News AMD outsells Intel in the datacenter for the first time in Q4 2024

[tamalero]

The Ryzen precedent required Intel to fall on its face and get stuck on the same node for 6 years. That obviously can't happen to Nvidia. In the most unlikely of scenarios if that did happen, AMD would be stuck in the same boat since they both use TSMC and no advantage would be gained. If anything Nvidia is in a better position since they have used Samsung as well recently and don't depend on just one producer like AMD does.

intels issues were more than just a node problem.

 

[jp7189]

Intel is still cutting labor and restructing; they've always been a much larger company than AMD (although it's the smallest gap ever now), so when revenues are similar between the two, you bet AMD will have better margins on that revenue. Next fiscal year should show better margins for Intel, assuming they can stop the DCAI hemorrhaging .

Gross margin is a metric of costs of good sold and doesn't factor in general overhead. It means Intel is spending more (poor internal fab yields, high TSMC contract) and getting less on the store shelf (as it were) compared to AMD. Net profit includes operating expenses and includes the factors you mentioned.

 

[DS426]

Gross margin is a metric of costs of good sold and doesn't factor in general overhead. It means Intel is spending more (poor internal fab yields, high TSMC contract) and getting less on the store shelf (as it were) compared to AMD. Net profit includes operating expenses and includes the factors you mentioned.

Ah, true -- good call.

 

[Ogotai]

The Ryzen precedent required Intel to fall on its face and get stuck on the same node for 6 years

your seem forgetting the other part of that.
intel suck the mainstream at 4 cores, when they could of added more, but they didnt. that forced those that wanted to use more cores, to go to their HEDT platform, ( inc me then ) costing hundreds more. amd gave that market more cores, that was more affordable. as Tamalero also mentioned, it wasnt just a node problem....

their 10nm issues.. just made it worse....

 

[spongiemaster]

your seem forgetting the other part of that.
intel suck the mainstream at 4 cores, when they could of added more, but they didnt. that forced those that wanted to use more cores, to go to their HEDT platform, ( inc me then ) costing hundreds more. amd gave that market more cores, that was more affordable. as Tamalero also mentioned, it wasnt just a node problem....

their 10nm issues.. just made it worse....

All BS. The problems all circle back to Intel's inability to get 10nm out. Intel's roadmaps had 10nm Cannonlake releasing in 2016. The year before AMD released Zen1.

Intel Desktop CPU Roadmap Confirms Broadwell-E In Q3 2015 - Broadwell K-Series in Q1 2015 Followed by Skylake In Mid-2015

The latest Intel 2013-2015 CPU roadmap has been leaked on the web detailing Intel's upcoming CPU platforms which include the HEDT Haswell-E and Broadwell-E has been leaked.

wccftech.com

Cannonlake was supposed to bring 8 cores to the mobile and mainstream desktop market before AMD did. We know this now due to multiple leaks and engineering samples that got out. That never happened because 10nm wasn't working then and desktop CL got cancelled and instead the pointless Kaby Lake was released and then Coffee Lake which hit 6 cores and mobile CL only saw a garbage dual core with no iGPU, so Intel could claim they released a 10nm CPU. Intel had no plans to release higher core desktop CPU's on 14nm for multiple reasons, but eventually was forced to as they had fallen so far behind AMD because of their 10nm node issues.

 

[bit_user]

The Ryzen precedent required Intel to fall on its face and get stuck on the same node for 6 years.

The difference between their CPU business @ 2017 and their GPU business @ 2024+ is that their CPUs were much further behind Intel, on a per-core basis.

If you look at the RX 7900 XTX, it has only 6144 shader cores. Compare this to the RTX 4090, which has16384 CUDA cores. This tells me AMD is actually doing pretty well, on a microarchitecture level. I think AMD could've pretty easily scaled up RDNA3 bigger than they did and brought some meaningful heat on the 4090.

That obviously can't happen to Nvidia.

It actually did, with Ampere using Samsung 8 nm, while AMD was on TSMC N7. AMD also out-innovated Nvidia with Infinity cache, and that's how the RX 6950X managed to snatch the crown of raster performance from the RX 3090 Ti. Weren't you paying attention?

Source: https://www.tomshardware.com/reviews/amd-radeon-rx-6950-xt-review/6
​

However, it was just a single-generation blip. A flustered Nvidia went huge on their RTX 4090, also used a more competitive node, and added a competitive amount of L2 cache. That's how Nvidia managed to regain its pride.

If anything Nvidia is in a better position since they have used Samsung as well recently and don't depend on just one producer like AMD does.

Let's hope next time with Samsung is better than the last.

 

[bit_user]

Cannonlake was supposed to bring 8 cores to the mobile and mainstream desktop market before AMD did.

