Info Intel 4Q23 Earnings

[trivik12]

I think only Mi300a from El Capitan will be counted in Q4 for AMD, which should be ~ $400m in sales. Which is still meaningful.

Mi300x only kicks-in in Q1 2024. AMD estimated $2 billion in sales for 2024. Patrick Moorhead just predicted $10 billion for 2024. So that's a pretty big range.

2B would be awful. Nvidia is looking at ~50B for 2024 depending on supply constraints. 10B would be really good. Would need 50K high end chips at 20K ASP.

 

[Joe NYC]

2B would be awful. Nvidia is looking at ~50B for 2024 depending on supply constraints. 10B would be really good. Would need 50K high end chips at 20K ASP.

It is hard to get the exact numbers, of how many Mi300a AMD is going to ship this quarter into the El Capitan supercomputer, but for comparison, the Frontier supercomputer apparently has 37,000 Mi250 GPUs.

I would estimate that number of Mi300a in El Capitan to be similar to number of Mi250 in Frontier.

So, if AMD can ship nearly 37k Mi250 / Mi300 in a quarter, shipping 50k even 100k should be within realm of possibilities for full year.

(if anyone has better numbers for the number of Mi300a / Mi250a in those supercomputers, I would be interested to know and correct this).

 

[gdansk]

The number of GPUs is not known but we have an estimated price of $600 million for El Capitan which, probably, makes the number of GPUs pointless for this discussion. That's a reasonable upper limit for how much it alone could contribute to AMD's 2023/2024 MI300 sales.

 

[Joe NYC]

The number of GPUs is not known but we have an estimated price of $600 million for El Capitan which, probably, makes the number of GPUs pointless for this discussion. That's a reasonable upper limit for how much it alone could contribute to AMD's 2023/2024 MI300 sales.

$600m includes HP portion (networking cabinets, blades). $400m is the estimate for AMD Mi300a, and will likely be all in Q4 earnings.

Government is most likely paying lower price than current market price of H100. If the number of Mi300a is 40,000, it would be about $10,000 per unit.

BTW, looks like @trivik12 is off by a zero in his estimate. It would take 500k units at $20k to reach $10 billion.

 

[turtile]

$600m includes HP portion (networking cabinets, blades). $400m is the estimate for AMD Mi300a, and will likely be all in Q4 earnings.

Government is most likely paying lower price than current market price of H100. If the number of Mi300a is 40,000, it would be about $10,000 per unit.

BTW, looks like @trivik12 is off by a zero in his estimate. It would take 500k units at $20k to reach $10 billion.

If I remember correctly, someone stated that the 400 million would also be a part of Q1. Considering that part of the deal involves developing ROCm software for AMD, a massive discount is worth it for AMD. There are estimates that the average cost of the MI300X will be around $16K. Considering that most units will be sold to Microsoft and Meta at volume discounts, the estimates sound about right. A small business will probably pay $20K a unit.

 

[Philste]

His tone is very foreign here. We're not gonna have a hangover because someone disappears after making big claims for months without proof.

I doubt he will disappear. Wrote similar stuff about RDNA3 on Twitter in the months before release and after it was obvious that it failed he was just like "yeah, just some Silicon that needed to get out and wasn't ready, but the next product is the one that kills everything". Now the next train is going, this time ZEN5. So if ZEN5 fails, that doesn't mean he will stop.

 

[Hitman928]

If I remember correctly, someone stated that the 400 million would also be a part of Q1. Considering that part of the deal involves developing ROCm software for AMD, a massive discount is worth it for AMD. There are estimates that the average cost of the MI300X will be around $16K. Considering that most units will be sold to Microsoft and Meta at volume discounts, the estimates sound about right. A small business will probably pay $20K a unit.

$400M in Q4 for AMD for El Capitan supercomputer and $400M in Q1 for AMD for DC GPU (MI300 based products) for other customers. These are based off of Lisa Su comments from the Q3 earnings call.

 

[Joe NYC]

$400M in Q4 for AMD for El Capitan supercomputer and $400M in Q1 for AMD for DC GPU (MI300 based products) for other customers. These are based off of Lisa Su comments from the Q3 earnings call.

Yes, that is how I remember it as well.

 

[Joe NYC]

I doubt he will disappear. Wrote similar stuff about RDNA3 on Twitter in the months before release and after it was obvious that it failed he was just like "yeah, just some Silicon that needed to get out and wasn't ready, but the next product is the one that kills everything". Now the next train is going, this time ZEN5. So if ZEN5 fails, that doesn't mean he will stop.

Since RDNA3 surprised AMD as well, it does not invalidate anyone's comments that were consistent with AMD expectations. On the contrary, it can increase the credibility of the comments predicting higher RDNA3 performance.

 

[adroc_thurston]

"yeah, just some Silicon that needed to get out and wasn't ready, but the next product is the one that kills everything"

Ugh, no.
They just missed.
And you don't miss at AMD.

 

[jpiniero]

Intel's tried to claim Gaudi is competitive, but nobody cares. The reason is that the H100 (and MI300X) are useful beyond just AI... and once Wall Street starts to lose interest in AI I imagine the spending will die down. But the chips can be useful for other things.

 

[Saylick]

Gaudi is such a poorly chosen name. You think the company that at one point considered Moore Noyce as its name would know better than to use a homophone for gaudy.