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

[DisEnchantment]

Speculate at will

 

[fastandfurious6]

New Turin instance on AWS cloud beats Genoa by 59%

View attachment 132120

AMD EPYC 9005 Brings Incredible Performance To The Cloud With Amazon M8a Benchmarks Review - Phoronix

www.phoronix.com

truly EPYC

 

[Bigos]

Btw this is with SMT disabled on both instances.

 

[StefanR5R]

Can anybody comment on the degree of performance determinism which is to be expected from AWS VMs? Also, while the article details the processor topology of m8a.8xlarge (1 NUMA node, 1 socket, 4× 32 MB L3$), the same details are not provided for m7a.8xlarge, and I haven't found respective documentation in an admittedly quick search.

Phoronix's bare metal tests have shown about ≈1.4× geomean performance of 9005 over 9004 at same core counts and Wattage. I suppose his result of ≈1.6× of M8a over M7a includes either tweaks to the VM config by AWS on top of the CPU and memory upgrade, higher socket power, VM performance variation (noisy neighbors), or a combination of those.

 

[Bigos]

From my experience working with VMs in AWS and GCP, there can be quite significant performance discrepancy of 5-10%. This is usually due to the "noisy neighbor" effect but also different topology like different amount of NUMA nodes [1] or maybe different IOD quadrants. This can be mitigated by using AWS dedicated host (and carving a portion of a host for a VM with the rest unused) or a similar solution for GCP (sole-tenant node).

The "5-10%" value is taken from memory, older hardware and a specific (proprietary) workload I was testing. It might be different on EPYC 4/5 machines and the PTS.

This should not invalidate the test results completely as the performance boost is a lot higher than the possible discrepancy.

[1] From my experience, the number of advertised NUMA nodes is the upper bound. So when a VM of a specific configuration says there are 2 NUMA nodes, it could be scheduled on actual 2 NUMA nodes or just one - which would enhance performance. In this case it seems like m8a advertises 1 NUMA node so it should never cross both, though. Not sure about m7a, but the PTS table shows "1x128GB" RAM for both so maybe it also uses 1 NUMA node.

 

[fastandfurious6]

incredible............

nothing screams more maximum power than those huge red lines

in one year... Medusa Halo Max Ultra X3D red lines will pierce through the graph and the monitor itself


get ready

 

[DrMrLordX]

@fastandfurious6

Strix Halo has a problem with fdiv at/above 32bit, is that going to be an issue?

 

[poke01]

in one year... Medusa Halo Max Ultra X3D red lines will pierce through the graph and the monitor itself

It’s more than a year… it come out in q2 2027

 

[511]

It’s more than a year… it come out in q2 2027

Q4 27 if not 28

 

[StefanR5R]

Now add the MI300A APU into the graphs — and suddenly, STXH's bars don't look so impressive anymore. :-P

(As will Medusa Halo's, once it arrives.)

 

[511]

Now add the MI300A APU into the graphs — and suddenly, STXH's bars don't look so impressive anymore. :-P

You can't put that thing in a desktop much less Lpatop

 

[StefanR5R]

Which leads us to the question how Medusa Halo is supposed to have monitor-piercing benchmark bars, while still fitting into laptops.
Will it have HBM? Certainly not.
(Wrong thread, but to my excuse, I am asking merely rhetorically.)

Edit,
apropos — is the bar graph shown above corresponding with laptop power levels? Sustained power levels even?
[Going by who it posted, it likely does not.]

Edit 2,
the source of #24,780 is https://chipsandcheese.com/p/amds-chiplet-apu-an-overview-of-strix. Folks, when you post stuff that's not your own, state your source.

The source does not say what exact power level the graph corresponds with, except giving a 75 W upper bound for the RTX 5070M, nor whether it's peak or sustained.

 

[deasd]

Krackan Point instead of Strix Point, coming to desktop, guess it's due to cost concern

https://videocardz.com/newz/amd-ryzen-9000g-series-may-debut-with-krackan-point-apus-agesa-1-2-7-0-update-hints

 

[LightningZ71]

There's enough mini-PCs out there with Strix Point to satisfy much of that market anyway. There's no world where a vanilla 9700x and a low end GPU don't pants Strix Point for similar costs. Also, it's not performant enough to make use of it's one possible advantage: having up to 128GB unified RAM.

 

[StefanR5R]

Has there been any info ever whether Krackan Point is made of one or of two core complexes?

(Edit: The early geekbench database entries of Krackan Point showed two CCXs, but folks claimed back at that time that this was merely geekbench reporting the topology incorrectly. I don't recall that anybody ever really confirmed it to be one or the other.)

 

[Josh128]

Anybody seen this yet?? Dual v-cache 9950X3D2 and 5.6GHz 9850X3d?! To combat Arrow Lake refresh??

