AMD Ryzen 9 9950X

[RogueSix]

@Dr. Dro

I thought we were talking about gaming? No one is saying that it is unfeasible for other scenarios like certain server applications that benefit from such configs. However, specifically for gaming, where low latency is king, I don't see multi 3D cache chiplets happening any time soon (if ever).
Especially not after these Zen 5 multi CCD reviews. Zen 5 should be regarded as proof that AMD has a lot of home work to do with regard to the regular scheduling, i.e. without adding any extra complexities on top of the wonky base.

 

[chrcoluk]

And so the plot thickens a little bit more:

Description: Windows Bug Found, Hurts Ryzen Gaming Performance

Same video twice in your post for some reason, I will do some digging to see if I can find anything that explains this mystery, but some of us on TPU have already discovered that the windows scheduler has issues.

The only thing I do recall at the time of this post is there is some differences on thread priority in between elevated and non elevated processes. One of the reasons why I think OBSS has a run as elevated option.

 

[Dr. Dro]

@Dr. Dro

I thought we were talking about gaming? No one is saying that it is unfeasible for other scenarios like certain server applications that benefit from such configs. However, specifically for gaming, where low latency is king, I don't see multi 3D cache chiplets happening any time soon (if ever).
Especially not after these Zen 5 multi CCD reviews. Zen 5 should be regarded as proof that AMD has a lot of home work to do with regard to the regular scheduling, i.e. without adding any extra complexities on top of the wonky base.

I don't think CPUs are ever designed for gaming in itself, either way, the segment that gets targeted for gamers is usually Ryzen 5 and Ryzen 7 (in contrast to i5/i7), the R9 chips have always been geared towards productivity/extreme enthusiasts, to whom a dual X3D chip makes sense even if there are some drawbacks to it. We wouldn't mind.

 

[Daven]

whoops

View attachment 359096

This is the best performance comparison IMHO as rarely is it recommended to upgrade every generation. While you can get close to these gains with a 7950X for $100 less, it still depends on the usage case as many have pointed out the insane gains in benchmarks such as V-ray.

If you love powerpoint, you can stick with…oh who cares, why would anyone upgrade to get better powerpoint performance!?!?!

 

[Zach_01]

And so the plot thickens a little bit more:

Description: Windows Bug Found, Hurts Ryzen Gaming Performance

Huh..?
He should have tested at least 1 Intel CPU as well

 

[Vincero]

Huh..?
He should have tested at least 1 Intel CPU as well

View attachment 359115

Agreed, but then people would probably bitch about which one... Arguably a 12100->12600 (non-K) would be ideal as it would exclude any E-core fun and games getting in the way (and Raptor Lake non-E-core parts are just Alder Lake H0 dies anyway), similar to how AMD is used.

I don't see this necessarily fixing the 2xCCD Zen5 CPUs lack of gaming performance increase compared to single CCD equivalents - unless this magically also gives a massive boost to the effectiveness of the core parking AMD PM software

 

[Beginner Macro Device]

To be fair, at least AMD doesn't require a new socket with every new generation.

This, frankly, is just an utter garbage argument. Upgrading a mobo alongside everything else opens up faster USB, better Thunderbolts, new advanced ports... you name it. Why would you love to get partially stuck on outdated tech? NB: upgrading more often than once a 5 years is currently a sign of severe ADHD and not a reasonable pattern. We don't get killer gen-to-gen improvements like we did in last century, they ain't even significant anymore.

And by the way:

(I will not comment on Zen 5 because this gen doesn't deserve that)

 

[oZ65]

Agreed, but then people would probably bitch about which one... Arguably a 12100->12600 (non-K) would be ideal as it would exclude any E-core fun and games getting in the way (and Raptor Lake non-E-core parts are just Alder Lake H0 dies anyway), similar to how AMD is used.

I don't see this necessarily fixing the 2xCCD Zen5 CPUs lack of gaming performance increase compared to single CCD equivalents - unless this magically also gives a massive boost to the effectiveness of the core parking AMD PM software

The reality is that everything should be retested.... What a headache

 

[sdedalus83]

W1zzard, have you tested other AM5 boards for idle power consumption? Guru3d has whole system idle numbers 25-30 watts lower than yours with what seems like similar hardware.

Review: Ryzen 9 9950X processor - Updated with KB5041587 for Gaming Charts (Page 27)

We review the ZEN 5 processor series, highlighting the powerful 16-core Ryzen 9 9950X. This model boasts the highest number of cores for consumers and operates at the fastest speeds. We're moving beyond high-end to the enthusiast tier of processors.

www.guru3d.com

 

[R0H1T]

This, frankly, is just an utter garbage argument.

