Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)

[DisEnchantment]

Speculate at will

 

[Joe NYC]

BTW, is Kraken going to have the same NPU?

 

[Timorous]

What it is interesting is that no test used for the IPC increase over Zen4 is 1t. They are all multiple threads, of course there are lower threaded tests (like FC6 - which has a strange behaviour with AMD) and higher threaded tests.
What is "strange" in all that is, multi-threaded performance increase is exactly where I'd put it (and I wrote it earlier in this thread that I expected around 10-15% in that) because nowadays it is mostly limited by thermals and power, which are also consequences of the production process. n-thread is exactly where the N4 to N5 process jump allows it to be. Single thread, I think we have to see the specific benchmarking. But all in all, the performance increase is not stellar, so in the desktop arena competition will continue as today. Server is another matter.

I can't sit at a computer right now and do the sums but those gaming tests Vs the 14900K eyeball to me that if compared to the 7950X the gaming uplift would be larger than 16% (which is pretty much the most important metric for the bulk of DIY sales).

I figured that the geomean IPC would be lower than 32% even if 32% was true because it would be an NT IPC comparison like prior reveals but the delta does seem too wide a gap to bridge. I am also assuming that 16% is about where Specint 2017 NT IPC would land, entirely possible AMD did not fit the geomean to spec but that would be a change in how they have done things prior.

 

[SteinFG]

Zen 5 was a jump on the next curve of progress. It could have gone better, it could have gone worse. I think It's normal that they got 16%. The next question is - can they iterate effectively upon that? If Zen 6 can deliver good PPC increase, it means they did something right with Zen 5.

 

[Gideon]

real world performance is always better than synthetic and potentially manipulated benchmarks

open bench is a joke because its real? man, we have user submitted benchmark threads here on the forums, those would easily be as legit... Are you thinking of userbenchmark? the site that is very not legit?

It is not a joke, but it most certainly is a poor substitute for real world performance. Try to run actual code on a real macbook to see.

I have a M1 Pro 14" Macbook as my primary work computer and a T14s with a Ryzen 6850U as a secondary one with 32GB LPDDR5 (Really wanted Zen 4 but it just wasn't t available at that time)

Spoiler: Proofz

Sorry for the dust, both have been in use for quite a while

I use both of them daily to write BE and FE code and run tests (and both unit- / and E2E tests are particularly good MT workloads to compare). I also have a 5800X3D at home and have played around with my fathers 7700X, a coworkers 12600K's, etc.

I'd say Apple's devices actually tend to run a bit better in any real world workloads (i wouldn't really consider Phoronix benches these) despite the sluggishness of MacOS (personal view). This is probably due to loads on memory bandwidth, and SLC on top of the wide cores. Everything javascript just loooves the high IPC cores and is a bit snappier against competition that might perform similar in all-core loads (I'd say the M1 pro has a more similar experience to my undervolted and memory tuned 5800X3D, rather than the 6850U)

 

[HurleyBird]

Zen 5 was a jump on the next curve of progress. It could have gone better, it could have gone worse. I think It's normal that they got 16%. The next question is - can they iterate effectively upon that? If Zen 6 can deliver good PPC increase, it means they did something right with Zen 5.

Nah, it's a miss. Not a big one, but a definite miss... at least for int. Remember that clocks are going down slightly despite the better node. On the same node as Zen 4 at nt fmax you're probably looking at ~10% best case.

 

[adroc_thurston]

Remember that clocks are going down slightly despite the better node.

They aren't?

 

[HurleyBird]

They aren't?

200MHz lower base @ 16c @ same tdp.

 

[adroc_thurston]

200MHz lower base @ 16c @ same tdp.

Base clock is a meme on AMD parts and we're talking 1t stuff anyway.

 

[HurleyBird]

we're talking 1t stuff anyway.

at nt fmax

 

[SteinFG]

200MHz lower base @ 16c @ same tdp.

That may be due to architectural differences, not node. They doubled AVX512 throuput, which probably causes 200MHz drop under such workloads.

 

[CouncilorIrissa]

Hopefully benchmark scores will start to leak out soon.

 

[SteinFG]

By the way, confirmed that they're using chipset lanes for USB4, not CPU (which was already known but now confirmed)
edit: hmm, maybe I got the wrong conclusion from this info. not sure

 

[HurleyBird]

That may be due to architectural differences, not node. They doubled AVX512 throuput, which probably causes 200MHz drop under such workloads.

I suppose that's possible, although Intel has already set the precedent of letting AVX loads go below base clock.

 

[DrMrLordX]

Back to the the slide rule or abacus I suppose.

Imma count with mah fingers. How long does it take to compile the Linux kernel that way?!?

It was also the only Zen to be on time so that was nice.

u wot m8

eh nevermind, that was years ago and not worth fighting about now. Needless to say AMD's launch cadence has been slowing down since Matisse. The most rapid release was Summit Ridge -> Pinnacle Ridge, even though some disregard it as not being important since that CCD never even made it into proper EPYCs.

That can still be made up in later generations?

Might be some low-hanging fruit up for grabs in Zen6 in the form of bug fixes. The real question is why OEMs got told what they did.

