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

[Kedas]

Well we know now how they will bridge the long wait to Zen4 on AM5 Q4 2022.
Production start for V-cache is end this year so too early for Zen4 so this is certainly coming to AM4.
+15% Lisa said is "like an entire architectural generation"

 

[Vattila]

Note that 5600X3D is reported/rumored to only add 32MB, so only half of the v cache. If true it's pretty much the dumping ground for all X3D dies where parts of the v cache broke, but at least half of it was still salvageable.

That is the only scenario that makes sense, as far as I can see. I have a hard time imagining AMD would stack perfectly good 64 MB V-Cache chiplets on top of salvaged 6-core CCDs. They must have built up such an inventory of partially faulty V-Cache chiplets that makes salvaging them worthwhile. However, that in turn means that the rumoured 12-core 5900X3D will most probably have two salvaged V-Cache chiplets stacked on top of two salvaged 6-core CCDs, for a total of 32 + 32 = 64 MB of V-Cache, and a total of 128 MB of L3 cache (64 MB on CCDs + 64 MB on V-Cache chiplets) — assuming all the salvaged V-Cache chiplets have half the capacity (i.e. 32 MB).

 

[Mopetar]

That's a pretty good use case. Heavy multi-taskers.

In theory it should be good at that. It's basically the same use case as cloud computing where different applications are being loaded and unloaded, but it's hard to quantify for a consumer. The use pattern is likely more sporadic and it's hard to say if it makes a noticeable difference. I don't think anyone even has a benchmark that aims to test something like this.

 

[Topweasel]

That is the only scenario that makes sense, as far as I can see. I have a hard time imagining AMD would stack perfectly good 64 MB V-Cache chiplets on top of salvaged 6-core CCDs. They must have built up such an inventory of partially faulty V-Cache chiplets that makes salvaging them worthwhile. However, that in turn means that the rumoured 12-core 5900X3D will most probably have two salvaged V-Cache chiplets stacked on top of two salvaged 6-core CCDs, for a total of 32 + 32 = 64 MB of V-Cache, and a total of 128 MB of L3 cache (64 MB on CCDs + 64 MB on V-Cache chiplets) — assuming all the salvaged V-Cache chiplets have half the capacity (i.e. 32 MB).

I think that makes assumptions about the manufacturing process that may or may not be the case. It's entirely possible that AMD has to make decisions on manufacturing before knowing the status of the cache. It may be why AMD didn't do one of their two die chips as part of the project. It could be a combination of them disabling the cores because they have to disable cache because of cores or vice versa.

 

[SteinFG]

That is the only scenario that makes sense, as far as I can see. I have a hard time imagining AMD would stack perfectly good 64 MB V-Cache chiplets on top of salvaged 6-core CCDs. They must have built up such an inventory of partially faulty V-Cache chiplets that makes salvaging them worthwhile. However, that in turn means that the rumoured 12-core 5900X3D will most probably have two salvaged V-Cache chiplets stacked on top of two salvaged 6-core CCDs, for a total of 32 + 32 = 64 MB of V-Cache, and a total of 128 MB of L3 cache (64 MB on CCDs + 64 MB on V-Cache chiplets) — assuming all the salvaged V-Cache chiplets have half the capacity (i.e. 32 MB).

This is unlikely. AMD uses Chip-on-Wafer for 3D cache, so they grab a whole CCD wafer and place a lot of Cache chiplets on top. With this, they can choose what cache chiplets to put on, but can't choose CCDs (since they're a whole wafer).

Honestly I can't see them releasing more X3D variants for AM4.

 

[DrMrLordX]

Why not now?

Just more upgrades for AM4 that are less likely to happen now that Raphael's release is pretty-much known.

It's going to be even more annoying if AMD launches a 7800X3D later this year and a 7900X3D in like April 2023. All the people that buy a day-one 7950X will be wondering if they should have waited . . .

 

[moinmoin]

However, that in turn means that the rumoured 12-core 5900X3D will most probably have two salvaged V-Cache chiplets stacked on top of two salvaged 6-core CCDs, for a total of 32 + 32 = 64 MB of V-Cache, and a total of 128 MB of L3 cache (64 MB on CCDs + 64 MB on V-Cache chiplets) — assuming all the salvaged V-Cache chiplets have half the capacity (i.e. 32 MB).

5900X3D is reported/rumored to contain 192MB of L3$ though, so it would be the new (and likely very last) marquee product for AM4.

 

[igor_kavinski]

All the people that buy a day-one 7950X will be wondering if they should have waited . . .

