Question Zen 6 Speculation Thread

[adroc_thurston]

A CPU that used a 2 year old process at the time of its launch

That was still bleeding edge.

 

[gdansk]

Zen 2 shipped on 7nm about 10 months behind Apple.
It'd be about the same for 2nm and Zen 6. But this time they'd be the first on N2P, which would be different but that doesn't seem unlikely (to me) given AMD's massive revenue growth since 2019.

 

[Io Magnesso]

Idk how we're still debating the nodes

Venice (MS server) uses N2P
Venice-Dense (Cloud server) uses N2P
Olympic Ridge (desktop) uses N2P
Gator Range (high-end laptop) uses N2P
Medusa Point 1 (premium laptop) uses both N2P+N3P for top SKUs and N3P only for lower end SKUs

All of this information has been given from AMD to OEMs and AIBs.

Where did this leak come from?
Surely, some OEM leaked out
Personally, I've never heard of it

 

[Win2012R2]

More than 1/2 of the video is Tom's tariff retardation

The other half is "and now an ad from a sponsor..."

 

[511]

The other half is "and now an ad from a sponsor..."

His secret sponsor is AMD( cause his AMD leaks are pretty accurate but his Intel leaks are hit or miss)

 

[Josh128]

That was still bleeding edge.

Bleeding edge means being the first to use the node, like Apple or Qualcomm, not using it when the node successor is about to commence volume production. This is how AMD has operated thus far, and them actually skipping a major node has never happened. I hope Im wrong, but done see AMD putting desktop on 2nm when a likely underutilized, mature, and cost-effective 3nm is ripe for the picking.

 

[soresu]

Bleeding edge means being the first to use the node

To be specific it's more like using it when the yields are still somewhat in the toilet, but the profits gained from the end product are so good (a la iPhones) that it doesn't really matter.

 

[soresu]

in AMD's case server/datacenter CPUs are a very good revenue stream increasing YoY, so they can afford to be a bit more aggressive on the node capacity front.

 

[Io Magnesso]

That said, which one is more profitable, the server or the desktop? I think it's a server

 

[Josh128]

Where did this leak come from?
Surely, some OEM leaked out
Personally, I've never heard of it


Venice (MS server) uses N2P
Venice-Dense (Cloud server) uses N2P
Olympic Ridge (desktop) uses N2P
Gator Range (high-end laptop) uses N2P
Medusa Point 1 (premium laptop) uses both N2P+N3P for top SKUs and N3P only for lower end SKUs

If this indeed turns out to be true, it'd represent the most aggressive move ever by AMD, a serious attempt at a "kill shot" against Intel.

If that is the case, Zen 6 is shaping up to be a much more impressive jump performance wise than Zen 5. 2nm + 12core CCD + new GPUIOD + potential clock boost should be far > new uarch on same node at same frequency that Zen 5 was. Now, Im not expecting miracles in latency and mem speed improvements from the new IO, but I do expect at least SOME improvement. Intels fancy looking "tiled" IO seemed to suffer from the same issues as Ryzens, leading me to believe that the trade offs are just inherent to chiplet designs and short of a true breakthrough, improvements will be incremental.

 

[soresu]

That said, which one is more profitable, the server or the desktop? I think it's a server

Not even a question - especially at the 2S SKU range.

 

[Io Magnesso]

To be specific it's more like using it when the yields are still somewhat in the toilet, but the profits gained from the end product are so good (a la iPhones) that it doesn't really matter.

TSMC doesn't usually fail... but There is also an example called N3B. After all, it may be just right to use it after waiting for a while when it comes to a new process. As time goes on, production increases

 

[Io Magnesso]

If this indeed turns out to be true, it'd represent the most aggressive move ever by AMD, a serious attempt at a "kill shot" against Intel.

If that is the case, Zen 6 is shaping up to be a much more impressive jump performance wise than Zen 5. 2nm + 12core CCD + new GPUIOD + potential clock boost should be far > new uarch on same node at same frequency that Zen 5 was. Now, Im not expecting miracles in latency and mem speed improvements from the new IO, but I do expect at least SOME improvement. Intels fancy looking "tiled" IO seemed to suffer from the same issues as Ryzens, leading me to believe that the trade offs are just inherent to chiplet designs and short of a true breakthrough, improvements will be incremental.

As for the chiplet, when AMD first introduced the chiplet, the degree of perfection was lower than it is now. It's just an accumulation of experience and improvements I think it's the same for Intel and AMD.

 

[jpiniero]

If this indeed turns out to be true, it'd represent the most aggressive move ever by AMD, a serious attempt at a "kill shot" against Intel.

If you think about it, Client CPU performance doesn't matter anymore. Only AI (and VCache for DIY).

The pricing will be crazy I am guessing.