The problem with OpenBenchmarking is that people misuse it: wrong compiler flags, obsolete versions of SW (for instance some of the benchmarks used to be only optimized for x86 intrinsics/assembly and only got Arm support recently), etc. I basically don't trust the database of results. But I agree the width of applications is great and much more pertinent than other benchmarks such as Geekbench and SPEC.

Over time, the data should improve once people start running newer versions of software and once ARM desktop/workstation platforms start getting better support.

 

[tsamolotoff]

It seems like the 9950X will still be quite a bit slower than the X3D zen 4s for gaming.

How so? according to these barcharts, it should be a bit faster than the current x3d. If anything, it looks like in gaming zen5 is around 25-30% faster than vanilla z4 (if we take the same-ish tests of HWU (xmp 6000, stock limits and good cooling solutions) -

)

 

[SteinFG]

How so? according to these barcharts, it should be a bit faster than the current x3d. If anything, it looks like in gaming zen5 is around 25-30% faster than vanilla z4 (if we take the same-ish tests of HWU (xmp 6000, stock limits and good cooling solutions)

Eh, It's better to just wait for independent results, but I agree that Zen5 seems to be better than Zen4x3d, but I only looked at the games that AMD showed on their keynote.

 

[surasak]

I feel like people are treating clock increase like “free performance”, that designer and just change the knob without doing any work, or just rely on tsmc to do the work.

My understanding is that that’s totally not the case. In order to support higher clock the cpu must be designed differently to benefit from that. Intel famously went for extremely long pipeline during P4 to get higher clock, that backfired. Being able to maintain IPC at higher clock is an achievement that we should be celebrating more.

 

[SteinFG]

By the way, confirmed that they're using chipset lanes for USB4, not CPU (which was already known but now confirmed)
edit: hmm, maybe I got the wrong conclusion from this info. not sure
View attachment 100310

Windows 10 still supported btw, probably the last AMD series that does this.

 

[igor_kavinski]

But at that level of performance, unless there's a strong efficiency improvement or very strong AVX-512 performance, I'll wait for X3D.

Professor! You game too???

 

[biostud]

Hopefully this means that Zen6 will also be on the AM5 platform. To me there is no reason to upgrade from Zen4 to Zen5.

 

[Timorous]

How so? according to these barcharts, it should be a bit faster than the current x3d. If anything, it looks like in gaming zen5 is around 25-30% faster than vanilla z4 (if we take the same-ish tests of HWU (xmp 6000, stock limits and good cooling solutions) -

)

The 6 game simple average on the AMD slides is 13% faster than the 14900k.

The 12 game average has the 7800X3D ahead of the 14900k by 7.5%.

This would indicate that yes Zen5 is slightly faster than Zen4 X3D in games and the 9950X is around 25% faster in gaming than the 7950X.

The caveat here is that in the AMD slides they use Intel Default settings and HUB at that time would have used the board default, also HUB used faster ram for the intel system so I would not be surprised if HUBs 14900K is getting better scores than AMDs 14900K meaning of HUB tested Zen 5 Vs their 14900K config it would not be 13% faster.

Seems like Zen 5 and Zen 4 X3D are going to be neck and neck on average but still think Zen 4 X3D will win in sim stuff and certain other genres that are not frequently tested.

So yea my eyeball reaction was wrong and at least for gaming that performance uplift is actually pretty decent if it holds out in reviews.

A 25% gaming performance uplift with a clock regression (needs to be tested and confirmed obviously) would be a pretty decent outcome, as fun as long bar charts are in a broad mix of workloads I only care about gaming. Once the 9800X3D is out I will probably jump on board.

 

[tsamolotoff]

h, It's better to just wait for independent results, but I agree that Zen5 seems to be better than Zen4x3d, but I only looked at the games that AMD showed on their keynote.

I agree, it remains to be seen how the newer wide arch would behave itself in games and light int-heavy workloads. I just noted that it's not all doom and gloom, but I guess the N3x hype flop did not teach some people anything

 

[Nothingness]

Professor! You game too???

Just a bit. I'm more interested in not having a CPU consume too much while having 16 cores et great ST/MT performance. This eases CPU cooling and reduce the electricity bill

 

[Jayzen]

Are you suggesting that I intentionally overhyped Zen5? Are you brain damaged?

Yes. You had access to the same, seemingly official slides which predicted a 10-15% IPC increase, and instead not only propagated unsubstantiated rumors of 30-40% IPC increase, but suppressed the release of said slides by other leakers so you could push your claims, probably for clout.

At least, it sure seems that way.

Member callouts are not allowed here. Do it on Twitter or someplace else where it is.

Mod DAPUNISHER

 

[Joe NYC]

Hopefully this means that Zen6 will also be on the AM5 platform. To me there is no reason to upgrade from Zen4 to Zen5.

It seems that way, since AMD extended the time AM5 will be supported to 2027 (from 2025).

But I wonder if there is a possibility of something like Strix Halo Zen 6 launching that will start a new, higher end socket, with more memory channels supporting CAMM2 memory...