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

Page 368 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Abwx

Lifer
Apr 2, 2011
11,157
3,855
136
You mean like almost every on prem enterprise customer. Licensing models heavily prioritise per core performance.
That s not the point that is discussed here, get back to DrMrLordX post, and even then your point was adressed in previous posts.


 

MangoX

Senior member
Feb 13, 2001
567
60
91
I use a 5600x. I'm not overestimating anything.

There's different tiers of buyers and they have different reasons. Those who want a 16 core are rarely doing it for the fun of it, as Igor said, they either don't care how much they spend or are professionals who will spend highly hoping to work faster. It is you who think that because you consider CPUs based on perf/price, others do. Most others who will splurge on a 14900k or a 7950x do so because they expect huge perf or because they expect returns.
Agreed. I built four 7950x systems. In the end still cheaper than a Epyc/TR and much faster overall when the encoding fps is combined vs a single 64 core machine because of faster clocks.
 

DrMrLordX

Lifer
Apr 27, 2000
21,779
11,110
136
That s based on urban legends
What urban legends? At least for Milan-X and Genoa-X, they shared identical CCDs with Vermeer-X and Raphael-X (respectively). The server parts launched first, and presumably the CCDs all came from the same wafers. Do you really think Turin-X will be any different?
 

MangoX

Senior member
Feb 13, 2001
567
60
91
You will be faced with four times the temptation to upgrade them to 9950X :p
That was all in the plan bro. Eagerly waiting for Zen 5, I'm lurking and reading. Not even looking at Intel. This "farm" can encode 2000~ 24min videos to AV1 1080p per day. A 30%+ MT perf increase at similar power? Dang I'm drooling. These beasts all do min 5.1Ghz all core MT encoding 24/7.
 

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,728
14,758
136
That was all in the plan bro. Eagerly waiting for Zen 5, I'm lurking and reading. Not even looking at Intel. This "farm" can encode 2000~ 24min videos to AV1 1080p per day. A 30%+ MT perf increase at similar power? Dang I'm drooling. These beasts all do min 5.1Ghz all core MT encoding 24/7.
Their Genoa counterparts do a minimum of 3.5 ghz all-core for 64 cores. Thats pretty killer for a server chip. Oh, and they have 12 channel memory, not 2 channel, so that could make a difference. Try one.
 

MangoX

Senior member
Feb 13, 2001
567
60
91
Their Genoa counterparts do a minimum of 3.5 ghz all-core for 64 cores. Thats pretty killer for a server chip. Oh, and they have 12 channel memory, not 2 channel, so that could make a difference. Try one.
I feel you, and respect you for what you do. That's just not what I'm looking for. The platform costs alone for Epyc/TR are insane compared to AM5. I'm good with 16C/32T, 64GB memory per machine. Tried to do 145ppt power limit. Max all core 4.7ish, 145w package power. But I need speed, so I run these PBO out of the box, 360m AIO keeps them cool, at 95C :D. They so fast makes my 5800x look stupid, even when comparing half core count. They way more than 2x faster.

Now EPYC: $1k mobos, 12 mem stick requirements for a single server processor? I don't see how that's worth it for me. Sure I know you can get ES samples for cheaper on eBay, but that's not something I'm into, or willing to take a risk with. I buy either OEM or retail.

If Zen5 can deliver, then perfect! A easy drop in upgrade for these machines. I can't wait. Even a 20% boost is good. 30%+ even better. Bring it AMD.
 

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,728
14,758
136
I feel you, and respect you for what you do. That's just not what I'm looking for. The platform costs alone for Epyc/TR are insane compared to AM5. I'm good with 16C/32T, 64GB memory per machine. Tried to do 145ppt power limit. Max all core 4.7ish, 145w package power. But I need speed, so I run these PBO out of the box, 360m AIO keeps them cool, at 95C :D. They so fast makes my 5800x look stupid, even when comparing half core count. They way more than 2x faster.

