AMD Ryzen 9 5950X3D & 5900X3D Historical Prototypes Demoed in Gamers Nexus Video

[Tomgang]

Count me in as well! AMD did mention that more X3D on AM4 wasn't completely out of the picture; they finally did bring out a second in the 5600X3D, and there's still a slim hope of at least an 5950X3D with 1 X3D CCD and 1 normal CCD like the 7950X3D.

Although I'd love for them to just go dual X3D CCDs, picking the top efficient bins of 5800X3D CCDs.

Yes I do hope that they will launch a cpu in not so near future.

Actually i hope they will make a much like Zen 4 3d setup. Meaning 1 3d cashe chiplet and one with out. This will give the best mixed gaming performance and workstation.

Reason is that not all games can use the 3d cashe to it's advantage and is better of with higher core clock.

So with a normal and a 3d-cashe ccd in one package. You can all ways get the best out of the game by using the that ccd best suited for the particular game. With a normal ccd also means you csn still have hight core boost on one ccd.

2 ccd both with 3d-cashe will only raise the price and lower the workstation performance even more do to lower core clock and not many things outside games really benefits from the ekstra v-cashe.

That's my reasoning for why I would rather like that setup. Well also because the cpu will be locket and so you can't do much to raise core clock your self.

 

[R0H1T]

16 Zen 4 cores with high frequencies and X3D would knock out the entire entry-level server market

And people would still whine about it because Asus boards would melt overpowering them, so it wasn't gonna happen any way.

The truth being that 5950X3D with the full 192 MB L3 is likely to stomp the daylights out of the 7950X3D in cache sensitive applications.

It could be more efficient than the current 5950x but no way in hell it'd be worse than 7950x3d except maybe certain edge cases.

 

[Dr. Dro]

And people would still whine about it because Asus boards would melt overpowering them, so it wasn't gonna happen any way.

It could be more efficient than the current 5950x but no way in hell it'd be worse than 7950x3d except maybe certain edge cases.

The 7950X3D doesn't have two X3D CCDs, so yeah, there's a high likelihood that it'd outperform it in cache sensitive applications. But nah, ASUS AM4 boards were solid. Never had a complaint about my B550-E and the only complaint I had about my Crosshair VI was AMD's fault - they were the ones who refused to allow X370 motherboards to be updated for Zen 3 to upsell until Alder Lake happened. So much for the friendly company there.

Yes I do hope that they will launch a cpu in not so near future.

Actually i hope they will make a much like Zen 4 3d setup. Meaning 1 3d cashe chiplet and one with out. This will give the best mixed gaming performance and workstation.

Reason is that not all games can use the 3d cashe to it's advantage and is better of with higher core clock.

So with a normal and a 3d-cashe ccd in one package. You can all ways get the best out of the game by using the that ccd best suited for the particular game. With a normal ccd also means you csn still have hight core boost on one ccd.

2 ccd both with 3d-cashe will only raise the price and lower the workstation performance even more do to lower core clock and not many things outside games really benefits from the ekstra v-cashe.

That's my reasoning for why I would rather like that setup. Well also because the cpu will be locket and so you can't do much to raise core clock your self.

I disagree. Dual X3D CCDs would be the superior option and if one really required to primarily operate a workload which doesn't benefit from X3D, the standard CPUs still exist - and the Ryzen 7 X3D (5800X3D and 7800X3D) more than service the gaming market (initial argument in defense of AMD for not releasing socket AM4-based dual X3D CPUs). For mixed workloads, the penalty that X3D incurs due to its slightly lower clock speeds is highly unlikely to be meaningful - if it's 10% slower here, it's also equally faster there, so it's a big whatever, really. So far the Ryzen 9 X3D parts all rely on a driver to minimize the scheduler fumbling and completely screwing up core order priorities. Which it still does anyway because it relies on an executable whitelist. It's a dirty hack. With dual X3D, this would not be a problem, as the CPUs would behave effectively the same as any other dual-CCD design (as every Ryzen 9 since the 3900X/3950X).

Only Intel's processors currently offer hardware-based, architecture-aware thread scheduling.

+1

I'm not sure if the 5950X3D would be that good.
The 7950X3D has double L2, miles higher operating frequencies and boosts and comes with DDR5....
You have to find a very specific bench/task that uses as much cache as possible to make the latter fall in front of the 5950X3D.

Possible, but I'm not entirely sold that the architectural improvements going from Zen 3 X3D to Zen 4 X3D are as extreme as to make up for that loss, especially with the inherently inefficient topology that the hybrid-cache Zen 4 X3D processors have. As stated on the video and as shown with the 7950X3D, residency is king, and the processor is always attempting to keep a workload in the CCD which is local to that core. Would it be 100% scalable? Probably not, but that wouldn't matter. Neither the 3950X, 5950X or 7950X offer 100% scaling. Software would be optimized around that. Either way, it's just AMD trying their very best to prevent self-owning itself. Why sell a dual-X3D 7950X3D for $799 when they can charge $4,000+ for the privilege? After all, as mighty as the i9-13900KS is, it's not a CPU that's oozing triple digit megabytes of cache.

