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

[DisEnchantment]

Speculate at will

 

[Tuna-Fish]

That's just 50 bucks extra for 3D chiplet in 16 core version, they can easily sell 9999X3D with both binned chiplets and 3D cache for $799 and make more profit

No. Because AMD cannot spend more money today to have more chiplets with 3D cache. The lead time for increasing capacity is measured in years, plural. For all intents and purposes, for the entire Zen 5 generation they have a fixed amount of chiplets with the added cache, they already know how many there will be, and given how they are selling currently, they probably expect to sell every last one at good prices.

Then, their choice is either selling one 2-vcache part, or two 1-vcache parts. In a sense, a 9950X3D is less of a 9950X with a vcache chiplet for +$50, than it is a 9800X3D with an extra core chiplet (which they can scale production of up and down to match demand) for +$220.

To match the revenue, they would have to price it higher. $999 would probably be fair.

 

[Win2012R2]

No. Because AMD cannot spend more money today to have more chiplets with 3D cache.

It's in stock now at MSRP - that means supply already exceeds current demand, so they can make more money by selling higher end version, yes feck it - price it at $999, sell it as server EPYC then, highly clocked heterogenous predictable two chiplets, awesome.

 

[Markfw]

It's in stock now at MSRP - that means supply already exceeds current demand, so they can make more money by selling higher end version, yes feck it - price it at $999, sell it as server EPYC then, highly clocked heterogenous predictable two chiplets, awesome.

9950x3d not even at amazon, and only scalper price at newegg at almost $1000

 

[Win2012R2]

9950x3d not even at amazon, and only scalper price at newegg at almost $1000

So 9999x3D would sell for more?

 

[MS_AT]

So 9999x3D would sell for more?

Why? You know it would get lukewarm reviews. In games it wouldn't behave better than 9950x3d on average. And for applications benchmarks it would need a specific selection of these to shine, which most reviewers are not using. So it would do slightly worse due to lower boost clock and have higher asking price. Actually you would have irritated these people who wanted to buy 9950x3d as they would argue that AMD is putting all x3d CCDs on 9999x3d as it has higher price. From PR point of view it's a bad move. But I guess Igor would be happier

 

[Win2012R2]

Why?

Because as per post above 9950X3D still gets scalped at $1000

 

[Markfw]

Why? You know it would get lukewarm reviews. In games it wouldn't behave better than 9950x3d on average. And for applications benchmarks it would need a specific selection of these to shine, which most reviewers are not using. So it would do slightly worse due to lower boost clock and have higher asking price. Actually you would have irritated these people who wanted to buy 9950x3d as they would argue that AMD is putting all x3d CCDs on 9999x3d as it has higher price. From PR point of view it's a bad move. But I guess Igor would be happier

Only if the app could use 2 threads, each pinned to a CCD using lasso (for example) and total performance might be good. Very limited wins IMO

 

[leoneazzurro]

A lot of 3D chips seems to be headed for the gaming laptop market, there are several models available soon (or in some cases, already there) with the 9955HX3D (yes, the top one) like MSI Vector 18 and Raider 18, Asus ROG Strix G16. Others are expected later. Dunno how AMD will keep up with the demand, because it seems quite strong.

 

[igor_kavinski]

Actually you would have irritated these people who wanted to buy 9950x3d as they would argue that AMD is putting all x3d CCDs on 9999x3d as it has higher price. From PR point of view it's a bad move. But I guess Igor would be happier

People are still unhappy because Genoa-X is too expensive and 9950X3D production is too low. Adding one more SKU for +300 USD isn't gonna ruffle a lot of feathers but those who want one, will at least have a chance. So some good will come out of it. As for slightly worse performance, AMD can clearly communicate that the ONLY benefit to this CPU is more V-cache. Then if the youtubers find cases where the extra cache helps, it will make the part look more attractive. If not, at least those people that hate dissimilar compute domains will get peace.

 

[fastandfurious6]

yes the discussion on dual x3d goes nowhere

1. limited x3d manufacturing capacity
2. higher profits/better market to make more mobile 9955hx3d instead of desktop dual x3d
3. unclear or minimal benefits of dual x3d

 

[igor_kavinski]

AMD could also let the Chinese make relatively cheap dual socket consumer mobos so people can put two X3D chips in them. They can keep it dual channel and limit the RAM to DDR5-6000 256GB max if they are so worried about such a setup encroaching into their TR sales. That would be a great compromise to see if people really want to pay extra for more V-cache in their system, as well as more PCIe lanes.

 

[MS_AT]

Only if the app could use 2 threads, each pinned to a CCD using lasso (for example) and total performance might be good. Very limited wins IMO

Not exactly. You would need something that is topology aware and does not rely heavily on inter thread communication and has a workload that does not normally fit in L3, but does fit in x3D extended one. Streaming workloads are mostly out of question as L3 is a victim cache. Some compute heavy stuff would fit. Based on the Genoa / Milan data, all FEM things simulations, also fit the criteria. So mostly useless for average user that uses PC for games and internet browsing.

 

[igor_kavinski]

So mostly useless for average user that uses PC for games and internet browsing.

Yes but then the average user isn't the target audience for dual V-cache CCDs. The target is people who want to pay for the best but don't want to pay server or workstation prices. You can't pay extra to get Extreme Edition or KS level frequency in the AMD space. What else can you pay for? More cache!

