News Ryzen 9 9950X3D purportedly surfaces with 5.65 GHz boost clocks — leaked specs inlcude 96MB 3D V-Cache on a single CCD and 170W TDP

[80251]

@spongiemaster
If Intel is raking in so much money then why can't they challenge the 3D vcache AMD parts in performance? Does all their profit go into the pockets of the board of directors and CEO?

 

[thestryker]

@spongiemaster
If Intel is raking in so much money then why can't they challenge the 3D vcache AMD parts in performance? Does all their profit go into the pockets of the board of directors and CEO?

It's not a matter of can't, but rather a matter of why would they. I'm sure their engineers would love to compete and have probably pitched just that. The problem is that Intel is a business and due to their CPU design there's no cost effective way to compete. AMD may have made a splash with X3D, but these aren't volume parts because retail isn't volume.

The reason AMD can afford to bring them to retail is that their CCD/IOD design means the only thing they had to figure out was the most effective way to add the cache. Comparatively speaking this is simple because the only thing that needs modification is a single CCD.

 

[Ogotai]

We have seen intel's answer for years now, they increase the cache bit by bit instead of all at once like amd does

and that has really gotten them where ? intel cant really compete with x3d.....

while amd has a real issue with selling their non x3d cpus because of how much better the x3d ones are for the mainstream users.

maybe.. but it seems intel cant sell their 13th or 14th gen cause of the issues those cpu have.. no one i know is even looking at intel now be cause of that. as for the " fixes " intel has issued... there will be no proof, they are fixed, for at least 6 more months, if not longer...

 

[Ogotai]

It's not a matter of can't, but rather a matter of why would they. I'm sure their engineers would love to compete and have probably pitched just that. The problem is that Intel is a business and due to their CPU design there's no cost effective way to compete. AMD may have made a splash with X3D, but these aren't volume parts because retail isn't volume.

The reason AMD can afford to bring them to retail is that their CCD/IOD design means the only thing they had to figure out was the most effective way to add the cache. Comparatively speaking this is simple because the only thing that needs modification is a single CCD.

didnt amd lay the ground work for x3d a few years ago by adding the capability for the extra cache when they designed the chiplets ?

 

[thestryker]

didnt amd lay the ground work for x3d a few years ago by adding the capability for the extra cache when they designed the chiplets ?

Yes, because Epyc and Ryzen share CCD design so when 3D V-cache was developed for Epyc the Ryzen CCDs had the TSVs required as well. This is another reason why the cost to put it on client parts is so much lower than it would be if they had to design it just for client parts.

 

[spongiemaster]

@spongiemaster
If Intel is raking in so much money then why can't they challenge the 3D vcache AMD parts in performance? Does all their profit go into the pockets of the board of directors and CEO?

thestryker covered the primary reason why. The market size doesn't justify Intel developing the product. AMD's modular design and the fact they use the same parts in all markets allows them to spread the development costs across their entire product portfolio. The 3DV dies that end up in consumer CPU's are binned Epyc dies that didn't make the cut. AMD didn't implement stacked IC's with the purpose of coming up with the best gaming CPU. They developed it for the enterprise market. The fact it was great at gaming was a side benefit that allowed them to offload dies that didn't make the grade. Intel's enterprise and consumer CPU's do not share the same die so if Intel wanted to use the technology for gaming, the consumer market would have to fund all the development costs.

It also needs to be pointed out that AMD didn't develop 3d chip stacking. 3D Vcache is AMD's marketing name for their implementation of TSMC's 3d packaging. AMD played no role in the development of the technology, and Intel could use it as well if they wanted to pay for it. AMD adapted the technology to their use case and developed a commercially viable product, but TSMC did most of the heavy lifting.

Lastly, Intel's CPU divisions are magnitudes more profitable than AMD's. The problem at Intel is their foundry business is currently bleeding billions a quarter as it tries to expand capacity and hit ex-CEO Gelsinger's 5 nodes in 4 years goal. The CPU divisions aren't nearly profitable enough to cover those losses, so any sort of side projects or niche products are going to get cut first or not started at all.

 

[80251]

Eventually, the server/workstation market is going to catch on that AMD has better CPUs and THEN where will Intel be? Isn't Intel's market share across the gamut of CPU markets being steadily eroded by AMD?

 

[Imaletyoufinish]

Is it unreasonable to expect something as smooth as the way threads slide around P and E cores on an Intel platform? X3d has been around for many years at this point.

