Info 64MB V-Cache on 5XXX Zen3 Average +15% in Games

Page 130 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

nicalandia

Golden Member
Jan 10, 2019
1,190
1,291
106
Why would you include a 16-core Epyc? They are not comparable at all.
I believe it was a valid comparison to make a point. An 5950X3D would have performed about the same than a 3D V-Cache EPYC 7373X, it would have outperformed both a 16 Core Milan and the 5850X.

Here is an example. The 72F3 8C/16T Zen3 EPYC has 256 MiB of Total L3 Cache. It has a High OC to 4.1 and it gets last place in this comparison.


1651154952438.png

1651156444568.png
 

jamescox

Senior member
Nov 11, 2009
506
823
136
I believe it was a valid comparison to make a point. An 5950X3D would have performed about the same than a 3D V-Cache EPYC 7373X, it would have outperformed both a 16 Core Milan and the 5850X.

Here is an example. The 72F3 8C/16T Zen3 EPYC has 256 MiB of Total L3 Cache. It has a High OC to 4.1 and it gets last place in this comparison.


View attachment 60805

View attachment 60807

No. The only thing those two have in common is the number of cores. The 7373X is an 8 CCX (chiplet) * 2 cores per CCX part. Each 2 cores has 96 MB L3. 5950X is 2 CCX * 8 cores per CCX, so 8 cores per 96 MB L3 in a hypothetical v-cache part. So it is 8x2 core config vs. 2x8; that is a big difference.

To get the same core config as a 5950X, you would need one of the 7773X parts and disable 48 cores on 6 of the chiplets. It is only the 7700 parts that have all 8 cores per CCX/chiplet enabled. Still not comparable though since 7700 parts base clocks are between 2 and 2.5 GHz and you have the 8 channel memory. Even the 72F3 doesn’t clock very high and any fine grained communication between threads / process will be slow since all 8 cores have an independent 32 MB L3 (8x1 core config; see the Wikipedia article for Epyc for all of the core configs).
 
  • Like
Reactions: Makaveli and Bigos

jamescox

Senior member
Nov 11, 2009
506
823
136
I was more on the safer side when it came to sizing up the possibilities of this CPU before launch.

Am I the only one who thinks that this product ended up being such a massive boost for A LOT of AMD platform owners, whose primary focus is on gaming, requiring zero amount of exotic cooling and more often than not even less power than the CPU it'd replace, from 2000 to 5000 series..., all that at the cost of a BIOS update? And getting equal performance to a brand new, state of the art 12900K gaming rig on average?
If anything, this CPU is not even the worst for Intel... it's a nightmare for mobo, RAM and other system part manufacturers :D

I'm down to saying that AMD could produce a million of these and have them all sold at its original MSRP eventually.
We are not in the stagnant, intel monopoly era. Lots of people will want to upgrade to Zen 4 and Zen 5 with new motherboards and memory, so even though AMD kept AM4 alive a long time, I think motherboard makers might view AMD more favorably than Intel at this point.
 

lobz

Golden Member
Feb 10, 2017
1,945
2,594
136
We are not in the stagnant, intel monopoly era. Lots of people will want to upgrade to Zen 4 and Zen 5 with new motherboards and memory, so even though AMD kept AM4 alive a long time, I think motherboard makers might view AMD more favorably than Intel at this point.
That was more of a satirical sting towards intel :)
 

jamescox

Senior member
Nov 11, 2009
506
823
136
The most likely reason AMD did not bother with 5900X and 5950X are as follows

1) 5900X and 5950X are productivity 1st parts and for most desktop usecases the 3d cache has little to no benefit so why would consumers pay extra for no gain in their primary workloads.

2) The concessions required for V-Cache currently mean a 5900X3D and 5950X3D would need lower clocks and better bins than their non 3D counterparts. With point 1 this probably means a regression in productivity performance (outside of a few edge cases) as we see with the 5800X3D vs 5800X.

3) It would use 2x the silicon and it seems unlikely AMD could charge $900 for a 5950X3D with the above caveats.

4) In gaming there is very little to choose between the 5800X, 5900X and 5950X. The same would be true for the 3D version of such parts.

So in conclusion why regress productivity performance on productivity 1st parts just to catch up to intel in gaming (on average) when you can release a gaming 1st part that does the same and is a lot cheaper and uses lower quality bins for the CCD and only 1 cache die leaving more usable parts for Milan-X.

I expect Zen 4 3D to have a different voltage domain for L3 cache vs core and it is possible AMD will stack the core on top of the cache as well rather than the other way meaning cooling is less of an issue. That would mean AMD can probably release parts with fewer caveats but it will still have the same issue of no productivity performance increase but it would mean AMD could release a no compromises all use case part which a 5950X3D would not be.
The idea of putting the cpu on top of the memory for cooling sounds great, but it likely doesn’t work. I think getting ~100 W through TSVs to the cpu is the problem. I think that is difficult to do, but I don’t know what the limit is. CPUs usually have a huge number of power and ground pins to handle the current draw, which is massive at such low voltages.

Edit: just watched a video on Moore’s law is dead YouTube channel about stacked GPUs. They seem to think that the compute die is stacked completely on top of the cache / IO die. They made a bunch of renderings with that, but that seems unlikely, although the power consumption of the compute die may not be that high if the whole thing is 150 W (2 gpu chiplets, base die or bridge die, and HBM). I would expect it to use EFB with an embedded cache / bridge die under the two gpu chiplets rather than a base die. Then they would also have bridge die (more likely to be just passive bridges) to connect the HBM. I have thought that they may use the same infinity cache die with Bergamo. It would likely be L4 since it is lower connectivity than SoIC used for v-cache die.
 