I'm guessing those type of people won't mind buying a X3D variant as an additional machine. I mean, it's not like normal people save up to buy a 5950X.

 

[DrMrLordX]

I'm guessing those type of people won't mind buying a X3D variant as an additional machine. I mean, it's not like normal people save up to buy a 5950X.

Personally I don't want both a 7950X and 7900X3D or whatever it is they call it. I'll probably want the 3D cache chip. Might have to wait a few months to see what AMD launches.

 

[nicalandia]

5900X3D is reported/rumored to contain 192MB of L3$ though, so it would be the new (and likely very last) marquee product for AM4.

The 5900X3D Prototype did have 192MiB of L3$

 

[LightningZ71]

My thoughts on this are somewhat on a different track. The 5800x3d suffers from limited boost frequency because of a combination of package thermals and total power draw. The 5600x3d, with only half the stacked cache powered and two less cores to draw power may actually outboost the 5800x3d by a nontrivial margin in the same power envelope if AMD relaxes the limits that are on the 5800x3d. It might have better performance on a few games over the 58x3d. Also, from a total cache per core perspective, it doesn't suffer as much compared to the 58x3d as it has about 10.08 MB per core compared to 12MB per.

The 5900x3d is a different animal. It's going to have a massive 16MB of L3 per core and should only loose 100-200MHZ all core boost to the 5800x3d in the same power/thermal envelope. It should have a significant MT throughput advantage over the 58x3d and might surprise against the regular 5900x. If boosting behaves properly, it should be every bit as good as the 58x3d in almost all games with sane code and modest threading.

 

[MadRat]

But 5800x3D and 5900x3D are more limited by power managememt than thermals. No amount of cooling is going to fix those limitations.

 

[Thunder 57]

My thoughts on this are somewhat on a different track. The 5800x3d suffers from limited boost frequency because of a combination of package thermals and total power draw. The 5600x3d, with only half the stacked cache powered and two less cores to draw power may actually outboost the 5800x3d by a nontrivial margin in the same power envelope if AMD relaxes the limits that are on the 5800x3d. It might have better performance on a few games over the 58x3d. Also, from a total cache per core perspective, it doesn't suffer as much compared to the 58x3d as it has about 10.08 MB per core compared to 12MB per.

The 5900x3d is a different animal. It's going to have a massive 16MB of L3 per core and should only loose 100-200MHZ all core boost to the 5800x3d in the same power/thermal envelope. It should have a significant MT throughput advantage over the 58x3d and might surprise against the regular 5900x. If boosting behaves properly, it should be every bit as good as the 58x3d in almost all games with sane code and modest threading.

I honestly don't see AMD releasing those parts.

 

[Insert_Nickname]

If boosting behaves properly, it should be every bit as good as the 58x3d in almost all games with sane code and modest threading.

Ah. That's where it get tricky...

We have seen plenty of the opposite over the years.

 

[LightningZ71]

I honestly don't see AMD releasing those parts.

I still won't believe it until either AMD makes it official or the hit the market. We know that at least a few 5900x3d parts existed for the presentation last year, and we know that they do stack 6 core CCDs, so, it's not a tall mountain to climb to release them. It's just, does it make sense financially?

 

[jpiniero]

Are there really enough partially busted vcache dies to make a product out of it? That's the part that would be a bit surprising.

 

[igor_kavinski]

I think they have some stringent criteria for enterprise V-cache dies like 1 in 10 million bit errors (just a stupid guess) during validation testing. Anything that fails that might still be good enough for consumer workloads. No one cares if some user gets a weird spreadsheet error or a random crash in some game once a year but the same thing happening to a critical enterprise workload could mean a loss in the millions. One reason enterprise chips are expensive is coz they are probably top 1% of the highest bins.

 

[nicalandia]

I think they have some stringent criteria for enterprise V-cache dies like 1 in 10 million bit errors (just a stupid guess) during validation testing. Anything that fails that might still be good enough for consumer workloads. No one cares if some user gets a weird spreadsheet error or a random crash in some game once a year but the same thing happening to a critical enterprise workload could mean a loss in the millions. One reason enterprise chips are expensive is coz they are probably top 1% of the highest bins.

That is the reason Sapphire Rapids keeps getting delay. It's not passing validation. Golden Cove/Raptor Cove are ready for mainstream. Yet for Xeons not quite.

 

[biostud]

You can get the 7373X Epyc which is 16 core with 768mb vcache, so they do make chiplets with less than 8 active cores with vcache, but granted it is another price segment.