X3D parts with dual CCDs have only been around since spring 2023 (about 20 months ago). Even 8 core X3D has only been around for since 2022 (about 2.5 years). "Many years" is a big stretch. In contrast Alder Lake's P and E cores debuted over 3 years ago with Intel working closely with Microsoft, as they always have for decades, and it still took a while for them to get it right (still has issues here and there last I heard). Microsoft is not as close with AMD as they should be so don't expect the same result as Intel in less time.

 

[thestryker]

X3D parts with dual CCDs have only been around since spring 2023 (about 20 months ago).

AMD announced the investigation of stacked cache in 2020, publicly debuted it June 2021 (dual CCD with cache on both), Milan-X launched March 2022 and 5800X3D June 2022. Realistically AMD probably had already nailed down the core configuration by the time those products actually launched.

"Many years" is a big stretch. In contrast Alder Lake's P and E cores debuted over 3 years ago with Intel working closely with Microsoft, as they always have for decades, and it still took a while for them to get it right (still has issues here and there last I heard). Microsoft is not as close with AMD as they should be so don't expect the same result as Intel in less time.

AMD's problems have existed since the launch of Zen 2 though. Everything is due to the dual CCD design and scheduling. While it's typically fine there always seem to be some random outliers (typically gaming type workloads). In an ideal world the highest clock speed parts should always be faster than any of the other non-X3D parts, but that isn't necessarily the case.

It mostly seems like there needs to be a lot more scheduler optimization for both AMD and Intel on Windows. Both companies have software solutions to work around the limitations and it would be much better if that could be avoided.

 

[Imaletyoufinish]

AMD announced the investigation of stacked cache in 2020, publicly debuted it June 2021 (dual CCD with cache on both), Milan-X launched March 2022 and 5800X3D June 2022. Realistically AMD probably had already nailed down the core configuration by the time those products actually launched.

AMD's problems have existed since the launch of Zen 2 though. Everything is due to the dual CCD design and scheduling. While it's typically fine there always seem to be some random outliers (typically gaming type workloads). In an ideal world the highest clock speed parts should always be faster than any of the other non-X3D parts, but that isn't necessarily the case.

Yeah I remember the public debut with what would've been the 5950X3D. I haven't seen it in years so don't remember them mentioning that stacked cache was on both CCDs. I guess you're saying AMD has been working on this stuff for many years which I would have to agree with. I was mainly talking about how long consumers have been able to buy an X3D part.

I know of Milan-X and assume it's the first data center chips that had stacked cache. Early 2022 was kinda crazy for me so forgive me for forgetting about the details of some of the releases around that time. I stand corrected.

I thought we were just talking about X3D dual CCD issues but it's definitely true that since AMD has been using chiplets latency has been a problem. High Yield and others have talked about Zen 6 being when AMD stacks the CCDs and IOD together which should drastically reduce latency and alleviate the issues we see with dual CCD X3D as well. I'm waiting here with my 7700X until I see a Ryzen 9 worth me upgrading to. Planned my build to handle 16 cores but if Zen 6 is using 12 core CCDs I may get a 11900X3D or whatever it's called.

 

[spongiemaster]

AMD's problems have existed since the launch of Zen 2 though. Everything is due to the dual CCD design and scheduling. While it's typically fine there always seem to be some random outliers (typically gaming type workloads). In an ideal world the highest clock speed parts should always be faster than any of the other non-X3D parts, but that isn't necessarily the case.

It mostly seems like there needs to be a lot more scheduler optimization for both AMD and Intel on Windows. Both companies have software solutions to work around the limitations and it would be much better if that could be avoided.

The problem with gaming on a dual CCD CPU goes all the way back to first generation Zen. The 16 core Threadripper 1950x had a game mode that required a reboot to disable one of the CCD's. That CPU is over 7 years old now and people still don't seem to get the issue which has not changed. Whether there is stacked cache on the die or not is completely irrelevant to the root issue which is latency between the CCD's is terrible for gaming performance. There is no work around solution to this problem to make this configuration work properly in gaming with both CCD's active. AMD has been unable to make this work with CPU's or GPU's (cancelled multi-chiplet RDNA4) and people need to stop thinking we're going to see X3D CPU's with dual stacked cache CCD's any time soon.

 

[80251]

I'm sorry to hear that AMD cancelled their multi-chiplet RDNA4. Did they get as far as a prototype?