What's your source on that? The Cannon Lake they actually launched had just 2 cores, suggesting their plan was to continue with a 2-core low-end die and 4 cores at the high end.

https://www.anandtech.com/show/13405/intel-10nm-cannon-lake-and-core-i3-8121u-deep-dive-review
​

Ice Lake launched only 4 cores for laptops, which is inconclusive. It doesn't mean they didn't plan another die that was either bigger or smaller, just that they weren't viable.

We know this now due to multiple leaks and engineering samples that got out.

I'd love to see your source on that, because it's not supported by the one you actually linked.

 

[spongiemaster]

What's your source on that? The Cannon Lake they actually launched had just 2 cores, suggesting their plan was to continue with a 2-core low-end die and 4 cores at the high end.

​

https://www.anandtech.com/show/13405/intel-10nm-cannon-lake-and-core-i3-8121u-deep-dive-review​

​

Ice Lake launched only 4 cores for laptops, which is inconclusive. It doesn't mean they didn't plan another die that was either bigger or smaller, just that they weren't viable.


I'd love to see your source on that, because it's not supported by the one you actually linked.

2015
Intel Cannonlake CPU With Up To 8 Cores And Coherent Fabric Spotted


Intel’s Unreleased 10nm Cannonlake 8 Core And 6 Core Laptop CPUs Spotted – Abandoned Due To Poor Yields or Inefficient Design?

 

[bit_user]

2015
Intel Cannonlake CPU With Up To 8 Cores And Coherent Fabric Spotted

Coherent Converged Fabric is the mesh-like interconnect they use in HEDT and Xeons. Their desktop CPUs have used a ring bus, ever since Sandybridge. So, no. That's not proof of an 8-core desktop die.

Check it:

"Even though Intel Corp. has been gradually increasing the number of general-purpose processing engines inside its microprocessors for servers, the company’s high-end chips for client PCs feature no more than eight x86 cores, whereas mainstream central processing units sport up to four cores. Apparently, Intel has no plans to change anything in the coming years as the company’s forthcoming Cannonlake-E/EP CPUs for high-end desktops and workstations will integrate up to eight cores."

Intel ‘Cannonlake-E/EP’ for desktops and servers to sport up to 8 cores - KitGuru

Even though Intel Corp. has been gradually increasing the number of general-purpose processing engin

www.kitguru.net

Intel’s Unreleased 10nm Cannonlake 8 Core And 6 Core Laptop CPUs Spotted – Abandoned Due To Poor Yields or Inefficient Design?

Yes, this does look like evidence of 8 & 6-core parts, although it was only spotted in a 3DMark Database, which is an extremely borderline source.

 

[spongiemaster]

Coherent Converged Fabric is the mesh-like interconnect they use in HEDT and Xeons.

No, it isn't. This is not related to the core topology.

Intel used 3 different core configurations. The low core count configuration was 4-10 cores. Skylake-X didn't have a 4 core so it used the LCC for 6-10 core variants. The Skylake X refresh didn't have 6 core either, so it used the LCC for just 8 and 10 core variants. So an employee saying he was working on 4, 6 and 8 core CPU's doesn't lineup with anything that would have been an HEDT product.

Intel's HEDT would not be called an SOC either which would typically have a graphics component. Nothing in that job posting points to a potential Cannonlake X platform that would have followed Skylake X.

 

[OldAnalogWorld]

I don't see AMD's success from the investors' point of view - its shares have already fallen almost 2 times in 2 years. It is starting to turn into a zombie company, especially after Intel rose from the ashes thanks to government subsidies, although it should have disappeared from the market. But it seems that AMD is now in trouble, its advantage in technological processes is rapidly melting away...

 

[bit_user]

No, it isn't. This is not related to the core topology.

Okay, let's say you're right. By 2015, Intel already should've heard rumors of Zen, possibly including that AMD was going to 8 cores. We know Jim Keller went to AMD in August 2012. The whole point of him going there was to spearhead the development of Zen. That's sure to have been making waves.

The relevant fact I see is that Intel didn't release any > 4-core mainstream CPUs until a year after Zen, which was an obvious response. And in that case, it was only a 6-core, which makes a lot more sense than them going directly from 4 to 8 cores. In fact, the only reason I can see why they'd have gone straight to 8 cores is that they were spooked by rumors of Zen.

Furthermore, the fact that they did up to 10 cores on 14 nm, in a mainstream socket, tells you all you need to know about whether Intel had been holding back. They'd been doing 4-core mainstream CPUs since Yorkfield (45 nm) in 2008. Even on 14 nm, they absolutely could've gone beyond 4 cores before 2017's Coffee Lake.