 

[Jism]

Loved this one. Shows a different aspect of what these engineers, testers, validators and what more have to cope with.

 

[R0H1T]

The 7950X3D doesn't have two X3D CCDs, so yeah, there's a high likelihood that it'd outperform it in cache sensitive applications. But nah, ASUS AM4 boards were solid. Never had a complaint about my B550-E and the only complaint I had about my Crosshair VI was AMD's fault - they were the ones who refused to allow X370 motherboards to be updated for Zen 3 to upsell until Alder Lake happened. So much for the friendly company there.

Yes but you said high frequency zen4 chips, so there would be a problem if they did have extra cache on both dies.

I don't consider them friendly, just budget friendly at best. No company is your friend & no one should ever make the mistake of thinking that way!

 

[ToTTenTranz]

AFAIK these are historical prototypes.

I am surprised they didn't do a dual CCD 3D cache - would be really interested to see how these stack up

Odds are the 5950X has significantly more gaming performance to gain than the 5800X did, because theres are 2x more CPU cores competing for access to the system RAM when there's a cache miss.
System RAM bandwidth per-core is half on the 5950X than on the 5800X. Decreasing the dependence on system RAM nets higher performance boosts on the model with lower RAM bandwidth per-core.


As for the other comments, I doubt there isn’t any market for a 5950X3D. There are plenty people on AM4 who play games and want the productivity of a 16-core CPU.

 

[R0H1T]

It's by far the best chip pre zen4, even with zen4 among the most efficient ever! Extra cache would only help IMO ~

 

[A Computer Guy]

It's by far the best chip pre zen4, even with zen4 among the most efficient ever! Extra cache would only help IMO ~

From the graphs two categories shown it's amazing how competitive the Core i5 12600 is. (sorry I just had to throw that out there)

In my mind GPU is still king in terms of gaming system cost balance and X3D CPU parts are the exception offing boosts in cache sensitive games primarily.

Releasing a single or dual 5950X3D wouldn't have made much sense after the launch of Zen4 with the obvious higher premium AMD would have charged for that.

Having said that a 5950X3D part is appealing to me and if available at the time I might have gotten one since my gaming pc hardware is my backup to my work pc but that's a very specific and niche use case where that kind of purchase sort of makes sense.

 

[Panther_Seraphin]

In the video they also shown the 64-core Threadripper 5990X "Chagall" processor (Milan HEDT) that they never released before pulling the rug under TRX40 chipset adopters, as well as a revision A0 sample of Zen 1 from the very first tray. It's been an enlightening video to watch, but to me, more of heavy confirmation bias towards one of the multiple grudges I've accrued with AMD over time.

THAT is the biggest mistake. The fact they admitted they had developed the Threadripper platform but decided that "nah EPYC is the golden goose, cant compromise that". Especially to people who bought into it with the "promise" of it being supported for many years. The fact they didnt just release the 5xxx series threadripper before abandoning it would have been a fine cut off.

 

[R0H1T]

They didn't admit anything like that, they said there was some issue with it. Whatever it's worth.

 

[Wirko]

One of the first working revisions worked on 18MHz memory

I can think of a reason why, at least if the cores and cache were running at speeds close to normal. You want to benchmark and analyse a system in which a single component - in this case, RAM - is the only limiting factor, with everything else being almost infinitely faster.

 

[Makaveli]

We will probably see the 5600X3D since its single CCD as for the 5900X3D and 5950X3D don't get your hopes up there is a reason they weren't released.

And there focus now is going to keep pushing people to AM5 with Zen 5 around the corner.

 

[Dr. Dro]

They didn't admit anything like that, they said there was some issue with it. Whatever it's worth.

Haha yeah the issue is that they cost the same to make as a $8000 EPYC but would sell for half that tops... In the middle of the global chip shortage... It just wasn't a good business move for them, so screw the TRX40 buyers

I can think of a reason why, at least if the cores and cache were running at speeds close to normal. You want to benchmark and analyse a system in which a single component - in this case, RAM - is the only limiting factor, with everything else being almost infinitely faster.

It's really just an exceptionally buggy early ES in this case. The design wasn't finished, firmware equally in prototype state, and clocks likely controlled directly in jtag at that stage.

 

[Panther_Seraphin]

Haha yeah the issue is that they cost the same to make as a $8000 EPYC but would sell for half that tops... In the middle of the global chip shortage... It just wasn't a good business move for them, so screw the TRX40 buyers

Pretty sure this was the time AWS/Google/Azure etc were buying up EPYC as quick as they were coming out the Fabs

The gall they had to release the Threadripper pro line and abandon TRX40 is just bullshit IMO.

 

[NC37]

Shame, I'd probably buy a 5900x3D if they did it.