zen 4 AMD Ryzen 7000X3D series coming February/April, 16-core Ryzen 9 7950X3D features 144MB cache: 21-30% higher gaming performance at 1080p (Update #3)

[Shimejii]

This entire page seems to be bantering about Date formatting that just seems irrelevant to the topic lol. Its how this stuff has been done if its launched in the US, You get the US Format. If its launched in the EU or Asia that use that format, they use that. Why is this surprising? 

 

699 for 7950x3D pretty surprising, thought it would be 800$ minimum. Nailed the 7800x3d price and 7900x one tho.

 

Why a month and a half delay though? Seems so dumb

[bcredeur97]

Why does the extra cache increase performance so much? It's almost clear the lack of cache is a significant bottleneck on CPU's

[leadeater]

6 minutes ago, bcredeur97 said:

Why does the extra cache increase performance so much? It's almost clear the lack of cache is a significant bottleneck on CPU's

Reduces pipeline stalls waiting for data. Cache is faster latency wise than system memory of course but to access system memory requires X number of clock cycles and that pulls it in to L3 and maybe also L2 cache at the same time, then cache access requires X number of clock cycles etc.

 

Way over simplified but for each layer up you go in cache and memory access the more stages and the more wait cycles are required and the more often CPU execution resources could be idle or under filled with work. That's why a lot of effort goes in to branch predictors and other cache optimizations so the required data is present when requested avoiding more costly data requests.

[bcredeur97]

2 minutes ago, leadeater said:

Reduces pipeline stalls waiting for data. Cache is faster latency wise than system memory of course but to access system memory requires X number of clock cycles and that pulls it in to L3 and maybe also L2 cache at the same time, then cache access requires X number of clock cycles etc.

 

Way over simplified but for each layer up you go in cache and memory access the more stages and the more wait cycles are required and the more often CPU execution resources could be idle or under filled with work. That's why a lot of effort goes in to branch predictors and other cache optimizations so the required data is present when requested avoiding more costly data requests.

That makes a lot of sense, thank you. 

Bring on the cache wars! 1GB of L3 cache or bust

[leadeater]

37 minutes ago, bcredeur97 said:

That makes a lot of sense, thank you. 

Bring on the cache wars! 1GB of L3 cache or bust

EPYC Genoa-X will have about 1GB 

[BiG StroOnZ]

Interesting:

 

 

[starsmine]

2 hours ago, bcredeur97 said:

That makes a lot of sense, thank you. 

Bring on the cache wars! 1GB of L3 cache or bust

awk part of the cache wars happening today is sram scaling is dead. 
There is ZERO scaling in sram density going from TSMC 5nm to TSMC 3nm, which is an oof when cost per mm^2 went up 33%. 

AMD is able to do these sram cache tricks because of chiplets using older nodes (n6, a derivative of n7) otherwise you will hit cost and retical limits due to the size of the chips. Intel is starting to go tile for this reason. And they can pull this trick off with navi31 and 32 MCD as well, even if it never comes out this gen, its there for them to learn from for next generation.

Which is a big reason why I personally am worried about nvidia, then went ham on cache going from samsung 8nm to tsmc 5(4). I know the architecture after blackwell is supposed to be chiplet/tile based. But Blackwell information on this has been tight lipped. With their costs using so much n5 already being a pain point, a monolithic blackwell sounds terrifying to me. (98MB on AD102)


https://fuse.wikichip.org/news/7343/iedm-2022-did-we-just-witness-the-death-of-sram/

Its chiplet or bust for all the companies now, for purely sram reasons. and because of the death of scaling sram, costs every gen will go up just because you need more cache to feed faster and faster logic, meaning more and more silicon aka total die size going up across the board. 

Analog scaling also died going from 7nm to 5nm (what is used in memory controllers)

[porina]

8 hours ago, bcredeur97 said:

Bring on the cache wars! 1GB of L3 cache or bust

L3 cache is one part of the overall performance consideration, which is feeding the right data to execution units at the right time. Chip designers will optimise L1, L2, L3, (sometimes L4), and ram configuration. For example, if ram was a LOT faster than it typically is now, designers may tend towards less cache.