Last edited:
  • Like
Reactions: lightmanek

ZGR

Golden Member
Oct 26, 2012
1,920
419
136
Was looking forward to seeing how a heavily modded build of GTA V runs on the 5800X3D. I have around 70 scripts running, 5x more AI pedestrians and traffic, and tens of thousands of additional props (mainly trees) that users have created over the years. It made the game a CPU nightmare.

I also copied my settings and save files over and using same resolution.

The game ran at about 29-31 fps on the 5775C at 2160p in front of Franklin's house. I rarely got above 35 fps in overall gameplay even if I dropped the resolution down to 1080p.

The 5800X3D is hitting around 77-78 fps in same location, with the 1080 ti nearing its limit.
 

blckgrffn

Diamond Member
May 1, 2003
8,427
1,740
136
www.teamjuchems.com
Was looking forward to seeing how a heavily modded build of GTA V runs on the 5800X3D. I have around 70 scripts running, 5x more AI pedestrians and traffic, and tens of thousands of additional props (mainly trees) that users have created over the years. It made the game a CPU nightmare.

I also copied my settings and save files over and using same resolution.

The game ran at about 29-31 fps on the 5775C at 2160p in front of Franklin's house. I rarely got above 35 fps in overall gameplay even if I dropped the resolution down to 1080p.

The 5800X3D is hitting around 77-78 fps in same location, with the 1080 ti nearing its limit.
I had to look that Broadwell part number up! Those things were completely off my radar.

Both results with the 1080ti?

Did you have a different AMD CPU before?
 

ZGR

Golden Member
Oct 26, 2012
1,920
419
136
Can you test that on a friend's 5600X/5800X? That way we can know for sure how much the V-cache is contributing to the increased fps.
The scripts folder is only 15MB total, but the mods folder is over 12GB due to all the props. If they do download it, I won't make them download the props, and I'll re run the test. But the game is really big, so it might be hard to convince them just to play a single player game.

If the props are removed, the total file size of everything gets really small at least.
 

ZGR

Golden Member
Oct 26, 2012
1,920
419
136
I had to look that Broadwell part number up! Those things were completely off my radar.

Both results with the 1080ti?

Did you have a different AMD CPU before?
Sorry I missed your reply, but yeah I used same GPU for both tests. And I did not have a Ryzen CPU before, but was able to test games on some.

edit: Congrats on those amazing results Det0x!
 
  • Like
Reactions: blckgrffn

DrMrLordX

Lifer
Apr 27, 2000
19,164
7,920
136
Am I the only one who thinks that this product ended up being such a massive boost for A LOT of AMD platform owners, whose primary focus is on gaming, requiring zero amount of exotic cooling and more often than not even less power than the CPU it'd replace, from 2000 to 5000 series..., all that at the cost of a BIOS update?
No, you aren't alone. Anyone on a 1xxx, 2xxx, or 3600/3700x/3800x with DDR4-3200 can really benefit from this CPU. Assuming their board will play ball, which fortunately a lot of them will (or will soon).
 

Timorous

Senior member
Oct 27, 2008
825
1,012
136
The idea of putting the cpu on top of the memory for cooling sounds great, but it likely doesn’t work. I think getting ~100 W through TSVs to the cpu is the problem. I think that is difficult to do, but I don’t know what the limit is. CPUs usually have a huge number of power and ground pins to handle the current draw, which is massive at such low voltages.

Edit: just watched a video on Moore’s law is dead YouTube channel about stacked GPUs. They seem to think that the compute die is stacked completely on top of the cache / IO die. They made a bunch of renderings with that, but that seems unlikely, although the power consumption of the compute die may not be that high if the whole thing is 150 W (2 gpu chiplets, base die or bridge die, and HBM). I would expect it to use EFB with an embedded cache / bridge die under the two gpu chiplets rather than a base die. Then they would also have bridge die (more likely to be just passive bridges) to connect the HBM. I have thought that they may use the same infinity cache die with Bergamo. It would likely be L4 since it is lower connectivity than SoIC used for v-cache die.
I think solving the power delivery issue is easier than solving the cooling issue.
 
  • Like
Reactions: Tlh97 and RnR_au

jamescox

Senior member
Nov 11, 2009
506
823
136
I think solving the power delivery issue is easier than solving the cooling issue.
Perhaps. The infinity cache die would actually need to be quite large, so stacking a compute (CDNA), graphics (RDNA), or cpu die on top may be reasonable. They may be too large to just be bridge chips. A 256 MB cache die could be around 150 mm2, so enough room for one gpu chiplet or 2 cpu chiplets. If they get used with Bergamo, then that would be 4 cache die for 1 GB total. I don’t know if they could go multi-layer on a bridge die implementation to get higher capacity with a small footprint.
 

blckgrffn

Diamond Member
May 1, 2003
8,427
1,740
136
www.teamjuchems.com
ha, indeed, I knew it existed and thought it was cool at launch. But actually seeing someone have it threw me for a bit of a loop. Pleb socket Broadwell was such a weird beast and release. Everyone had a 4770/4790K it seemed like and just stayed there.

I am glad this is performing well and
getting embraced and wondering if I should have gotten this rather than the 5900K I picked up here rather inexpensively (it was the CPU I had always wanted but you all remember the release woes).

All I do on this PC is game… decisions… :)
 

ASK THE COMMUNITY