 

[Io Magnesso]

If you think about it, Client CPU performance doesn't matter anymore. Only AI (and VCache for DIY).

The pricing will be crazy I am guessing.

If your server/AI earnings increase...
V-cache for consumers will also be lost in sight...

 

[OneEng2]

One possibility is that they sell the 48 thread part for $1000 while the 32 thread part maintains its current price more or less.

And how many people would buy a 1k cpu?

That said, which one is more profitable, the server or the desktop? I think it's a server

Most certainly!

I think AMD is all in on Server first strategy. But you can't ignore their increased margins in high end desktop even though the volumes are much lower than corporate laptop.

 

[Io Magnesso]

And how many people would buy a 1k cpu?

Most certainly!

I think AMD is all in on Server first strategy. But you can't ignore their increased margins in high end desktop even though the volumes are much lower than corporate laptop.

If the server/AI is more on track, Consumer markets may not be so important...
AMD also plans a rack system. You can sell more EPYC

 

[Win2012R2]

And how many people would buy a 1k cpu?

Almost everybody who will buy 6090, so that's a few millions at least

And given that Unreal Engine 5 will be everywhere soon a fast CPU will be essential to enjoy gaming, does not need to be 1k version however

 

[igor_kavinski]

And how many people would buy a 1k cpu?

A lot of people buy the highest end because of the prestige of owning a flagship CPU and also due to FOMO because of scalpers.

I can almost bet that at least 50% of current 9950X3D users do not need and probably don't use all the threads but it just makes their heart feel really warm inside to own the CPU.

 

[Joe NYC]

No, at least in the area where I live
9800x3D is traded at a fairly high price

Amazon US:
9800x3d: $472
7900x3d: $369
14900k: $438
285K:$559

Intel is already charging more for weaker gaming performance. Thinking that Intel will charge less for a CPU with 2 large N2 dies is delusional.

And it is not like Intel will not NEED to charge even higher prices. Intel is still in the middle of a financial downfall, the turnaround is still in the distant future (if ever).

This NVL with big cache will probably well > $700

 

[Joe NYC]

Ryzen7's X3D model is... The concept was to boast the same performance as a high-end CPU at a middle range price… 5800X3D…
It's a good dollar box for AMD, so I can't blame it... I think the current 9800x3D is out of that concept.
In the first place, the packaging lanes for implementing V-cache are limited, so there are some places that can't be helped...

The capacity is limited only by what you place orders for. There was, maybe a 1+ year long capacity shortfall for CoWoS type packaging, but that is probably easing too. So there is nothing in the way of AMD placing orders for greater capacity, and TSMC completing that capacity.

If there is a 9600x3d that is universally available (not limited to only certain retailers like 5600x3d was) and AMD prices it between $250 and $300, then AMD will double the shipments of x3d parts.

Which indicates that AMD probably already ordered that capacity increase...

 

[Joe NYC]

This is just a story about AMD using N2...
AMD is not the first to order N2
Some manufacturers may have ordered it at the same time as AMD, or have already ordered it before AMD ordered it.

This is just an announcement that it was tape-out and produced (probably engineered sampling).

I think this time around, AMD will be definitely in top 3 users of the most advanced node (N2). Products that will use N2:
- all Venice generations server CPUs
- Mi400 generation of datacenter GPUs
- premium client CPUs

Those numbers will start to add up, since AMD will be close to 50% server CPU volume, and most likely, AMD datecenter GPU will be on a strong upswing with Mi400 generation (from low single digit market share).

And there will likely be a big shift in the premium client CPUs in AMD direction...

 

[Joe NYC]

Chances of us encountering alien life in the next 5 years are higher than that.

There is some way for Intel to even come close to catching up to AMD L3 performance.

You have to wonder if Nova Lake is to be another "leapfrog" concocted by Gelsinger that will fizzle out just like like all his other "leapfrogs" have fizzled out so far.

 

[Joe NYC]

Almost everybody who will buy 6090, so that's a few millions at least

And given that Unreal Engine 5 will be everywhere soon a fast CPU will be essential to enjoy gaming, does not need to be 1k version however

A lot of games are already on first implementations of Unreal 5 engine, but it is the newer ones and upcoming versions that will, apparently, reshuffle the CPU-GPU interactions and increase parallelism.

Are there any more detailed predictions on the effects of this?

 

[Win2012R2]

A lot of games are already on first implementations of Unreal 5 engine, but it is the newer ones and upcoming versions that will, apparently, reshuffle the CPU-GPU interactions and increase parallelism.

They've only improved it just about now in 5.6 - most big game dev use much older versions as it takes 5-6-7 years to make game these days, and upgrading isn't trivial so they will ship using older versions most certainly. And frankly 5.6 isn't exactly solving the problem completely, they only hope to achieve it in UE 6 - so that's for games a decade from now.