AMD Ryzen 9 9950X3D2 CPU To Feature Dual X3D CCDs For Up To 192 MB Cache, Ryzen 7 9850X3D Gets Boosted 5.6 GHz Clocks

AMD's upcoming Ryzen 9 9950X3D2 and Ryzen 7 9805X3D CPUs have been leaked, featuring up to 192 MB cache and faster 5.6 GHz clocks.

wccftech.com

 

[Thunder 57]

I hope they release it so people can see just how pointless it is. As if there is no inter-CCD latency. What's more interesting is that if this leak is true AMD seems to have figured out how to get (near) full speed (GHz) with X3D chips.

 

[MS_AT]

Oh no, here we go again... Still if by any chance this turns out to be true, this will be the fastest x64 CPU on the market till the next gen ships. It will even be able to compete with M4 in SPECint.

 

[ToTTenTranz]

Here's the original post:

https://twitter.com/x/status/1980605233987559514

I'm honestly anticipating this. I'm holding off on it for a long needed system upgrade from my current 5900X setup.

I hope they release it so people can see just how pointless it is. As if there is no inter-CCD latency.

Its performance might surprise some people because of the core parking issues in the 9950X3D.

It could also become a champion of Mixture-of-Experts models running on GPU+CPU.

 

[Josh128]

What's more interesting is that if this leak is true AMD seems to have figured out how to get (near) full speed (GHz) with X3D chips.

100%. I think it was possible all along, but 9800X3D was a pretty early release in the Zen 5 lifecycle, used brand new design and location for the X3D die, and so they had to be conservative for safety and reliability reasons.

Now that they've had almost a year more time to R&D, they most likely:

1. Were able to get enough /divert enough good bins to allow it (remember even vanilla 9700X was limited to 5.5GHz)
2. Have more confidence in the reliability of the design at the required voltage and current for these frequencies.

X3D2 is very interesting as it has both high clock and dual X3D. Likely will be pointless over 9850X3D for gaming, but will command a new high price for the elitist technophiles.

 

[Det0x]

Anybody seen this yet?? Dual v-cache 9950X3D2 and 5.6GHz 9850X3d?! To combat Arrow Lake refresh??

AMD Ryzen 9 9950X3D2 CPU To Feature Dual X3D CCDs For Up To 192 MB Cache, Ryzen 7 9850X3D Gets Boosted 5.6 GHz Clocks

AMD's upcoming Ryzen 9 9950X3D2 and Ryzen 7 9805X3D CPUs have been leaked, featuring up to 192 MB cache and faster 5.6 GHz clocks.

wccftech.com

View attachment 132352

Wow new toys incoming !?
Can be fun to tryout together with new a-die 24GB sticks

https://twitter.com/x/status/1980583152486871438

 

[eek2121]

Here's the original post:

https://twitter.com/x/status/1980605233987559514

I'm honestly anticipating this. I'm holding off on it for a long needed system upgrade from my current 5900X setup.




Its performance might surprise some people because of the core parking issues in the 9950X3D.

It could also become a champion of Mixture-of-Experts models running on GPU+CPU.

The rumor that just won’t die! 🤣

If it comes out, I’ll buy it.

 

[Joe NYC]

Has there been any info ever whether Krackan Point is made of one or of two core complexes?

I was wondering the same thing but have not seen the question answered.

 

[Joe NYC]

Anybody seen this yet?? Dual v-cache 9950X3D2 and 5.6GHz 9850X3d?! To combat Arrow Lake refresh??

AMD Ryzen 9 9950X3D2 CPU To Feature Dual X3D CCDs For Up To 192 MB Cache, Ryzen 7 9850X3D Gets Boosted 5.6 GHz Clocks

AMD's upcoming Ryzen 9 9950X3D2 and Ryzen 7 9805X3D CPUs have been leaked, featuring up to 192 MB cache and faster 5.6 GHz clocks.

wccftech.com

View attachment 132352

If these CPUs are released, the doubters who doubt the benefit of V-Cache across the board (such as @adroc_thurston ) will finally be convinced - hopefully.

V-Cache CPUs that run at a clock speed deficit of 200-500 MHz first have to make up the clock speed difference and then surpass the performance of CPUs clocked much higher.

So, with these CPUs instead of first subtracting and then adding, you would be only adding. The apps (that were said to not benefit from V-Cache) were just more sensitive to subtracting than to adding. With no more subtracting, these apps will also benefit.

 

[Joe NYC]

I hope they release it so people can see just how pointless it is. As if there is no inter-CCD latency. What's more interesting is that if this leak is true AMD seems to have figured out how to get (near) full speed (GHz) with X3D chips.

There are plenty of workstation / server apps that are note noticeably penalized by multiple CCDs.

Gains in gaming of 9950x3d2 will likely be minimal (vs 9850x3d), but in other apps, like those that Phoronix typically tests, we will see benefits.

 

[Thunder 57]

There are plenty of workstation / server apps that are note noticeably penalized by multiple CCDs.

Gains in gaming of 9950x3d2 will likely be minimal (vs 9850x3d), but in other apps, like those that Phoronix typically tests, we will see benefits.

The benefits seem to be rare enough to not be worth it as it looks like AMD stopped putting 3D cache on Epyc with Zen 5.