 

5 hours ago, starsmine said:

Its chiplet or bust for all the companies now, for purely sram reasons. and because of the death of scaling sram, costs every gen will go up just because you need more cache to feed faster and faster logic, meaning more and more silicon aka total die size going up across the board. 

It'll be interesting to see how that goes. RDNA3 having cache/MC chiplets is an interesting way to deliver that. It isn't so clear cut with AMD CPUs. V-cache wont necessarily be cost effective for lower CPUs which I see as remaining monolithic for a long time to come. Crossing silicon will have electrical (power, latency) and manufacturing costs associated with it.

[leadeater]

6 hours ago, starsmine said:

AMD is able to do these sram cache tricks because of chiplets using older nodes (n6, a derivative of n7) otherwise you will hit cost and retical limits due to the size of the chips.

Also the SRAM in the V-Cache dies is denser than in the CPU die as they are able to layout differently due to not having logic and connectivity logic throughout the die. It's just SRAM and TSV. AMD did say the density gains but I can't remember

[GOTSpectrum]

I think what we can say here is AMD has done something incredible. They have once again brought about a revolution in the silicon/x86 space using a fraction of the budget as Intel. They brought about multicore, they brought about AMD64 the 64 bit extension to x86, they made the sub 100USD Quad-Core, they brought about MCM and now as an extension of that stacked cache... 

[Delicieuxz]

On 1/4/2023 at 5:10 PM, BiG StroOnZ said:

 

 

 

Would the 7700X compare more favourably to the 7800X3D than the 5800X3D in that image, or basically identical (or even worse)?

 

How might the 7800X3D's gains scale at 1440p?

[mdk777]

5 minutes ago, Delicieuxz said:

 

Would the 7700X compare more favourably to the 7800X3D than the 5800X3D in that image, or basically identical (or even worse)?

 

How might the 7800X3D's gains scale at 1440p?

Complex sentence there, but I think you are asking;

 

In gaming the 7700X is about equal to or a little worse than the 5800x3d.

 

This of course meant that paying for a new CPU, MB and ddr5 memory got you very little for your money as compared with simply dropping the 5800x3d into your existing rig.

 

RE scaling, as you go up the scaling is dependent on how GPU bound you are.

The higher the resolution, the more likely you are to be GPU bound rather than CPU bound.

[BiG StroOnZ]

5 minutes ago, Delicieuxz said:

Would the 7700X compare more favourably to the 7800X3D than the 5800X3D in that image, or basically identical (or even worse)?

 

How might the 7800X3D's gains scale at 1440p?

 

According to the benchmarks @ 1440p, the 7700X performs minorly better on average than a 5800X3D (about 1%):

 

 

At 1080p it seems the 7700X is a little faster than the 5800X3D (about 7%):

 

 

It's pretty close though, so I imagine any gains seen with the 5800X3D to the 7800X3D will be basically identical with a 7700X to 7800X3D. You will definitely see better results at 1080p compared to 1440p (where 1440p is more GPU bound) but probably not that much.

 

With the 5800X3D vs the 5800X @ 1440p the 5800X3D was on average about 7% faster:

 

 

Compared to at 1080p the 5800X3D was about 8% faster than the 5800X:

 

 

Obviously, on average performance/relative performance paints a different picture; as the 3D V-Cache benefit will be game dependent (as even seen from AMD's own slides). 

[Delicieuxz]

18 minutes ago, BiG StroOnZ said:

 

According to the benchmarks @ 1440p, the 7700X performs minorly better on average than a 5800X3D (about 1%):

 

 

At 1080p it seems the 7700X is a little faster than the 5800X3D (about 7%):

 

 

It's pretty close though, so I imagine any gains seen with the 5800X3D to the 7800X3D will be basically identical with a 7700X to 7800X3D. You will definitely see better results at 1080p compared to 1440p (where 1440p is more GPU bound) but probably not that much.

 

If 7700X vs 5800X3D at 1080p is 7% apart, while at 1440p is 1% apart, is it reasonable to think that AMD's 7800X3D slides would also show a significant contraction in relative performance at 1440p, if they were being compared to a 7700X instead of the 5800X3D?

