must log in or register to reply here.
Back
Top