WTH are you even talking about? You can probably get a second hand flagship AM4 board for $50~100 & at least a mid range B550 or x570 mobo (brand new) close to $100~150 with all bells & whistles!

Upgrading a mobo alongside everything else opens up faster USB, better Thunderbolts, new advanced ports... you name it.

Hey wait till you find out what you just got jumping from x370 to B550 or x570 all with the same socket

NB: upgrading more often than once a 5 years is currently a sign of severe ADHD and not a reasonable pattern.

Nah people tooting Intel's horn for changing socket every 2 years are "disingenuous" at best & shills at worst

Even replacing an AM4 board/buying a new one's way cheaper than anything Intel from 5-10 years back!

 

[Hecate91]

This, frankly, is just an utter garbage argument. Upgrading a mobo alongside everything else opens up faster USB, better Thunderbolts, new advanced ports... you name it.

It isn't, and the cost of a B550 or X570 motherboard is still cheaper while getting all the features most people would need.
Things like thunderbolt, USB 3.2 or PCI-e gen 5 don't matter for most users in a gaming system.

Why would you love to get partially stuck on outdated tech? NB: upgrading more often than once a 5 years is currently a sign of severe ADHD and not a reasonable pattern.

And why would you love to upgrade the board every 2 years? When the socket doesn't change for more than 2 years upgrading makes a lot more sense, for example you can start with a cheap cpu if your budget is limited then upgrade to a flagship or X3D cpu later.

We don't get killer gen-to-gen improvements like we did in last century, they ain't even significant anymore.

Going from a Ryzen 3600 to a 5800X3D is a 15-20% performance increase, you don't get that level of performance increase on Intel because they change sockets too often.
Edit- more of a 30% or higher increase in 1080P gaming.

 

[AVATARAT]

Some info on the 9950x "feature's", congrats to the winners

 

[Beginner Macro Device]

You can probably get a second hand flagship AM4 board

Why should I do that if I'm not on a prehistoric PC (Haswell/AM3+ and older)?

Hey wait till you find out what you just got jumping from x370 to B550 or x570 all with the same socket

A great hole in my budget. And maybe, just maybe, a one-gen improvement in about one or two interfaces.

Even replacing an AM4 board/buying a new one's way cheaper than anything Intel from 5-10 years back!

If you're buying a 5+ y.o. system for an upgrade it's a clear indication there's something horribly wrong going on.

My point is you, as an end user, want to have "wow, it's illegal to be that fast" from your new PC. It is legit not possible if you upgrade every time they release something new. This itch for upgraded demotivates them from making gen-to-gen improvements. Why bother if it sells anyway?
However, if you wait for long enough and, say, upgrade your i7-6700K to i7-14700K (Ryzen 1600 to Ryzen 7700X for AMD fans) it's a step-up you just can't miss.
I currently own an i5-12400F. Do next-gen CPUs outperform it? Of course! Is this margin huge? OF COURSE NOT! I also don't miss out because everything I work with works fine. When I finally upgrade this will be a 9 y.o. CPU and neither LGA1700 nor AM5 will be relevant by then.

And why would you love to upgrade the board every 2 years?

I wouldn't. Way too often.

Going from a Ryzen 3600 to a 5800X3D is a 15-20% performance increase

It's nowhere near 150% (which is a bare minimum if you want your upgrade to be A REAL UPGRADE).
Unless you get paid a lot for your calculating power, just read what I wrote above quoting you. In case you really itch for some next-gen game that doesn't run smooth enough you can, instead of wasting hundreds of dollars on upgrading, just go to some gaming centre and pay a dozen dollars for their VIP gaming computers for rent.

 

[9087125]

If it turns out that AMD fix the drivers that miraculously give a big increases in performance and users report an easy upgrade from their 7950X, then I will consider it, but this Zen 5 is so far off the map that im pretty much digging my self in here, going to ride out my am5 system. And tbh, I fully anticipated this, I already assumed with my 7950X build that this would be the final cpu ever.

 

[Zach_01]

This, frankly, is just an utter garbage argument. Upgrading a mobo alongside everything else opens up faster USB, better Thunderbolts, new advanced ports... you name it. Why would you love to get partially stuck on outdated tech? NB: upgrading more often than once a 5 years is currently a sign of severe ADHD and not a reasonable pattern. We don't get killer gen-to-gen improvements like we did in last century, they ain't even significant anymore.

And by the way:
View attachment 359120

(I will not comment on Zen 5 because this gen doesn't deserve that)

And I have this AM4 system for 5 whole years now and will still have it for at least 1 year more. And many many users have AM4 even longer and enjoying very nice performance with all the 5000series
What is the point of the above conversation? To de-value the best platform ever?
I dont quite get the narrative here

Going from a Ryzen 3600 to a 5800X3D is a 15-20% performance increase, you don't get that level of performance increase on Intel because they change sockets too often.