[BiG StroOnZ]

2 minutes ago, Delicieuxz said:

If 7700X vs 5800X3D at 1440p is 1% apart while at 1080p is 7% apart, is it reasonable to think that AMD's 7800X3D slides would also show a significant contraction in relative performance at 1440p, if they were being compared to a 7700X instead of the 5800X3D?

 

Yes, if they were being compared to a 7700X at 1440p, there would be a reduction in overall performance gain seen. I think that's why they used 1080p and a 5800X3D instead. To show higher performance uplift.

[Deadpool2onBlu-Ray]

I am actually a bit underwhelmed TBH, given those slides. I might just skip the Zen 4 x3d entirely and wait for Zen 5 x3d.

[starsmine]

20 minutes ago, ShawtyT30beTHICCC said:

I am actually a bit underwhelmed TBH, given those slides. I might just skip the Zen 4 x3d entirely and wait for Zen 5 x3d.

when I see posts like this, I have to wonder, what were you expecting? Because the only new news here is specific release date and prices, but even then we knew ball park before (Q1 launch, +100ish dollars) since before the retail launch of Zen 4. Which you had to have been paying attention to given your build. What were your priors?

[Deadpool2onBlu-Ray]

18 hours ago, starsmine said:

when I see posts like this, I have to wonder, what were you expecting? Because the only new news here is specific release date and prices, but even then we knew ball park before (Q1 launch, +100ish dollars) since before the retail launch of Zen 4. Which you had to have been paying attention to given your build. What were your priors?

The few slides they provided on performance compared to the 5800x3d. It looks like the 7800x3d is going to be around 12-15% faster in a base case scenario compared to the 7700x at 1440p

[starsmine]

5 hours ago, ShawtyT30beTHICCC said:

The few slides they provided on performance compared to the 5800x3d. It looks like the 7800x3d is going to be around 12-15% faster in a base case scenario compared to the 7700x at 1440p

I feel like you cant expect more. Its the exact same architecture, and running at lower clocks.
So a 5800x -> 5800x3d is a fair thing to have compared it to. 

But yea, its not worth it if you are already zen 4.

I bought this platform to wait for zen5x3d or zen6x3d and run it for a decade.

[Deadpool2onBlu-Ray]

1 minute ago, starsmine said:

I feel like you cant expect more. Its the exact same architecture, and running at lower clocks.
So a 5800x -> 5800x3d is a fair thing to have compared it to. 

But yea, its not worth it if you are already zen 4.

I bought this platform to wait for zen5x3d or zen6x3d and run it for a decade.

Yeah I just really thought this time, with the 2nd gen 3d cache, the clock speed issues would be negated and the jump from 7700x to 7800x3d would be greater than 5800x to 5800x3d. I know official benchmarks aren't out, but that isn't looking like the case. It's just mildly disappointing. I guess it saves me a few bucks and I'll ride this 7700x for a few years then throw a 9800x3d or whatever it'll be called in there.

[Fendrick]

13 hours ago, starsmine said:

I feel like you cant expect more. Its the exact same architecture, and running at lower clocks.
So a 5800x -> 5800x3d is a fair thing to have compared it to. 

But yea, its not worth it if you are already zen 4.

I bought this platform to wait for zen5x3d or zen6x3d and run it for a decade.

It's not worth it for 5800 X owners either.

[BiG StroOnZ]

Looks like you will gain some granular control over the CCDs with 7000X3D CPUs ~

 

Beta BIOS Enables Prioritization of CCDs on AMD 7000X3D Processors:

 

 

 

Quote

A new Asus X670 Beta BIOS has advanced settings that enable you to control the X3D cores. The screen shots shared by Twitter user HXL (@9550pro) show a range of algorithm conditions and actions that can activate the correct CCD core, which is OS agnostic since it is built into the BIOS.

 

This report suggests that many, if not all, X670 motherboards, including some of the top motherboards, will soon receive BIOS updates that enable the fine-tuning of which workloads are assigned to which CCD. Possibly other chipsets, such as B650, will get updated also.

 

https://www.guru3d.com/news-story/beta-bios-enables-prioritization-of-ccds-on-amd-7000x3d-processors.html