AMD demonstrates Ryzen 9 5900X prototype with 3D V-Cache stack chiplet design

[Journey]

Probably wishful thinking though and just on 7nm

They will be the same chiplets used for Milan-X, and the same I/O die they use now, so obviously it will be 7nm for the chiplets. I'd say the new 5600X-3Dcahce will be the best PPW, and maybe even the best VFM if priced well enough against the 12600K.

 

[g67575]

Old news, but each CCD will apparently have 96MB of cache according to Tomshardware! Link here:
https://www.tomshardware.com/uk/new...mb-of-l3-cache-per-chip-15-gaming-improvement

That's 3 times the L3 cache of the 5800X. For an 8 core part, could be looking at a maximum of 44 watts for the interconnect and L3 cache (based on Humbug's earlier 14.8 watt figure for the power consumption of a 5800X's Cache +interconnect), for a 5800X equivalent with V-Cache.

It's likely to be less than this though, do to an interconnect with '3x the energy efficiency'.

If there's a 16 core part, the article mentions a total of 192MB L3 cache. Impressive if true.

 

[Journey]

"The technology also doesn't require any specific software optimizations and should be transparent in terms of latency and thermals (no significant overhead in either)."

Might also want to read the link HERE! about the AMD I/O uncore to learn a thing or two about the way the chips assign power.

 

[pcuser]

If memory does not mislead me AMD are usually boosting the hype when they have something in their sleeve but seem way too quiet at the moment at least to me. Hope they will manage to get the 3D cache processors for next year as Intel seems to be on the rise especially with the recent Raptor Lake leaks.

 

[LePhuronn]

If memory does not mislead me AMD are usually boosting the hype when they have something in their sleeve but seem way too quiet at the moment at least to me.

There's nothing to hype right now. We saw the 15% average gaming uplift in June, we know production starts "end of 2021" and AMD don't discuss unannounced or unreleased products in earnings calls. If AMD don't talk about any v cache products at CES then maybe there's an issue.

And what is there to leak if the actual chips are only just being made?

 

[Joxeon]

There's nothing to hype right now. We saw the 15% average gaming uplift in June, we know production starts "end of 2021" and AMD don't discuss unannounced or unreleased products in earnings calls. If AMD don't talk about any v cache products at CES then maybe there's an issue.

And what is there to leak if the actual chips are only just being made?

You'd expect engineering samples to be out there given that we were seeing Alderlake samples popping up over 6 months ahead of launch.

 

[LePhuronn]

No, you wouldn't, because AMD isn't leakier than a sieve like Intel. In 5 years we've never had more than a handful of ES chips show up.

 

[Grim5]

Old news, but each CCD will apparently have 96MB of cache according to Tomshardware! Link here:
https://www.tomshardware.com/uk/new...mb-of-l3-cache-per-chip-15-gaming-improvement

That's 3 times the L3 cache of the 5800X. For an 8 core part, could be looking at a maximum of 44 watts for the interconnect and L3 cache (based on Humbug's earlier 14.8 watt figure for the power consumption of a 5800X's Cache +interconnect), for a 5800X equivalent with V-Cache.

It's likely to be less than this though, do to an interconnect with '3x the energy efficiency'.

If there's a 16 core part, the article mentions a total of 192MB L3 cache. Impressive if true.

192mb looks small next to EPYC, the power of stacked memory - can't wait till we can get rid of RAM and chuck everything in L3 cache (I know it won't happen but one can dream)

 

[Perfect_Chaos]

Ian Cutress was on Tech Deals earlier, he didn't seem very excited about Zen 3D saying they will cost about 50% more as well.

 

[humbug]

Ian Cutress was on Tech Deals earlier, he didn't seem very excited about Zen 3D saying they will cost about 50% more as well.

50%? So £1250 for the 6950X?

It had better be 50% faster...

 

[Grim5]

That doesn't make any sense, CPUs are not in low supply, AMD won't sell many units if they did a 50% price increase for 5-10% extra performance and the store shelves are filled few CPUs they can't move.

the only way they could get away with a 50% increase is if Zen 3D is a "special edition", like a one off batch run (like 10,000 units per model or something and you called it 5950XT etc).

Again, CPUs are not in low supply. You can get away charging those prices on a new GPU but not a new CPU

 

[Twinz]

I have not seen any info that would indicate performance from Vcache would allow them to charge another 50%.

Maybe they just want to brag about how their CPU's remain "in stock" and "available."

 

[Perfect_Chaos]

Could have sworn he said 50% or up to 50% more expensive somewhere in the video, can't find it now but he talks about the performance at 1hr 35 mins.

https://youtu.be/A9hUHuQcyiQ?t=5700

 

[Joxeon]

Maybe they think they can increase prices by 50% due to these V cache CPUs mining performance.

The 6600XT that was released during the ETH boom was almost a 40% MSRP increase from the 5600XT despite lacklustre performance increases and even the 6600 non XT went up by 20% despite only being 5% faster.

 

[g67575]

Maybe the CPU/CPUs with 192MB (96MB per CCX) of L3 cache will be expensive, with the others being more affordable?

I'd guess there will be cheaper models with 64MB per CCX too.

A significant increase in price wouldn't be too surprising though, isn't CPU cache the fastest (lowest latency) and most expensive type of memory available?

You have to wonder how much a similar amount of L1 cache would cost?

 

[Purgatory]

Could have sworn he said 50% or up to 50% more expensive somewhere in the video, can't find it now but he talks about the performance at 1hr 35 mins.

https://youtu.be/A9hUHuQcyiQ?t=5700

It's here he talks about 50% price increase :- https://youtu.be/A9hUHuQcyiQ?t=6228

 

[KompuKare]

Is it possible that Ian meant the cost to make for AMD?

Manufacturing and production costs are after all something Ian knows about.

Back when I did the wafer costs calculations, the wafer cost of a CCX was about $12 or so. Half the size of SRAM should be cheaper (better yields and less complex) but assembly costs will be higher for such a new thing.

$12 CCX die
$5 SRAM cache die
$?? IOD die
$?? normal assembly costs
$?? new process assembly costs

However, even a deliberately cut down Ryzen 5 5600X (yields should be good enough that almost all dies would be good for 8C/16T) has far higher margins than lots of other things AMD do like GPU, console chips etc., and once the lower end ADL chips and chipsets come out I can see Zen3's marketshare declining unless they cut prices anyhow.

 

[Dave2150]

You'd expect engineering samples to be out there given that we were seeing Alderlake samples popping up over 6 months ahead of launch.

This. I mean we have Raptor Lake engineer samples submitting benchmarks already!

 

[Dave2150]

Ian Cutress was on Tech Deals earlier, he didn't seem very excited about Zen 3D saying they will cost about 50% more as well.

Do you have a source for this? 50% more expensive sounds madness. I know TSMC increased their prices in the last few months, but......

 

[subbytna]

One more thing, remember the FX-9590? that was also on a PGA Socket, AM3+, it had a package power of 220 Watts.

AMD are willing to throw the PGA 142 Watt limit out of the window, it literally fried cheaper not well made boards, mainly those from MSI but i ran one for a year OVERCLOCKED to 5Ghz all core on my Asus Sabertooth 990FX Rev2 and it never bothered it in the slightest.

There is no reason why AMD can't release an X590 board with higher power directives for vendors to adhere to, that would unleash the high core count Ryzen 5000 which are capable of running much higher clocks all core than they currently do with the 130 watt limits in place.

They'll not use the x590 option as it's too close to Z590 intel naming, but I see your point