 

[inquiss]

yes the discussion on dual x3d goes nowhere

1. limited x3d manufacturing capacity
2. higher profits/better market to make more mobile 9955hx3d instead of desktop dual x3d
3. unclear or minimal benefits of dual x3d

It goes nowhere and it should go nowhere. There's no need for this part. I expect these two to be the first in line to be disappointed when the part has minimal uplift in gaming. AMD software would still, rightly, peg the game to one CCD.

 

[inquiss]

Yes but then the average user isn't the target audience for dual V-cache CCDs. The target is people who want to pay for the best but don't want to pay server or workstation prices. You can't pay extra to get Extreme Edition or KS level frequency in the AMD space. What else can you pay for? More cache!

There is no target audience. You've stated it yourself. People who don't want to pay extra.

 

[igor_kavinski]

People who don't want to pay extra.

Middle ground between 9950X3D and TR. There's a market waiting to tapped.

Let's suppose a non-gaming workload. If I could get the dual V-cache chip installed in a cheap $170 mobo and get my compilers and browsing super accelerated, why wouldn't I want to spend extra, as long as it's a reasonable extra dollars?

 

[GTracing]

Middle ground between 9950X3D and TR. There's a market waiting to tapped.

Let's suppose a non-gaming workload. If I could get the dual V-cache chip installed in a cheap $170 mobo and get my compilers and browsing super accelerated, why wouldn't I want to spend extra, as long as it's a reasonable extra dollars?

Unless I'm looking at the wrong benchmarks, web browsing gets practically zero speedup from V-cache.

 

[MS_AT]

Let's suppose a non-gaming workload. If I could get the dual V-cache chip installed in a cheap $170 mobo and get my compilers and browsing super accelerated, why wouldn't I want to spend extra, as long as it's a reasonable extra dollars?

It boils down to what you consider reasonable. Plus the compiler benchmarks you are able to find on the net require some context. They usually test clean build performance. Something an average developer is rarely doing since you build code incrementally, for that usually higher frequency triumphs all if you have one or two files to recompile and link. But then it also depends on what you link, what you compile, on which platform where x3D can have various degree of effect. So when you can see that on llvm compilation 9950x3d gives you 6% improvement, dual x3d part might give you... regression even since you don't know the cpu load profile of the whole job. But it could also give you double digits gains under special circumstances.

 

[igor_kavinski]

Unless I'm looking at the wrong benchmarks, web browsing gets practically zero speedup from V-cache.

Spoiler

From here: https://www.techpowerup.com/forums/threads/post-your-jetstream-2-speeds.297345/page-7

Also: https://babeltechreviews.com/battle...n-7-9800x3d-an-in-depth-performance-analysis/

 

[MS_AT]

The one graph not under the spoiler is 3 percent difference, despite gigantic blue bars Since we do not know the details, it might be due to the fact that x3d CCD is higher clocked on the 9950x3d or because it landed on frequency CCD. Your other screenshot might be as well random numbers, it seems each person owned a different machine running it under different OS and possibly OS version.

 

[igor_kavinski]

Your other screenshot might be as well random numbers, it seems each person owned a different machine running it under different OS and possibly OS version.

We can test with my "thermally constrained" 9950X3D in Jetstream vs. someone's stock 9950X here and compare the subtest scores.

I'll try to post the run here tonight. If the 9950X has more wins, fine. My 9950X3D sucks

 

[MS_AT]

I'll try to post the run here tonight. If the 9950X has more wins, fine. My 9950X3D sucks

But you don't want to compare if 9950X has more wins than 9950X3D because it tells us nothing if 9999X3D would be beneficial for browser workloads compared to 9950X3D, as this what you are arguing Plus since your setup has issues it wouldn't be representative info. You could rather process lasso the benchmark to normal CCD and compare to X3D CCD if it is possible to set number of threads for the benchmarks to use, to avoid situations it would try to fit 32 workers on single CCD. This at least takes the thermals, OS and other setup issues out of the equation. If you would compare stock config (both CCDs enabled), against x3D CCD against normal CCD you could then have some basis to reason about the feasibility of 9999X3D. In that particular benchmark at least

 

[gdansk]

We can test with my "thermally constrained" 9950X3D in Jetstream vs. someone's stock 9950X here and compare the subtest scores.

I'll try to post the run here tonight. If the 9950X has more wins, fine. My 9950X3D sucks

Compare your own CCDs.
I have done so in some web benchmarks. Vcache wins some and loses others. It was not a clear win.

 

[GTracing]

Spoiler

View attachment 122076
View attachment 122077
View attachment 122078

From here: https://www.techpowerup.com/forums/threads/post-your-jetstream-2-speeds.297345/page-7

Also: https://babeltechreviews.com/battle...n-7-9800x3d-an-in-depth-performance-analysis/

View attachment 122079

That's a 2-3% difference. The gap between a 9950X3D and a CPU with 2 V-cache CCDs would be even smaller.

edit: Also, looking at Techpowerup's testing, the 9700X is faster than the 9800X3D in Jetstream, which implies that the benchmark benefits more from the slightly higher clockspeeds than it does from the large cache. A CPU with 2 V-cache CCDs would performance worse than the 9950X3D in Jetstream.

 

[igor_kavinski]

That's a 2-3% difference. The gap between a 9950X3D and a CPU with 2 V-cache CCDs would be even smaller.

That still doesn't make me want a dual V-cache CCD less. I would feel satisfied that no thread is being discriminated against by the scheduler by being put on the non-V-cache CCD.