 

[Vattila]

I think that makes assumptions about the manufacturing process that may or may not be the case. It's entirely possible that AMD has to make decisions on manufacturing before knowing the status of the cache.

Good point. I am no expert on this. I have read some articles about wafer testing (probing) and known-good-die (KGD) schemes, in which it is stated that these are crucial for chip stacking to work out at production scale, and a key benefit of chiplet design. However, these schemes probably do not provide 100% test coverage, so as you say, there may be lack of certainty and hence a statistical factor involved. In particular, wafer probing may still be limited in what it can do, I guess. I presume AMD is using a die-on-wafer bonding approach, so they may have more certain knowledge about the V-Cache chiplet (which already has been diced and tested to KGD standards, I presume) than the CCD wafer (which has not yet been diced, and hence can only be probed). But, I would think AMD has pretty good knowledge of the functional status of the chiplets before stacking them together.

That said, as @nicalandia pointed out, AMD's V-Cache prototype, shown at Computex last year, was indeed a 12-core dual-CCD chip, with fully functional 64 MB V-Cache on each 6-core (salvaged) CCD, for a full 192 MB total L3 cache. However, this was a prototype, which has not, so far, turned into full scale production. Limiting the demo to 12-core may have been down to package power and thermal limits, I guess, since 16-core is already constrained in the AM4 socket.

 

[ZGR]

I just wanna try out some games with a 1GB L3 Milan-X CPU and see how they compare to the 5800X3D. Anyone wanna donate $30,000 for me to buy the parts? Thanks.

 

[Mopetar]

Are AMD even selling them for individual purchase yet? I'd thought they were still all being sold direct to various cloud providers.

 

[nicalandia]

I just wanna try out some games with a 1GB L3 Milan-X CPU and see how they compare to the 5800X3D. Anyone wanna donate $30,000 for me to buy the parts? Thanks.

If the 5900X3D Prototype is any indication. Unlike Fluid Simulation where 3D Cache does make a difference. Games will not take advantage of the extra Cache on other CCDs....

 

[jamescox]

Unless its correct that these chips failed to go into Milan CPU's, then it would make sense to use everything they can to make money and sell chips.

They may not know the quality of the ccd since it is a wafer level stacking tech, so it does make sense that they have some salvage. It is a mature process though, so it is unclear how many would really be unsuitable for Milan. They have a wide range of core counts used in Milan processors, so it seems like they could still use most parts for Milan. Of course, some will fail for other reasons than just core count; power consumption / frequency curve seems most likely.

If the 32MB cache chips are happening, then these are likely chips where something went wrong in hybrid bonding. The cache die are diced, so I would think they would be tested before being placed into a reconstituted wafer for stacking. I doubt that they would make a special set of Milan SKUs for these smaller cache size parts, so selling them in the consumer market makes some sense, especially if there will be a high cost for the Zen4 / DDR5 platform. This seems the most likely reason to release these parts unless they just have a large number of die that failed Milan validation, but still have the full 64MB of cache.

 

[jamescox]

My thoughts on this are somewhat on a different track. The 5800x3d suffers from limited boost frequency because of a combination of package thermals and total power draw. The 5600x3d, with only half the stacked cache powered and two less cores to draw power may actually outboost the 5800x3d by a nontrivial margin in the same power envelope if AMD relaxes the limits that are on the 5800x3d. It might have better performance on a few games over the 58x3d. Also, from a total cache per core perspective, it doesn't suffer as much compared to the 58x3d as it has about 10.08 MB per core compared to 12MB per.

The 5900x3d is a different animal. It's going to have a massive 16MB of L3 per core and should only loose 100-200MHZ all core boost to the 5800x3d in the same power/thermal envelope. It should have a significant MT throughput advantage over the 58x3d and might surprise against the regular 5900x. If boosting behaves properly, it should be every bit as good as the 58x3d in almost all games with sane code and modest threading.

If the clock speed limitation is at least partially due to voltage limitations on the cache die, then that is unlikely to change for these parts. A 5600X3D should have more thermal headroom, but that may not be the only limitation. The parts that end up in the desktop market are also likely to be leakier, higher clocking parts, so the essentially low power, low clocked cache die are a bit mismatched. The v-cache parts were almost certainly designed almost entirely for the Epyc market where clock speeds at a lot lower.

 

[jamescox]

I just wanna try out some games with a 1GB L3 Milan-X CPU and see how they compare to the 5800X3D. Anyone wanna donate $30,000 for me to buy the parts? Thanks.

I surprised LTT hasn’t tried it yet.