Actually its from ~10% to ~50% increase depending the resolution and GPU

 

[R0H1T]

This itch for upgraded demotivates them from making gen-to-gen improvements. Why bother if it sells anyway?

You do know that ~16% IPC on the same node, same RAM, same everything is nearly unheard of in the last 15-20 years right? And yes the IPC numbers are fairly accurate! And they did it twice with zen2>3 & now zen4>5 although for zen3 there was higher clock speeds as well IIRC. Zen5 is worse because it's clocked slightly lower than zen4 I guess?

If you're buying a 5+ y.o. system for an upgrade it's a clear indication there's something horribly wrong going on.

We were talking about old parts being replaced or upgraded. This is simply not a good/cheap option on Intel & hasn't been since LGA 775 died for Intel's insatiable greed.

AMD's getting there but they're still only about half way atm.

 

[TumbleGeorge]

^^^Now, a really in US, used motherboard, a flagship model from just 3-4 years ago can sell for as price for a pizza. However, don't think that it's like that everywhere.

 

[R0H1T]

That's probably true everywhere else except US & most(?) of Western Europe. Second hand (Intel) mobos aren't exactly plentiful or cheap anywhere else at least AFAIK.

 

[mahirzukic2]

Hmm... focusing on application performance under Linux, as seen on Phoronix, paints a slightly different picture.
The performance increases from this generation are more pronounced, and in benchmarks utilizing AVX-512 the new full-width backend shows its strength by the 12c 9900X sometimes beating the 16c 7950X. All that while taking less power than Zen 4.

I was just gonna reply how I wasn't that impressed with 9950x even though it should have had more performance increases than it actually did, but this was only reading the review from TPU, after reading the Phoronix review, man this is an absolute beast, especially in AVX capable workloads.

Which would mean that for my use cases this CPU would be good as well, but my initial (even before the reviews came out for Zen 5) decision to migrate to Zen 6 will be even more correct and justified.

 

[TumbleGeorge]

Hmm. I wonder if the only reason they keep the IO chip unchanged is just to keep costs down...Or limitations in the socket itself. Intel, perhaps not by accident, increases the number of contacts from 1700 to 1851, although it reduces power consumption, so it's not for more wiring for power supply...While staying AM5 may have purely physical obstacles to push signaling with higher bandwidth to RAM.

 

[Event Horizon]

Interesting findings on the Windows 11 bug from HUB Steve.

 

[Vincero]

Hmm. I wonder if the only reason they keep the IO chip unchanged is just to keep costs down...Or limitations in the socket itself. Intel, perhaps not by accident, increases the number of contacts from 1700 to 1851, although it reduces power consumption, so it's not for more wiring for power supply...While staying AM5 may have purely physical obstacles to push signaling with higher bandwidth to RAM.

It's planned - on AM4, same IO Die was used on Zen2 and Zen3.
I suspect AMD have plans for a newer IO die where some new requirement forces it, such as USB or PCIe spec changes, or maybe improved DDR5 performance - that's one area I think that would be of interest to some especially if some of the lack of performance improvements of Zen4>5 are bandwidth starvation.
Wouldn't necessarily require a new socket for those features, but likely motherboards with new chipsets that support the required newer signalling, kind of like AM2>AM2+, AM3>AM3+ where different hyper-transport speeds were supported, etc., so maybe there could be an AM5+ which still offers some backwards compatibility if such a thing happened.

Along the memory limitation lines, if the desktop moved to CAMM memory, one CAMM module provides 2x64-bit memory channel access, so it would be possible to move to 'quad channel' on the desktop with a twin-CAMM motherboard (maybe one CAMM on each side). That would need more socket pins and a new IO die, but the flexibility of the chiplet design means that the CCDs could remain unchanged. At this point in time, that's probably the most likely thing that would force an AM5+/AM5b platform, unless it co-incided with DDR6 and PCIe6, in which case with such wholesale change it would probably mean a whole new AM6 platform.

 

[Dr. Dro]

This, frankly, is just an utter garbage argument. Upgrading a mobo alongside everything else opens up faster USB, better Thunderbolts, new advanced ports... you name it. Why would you love to get partially stuck on outdated tech? NB: upgrading more often than once a 5 years is currently a sign of severe ADHD and not a reasonable pattern. We don't get killer gen-to-gen improvements like we did in last century, they ain't even significant anymore.

And by the way:
View attachment 359120

(I will not comment on Zen 5 because this gen doesn't deserve that)

To be honest, at that point I would probably do the same thing I did with my 13900KS build, invest in a quality board like the Crosshair VIII Dark Hero and keep using that thing. A 5800X3D gaming PC is about as good as a gaming PC gets tbh.