Now EPYC: $1k mobos, 12 mem stick requirements for a single server processor? I don't see how that's worth it for me. Sure I know you can get ES samples for cheaper on eBay, but that's not something I'm into, or willing to take a risk with. I buy either OEM or retail.

If Zen5 can deliver, then perfect! A easy drop in upgrade for these machines. I can't wait. Even a 20% boost is good. 30%+ even better. Bring it AMD.
Just in case you change your mind..... EPYC mobo $715, memory, was $720, and last 9554 I got (best price on ebay) was $1200, and that was quite a while ago, but I now see they have gotten popular, so $2500.
 

Abwx

Lifer
Apr 2, 2011
11,157
3,855
136
What urban legends? At least for Milan-X and Genoa-X, they shared identical CCDs with Vermeer-X and Raphael-X (respectively). The server parts launched first, and presumably the CCDs all came from the same wafers. Do you really think Turin-X will be any different?

That contradict your saying that X3D chips are second rate, the first manufactured dies are not forcible the most performing frequency wise, beside they launched servers parts first because that s a much more profitable market.
 

Ghostsonplanets

Senior member
Mar 1, 2024
513
906
96
He is remarkable terse on this important issue...

View attachment 96916

But maybe we'll get some tidbits sooner rather than later...
Chinese netizen said they have STH in hands. But they don't have official firmware/drivers support yet, so performance is a bit lackluster.

A lot of rumors that come this way are basically rehashed from China forums, social media, etc.
 

soresu

Platinum Member
Dec 19, 2014
2,918
2,136
136
Might make for a decent Steam Deck 2 SoC though, assuming the GPU clocks can improve substantially without a big power add.
Van Gogh is already 8 CU though, so it would have to be a major clock boost which is unlikely to come economically to power.

The Deck has been popular enough to warrant another semi custom chip.

We might just see a 12 CU RDNA4 design.
 

Mahboi

Senior member
Apr 4, 2024
741
1,316
96
You just described running a Steam Deck at 540p.
Tangent following that thought, the Deck is truly an impressive thing, Valve may deserve to use the phrase "Bespoke Hardware" far more than Nuvidia and its silly Tensor Cores that apparently do everything cool in the GPU despite being just a matmul engine.
They cleaned their Z2/RDNA2 APU enough that they got a 10%~ battery bonus off the base model, then ran that for about 2 years before shrinking it, optimising every bit of the hardware, and releasing something that was fixing every single last little thing wrong. In just 2 years.

They passed on Z4/RDNA 3, especially with the zillion clones that were coming out that I won't bother naming. Almost all of them were just mercenary designs with the same PHX internals. Little customisation effort, and the Deck was still a viable baseline, to the point that the upgraded 6nm Z2/RDNA2 Deck was competitive with the Z4/RDNA 3 models in almost every field.

But I wonder if Z5 isn't going to be the occasion to do the next jump. I wouldn't be surprised if they wait for Z6 and possibly some LPCAMM2 option, but I wonder just how much they'll be willing to hold on to the Deck in the first place, it's been refreshed and will easily last till 2025 but if they wait until Z6 comes out, I don't expect a Deck evolution to come out WITH Z6, rather a year later at least, so that'd take us into late 2026 or 2027.

Strix Point is probably going to be a solid base for a Deck 2, especially since the original one managed to do some 900p or 720p on a small screen and pretty much everyone accepted the resolution. If they jump from a custom RDNA2/Z2 APU that's PS5 era tech to a RDNA 3.5/Z5, and go full 1080p, that'd be very impressive already.
But considering Valve, I wonder if they wouldn't try something more interesting, like trying to get RDNA 4 and upgraded RT in there. Doubtful, but not impossible. Or cutting down the Strix Point APU, or using Kraken instead. I wouldn't be surprised at some "best of both worlds" option, something like 2 Z5 + 4 Z5c or 4/4 and 12 RDNA 3.5 CUs. Would basically cover their exact needs, but I wonder if they can get something this custom from AMD.
 
  • Like
Reactions: Tlh97 and Vattila