Along the memory limitation lines, if the desktop moved to CAMM memory, one CAMM module provides 2x64-bit memory channel access, so it would be possible to move to 'quad channel' on the desktop with a twin-CAMM motherboard (maybe one CAMM on each side). That would need more socket pins and a new IO die, but the flexibility of the chiplet design means that the CCDs could remain unchanged. At this point in time, that's probably the most likely thing that would force an AM5+/AM5b platform, unless it co-incided with DDR6 and PCIe6, in which case with such wholesale change it would probably mean a whole new AM6 platform.

CAMM is coming with Arrow Lake and prototype Z890 motherboards have already been developed, although it seems that it will be restricted to variants of the highest end boards, the DIMM format is still expected to be the primary target. Considered these 9000 series chips have the exact same IOD as the 7000 series, CAMM support on AMD is unlikely. Their maximum frequencies attainable are below the limit of the DIMM standard as it is, so even if CAMM X870E boards are developed, you aren't going to benefit from it any.

 

[mkppo]

It's planned - on AM4, same IO Die was used on Zen2 and Zen3.
I suspect AMD have plans for a newer IO die where some new requirement forces it, such as USB or PCIe spec changes, or maybe improved DDR5 performance - that's one area I think that would be of interest to some especially if some of the lack of performance improvements of Zen4>5 are bandwidth starvation.
Wouldn't necessarily require a new socket for those features, but likely motherboards with new chipsets that support the required newer signalling, kind of like AM2>AM2+, AM3>AM3+ where different hyper-transport speeds were supported, etc., so maybe there could be an AM5+ which still offers some backwards compatibility if such a thing happened.

Along the memory limitation lines, if the desktop moved to CAMM memory, one CAMM module provides 2x64-bit memory channel access, so it would be possible to move to quad channel on the desktop with a twin-CAMM motherboard (maybe one CAMM on each side). That would need more socket pins and a new IO die, but the flexibility of the chiplet design means that the CCDs could remain unchanged. At this point in time, that's probably the most likely thing that would force an AM5+/AM5b platform, unless it co-incided with DDR6 and PCIe6, in which case with such wholesale change it would probably mean a whole new AM6 platform.

Regardless of DDR6 or other spec changes, they need a new IO die which is rumoured to come with Zen 6 because from what it looks like they are extremely beefed up on the front end/execution side but bottlenecked at the back end/IO. The latter makes a good deal of difference in games.

If AMD wanted, they could've spent less time and money and just created a new IO die, fab Zen 4 on N4P and it would pretty much result in much higher gains in games. That's not where the money is though and they have X3D to somewhat 'mitigate' the memory side issues and bring gaming performance in line.

 

[Vincero]

Along the memory limitation lines, if the desktop moved to CAMM memory, one CAMM module provides 2x64-bit memory channel access, so it would be possible to move to 'quad channel' on the desktop with a twin-CAMM motherboard (maybe one CAMM on each side). That would need more socket pins and a new IO die, but the flexibility of the chiplet design means that the CCDs could remain unchanged. At this point in time, that's probably the most likely thing that would force an AM5+/AM5b platform, unless it co-incided with DDR6 and PCIe6, in which case with such wholesale change it would probably mean a whole new AM6 platform.

CAMM is coming with Arrow Lake and prototype Z890 motherboards have already been developed, although it seems that it will be restricted to variants of the highest end boards, the DIMM format is still expected to be the primary target. Considered these 9000 series chips have the exact same IOD as the 7000 series, CAMM support on AMD is unlikely. Their maximum frequencies attainable are below the limit of the DIMM standard as it is, so even if CAMM X870E boards are developed, you aren't going to benefit from it any.

Kind of missing the point of that post - that being 'hypothethical reasons for a new IO die' - if they pushed to a quad-channel on the desktop it would need a new socket and IO die. It's purely an example - I'm sure they have a roadmap, and at this point CAMM vs DIMM is currently just packaging standards with very little electrical differences (although the idea is to benefit from shorter traces between chip and memory module, so that's a knock-on electrical benefit).
Also, the memory speed limits would be potentially mitigated with quad-channel capability (although yes, that always comes with some latency tradeoff and will not really offer 2x the bandwidth but then neither does frequency increases due to latency increases).
Yes, I'm sure a new IO die will be coming with Zen6... at least if they want to sell anything that has multi-CCDs and more than 8-cores as the current performance bottleneck would only get worse.

Until we see the same 9xxx Zen5 cores on a server product being tested, we will not know for sure just how much additional bandwidth will help out core performance as core counts scale - all we can see right now is how it is on the desktop AM5 platform as it is with 9700X vs 9950X, and the answer in this scenario is 'not great'...