Question Zen 6 Speculation Thread

[Doug S]

not bandwidth ffs, capacity.
Omegafat sockets need a ton of DRAM (as in actual DRAM bits) to be useful!

Same objection. Look at capacity per core over time and tell me why today's capacity per core is a requirement for such a large percentage of workloads that there is no value in having double the cores at half the available DRAM each.

Heck try and make me believe even 1% of whatever the most cores per socket you can currently buy are configured at the maximum supported RAM. Most probably don't even have 1/4 of the max. Leading edge density is a huge price jump, you have to absolutely REQUIRE it to go there.

 

[adroc_thurston]

Look at capacity per core over time and tell me why today's capacity per core is a requirement for such a large percentage of workloads that there is no value in having double the cores at half the available DRAM each.

Because favelas sell fixed vCPU@DRAM ratios and they are the customers for fat sockets.

Heck try and make me believe even 1% of whatever the most cores per socket you can currently buy are configured at the maximum supported RAM. Most probably don't even have 1/4 of the max.

"Maximum supported RAM" == bazinga expensive 2DPC 3DS LRDIMM misadventures aka you do not understand the server market at all.
please stop.

 

[igor_kavinski]

please stop.

I love it when you plead like that. You can do more to not allow people to make you plead, i.e. spill more!

 

[Doug S]

Because favelas sell fixed vCPU@DRAM ratios and they are the customers for fat sockets.

And your contention is that they always max out DRAM regardless of cost. So post their vCPU/DRAM ratios they are selling at, and compare with the densest AMD/Intel sockets and their maximum RAM config to prove your point.

Though I'm guessing here is where you say "I'm not going to do your research for you" and ask ME to prove YOUR ridiculous claim.

 

[adroc_thurston]

And your contention is that they always max out DRAM regardless of cost

No, they max out what they can without resorting to 3DS LRDIMMs.

So post their vCPU/DRAM ratios they are selling at, and compare with the densest AMD/Intel servers and their maximum RAM config to prove your point.

i mean opening the aws instance SKU list ain't hard.

Though I'm guessing here is where you say "I'm not going to do your research for you" and ask ME to prove YOUR ridiculous claim.

The claim is simple enough to be uttered by anyone working at any major CSP.
Stick to phones.

 

[igor_kavinski]

And your contention is that they always max out DRAM regardless of cost.

Well, if they are like me, they would do that. I would feel very flattered if they follow my thinking

 

[igor_kavinski]

No, they max out what they can without resorting to 3DS LRDIMMs.

I've seen portable servers (ones you can carry around) being advertised with 6TB RAM. How come they are getting 3DS working?

 

[adroc_thurston]

I've seen portable servers (ones you can carry around) being advertised with 6TB RAM

I mean if it has a normal socket it can has as much DRAM as it fits.

How come they are getting 3DS working?

me after my wife splurged on 6TB worth of LRDIMMs

 

[eek2121]

Thanks,

This looks interesting. Actually I am paying that 30€ for Copilot & 20€ for Chatgpt+. I was hoping to find a shortcut to finally arrange family photos on my laptop🤣

You could save a lot of money by downloading LMStudio. Just sayin.

 

[DrMrLordX]

All you've said is just hypothetical nerd wank;

That's normal for a speculation thread!

 

[blackangus]

That's normal for a speculation thread!

Well there is speculation - Something based on known information.
Then there is fantasy - Some that someone wants because its cool or dreamy.

IMHO speculation should have some basis in known and current info. - This is of value to a thread.

Anyone can spout fantasy with no basis in fact - This detracts from a thread.

For instance - AMD is going to team up with nvidia and intel and fab nvidia designed GPUs with AMD chiplets for a "perfect" SOC on INTEL A14 for Zen 6.

 

[Josh128]

View attachment 131354

Don't really think it's a leak but what do I know?

Its a picture. Looks like a "Gamer Meld" -style YT thumbnail regurgitating 3 month old MLID info and passing it off as something new and accurate. AMIRIGHT?

 

[igor_kavinski]

AMIRIGHT?

Very possibly yes!

 

[Kaffeekenan]

Very possibly yes!

And the "2,89 times faster" is about the RAM-Speed. God, how I hate these stupid videos with those stupid clickbatetitles. Pure garbagecontent...

 

[igor_kavinski]

And the "2,89 times faster" is about the RAM-Speed.

I could kinda believe such a number with multicore benchmark scores but with RAM speed??? That's like saying Zen 6 will go up to DDR5-17500 or get quad channel DDR5-8800 support.

 

[adroc_thurston]

That's normal for a speculation thread!

Naa, speculation implies understanding the limits of what's possible instead of going "512c sockets are viable rn because IPHONE PRO MAX you see".

 

[Kaffeekenan]

I could kinda believe such a number with multicore benchmark scores but with RAM speed??? That's like saying Zen 6 will go up to DDR5-17500 or get quad channel DDR5-8800 support.

My bad. It's memory writespeed in the video. So maybe my wording "RAM speed" was not 100% correct. But still the video sucks. xD Sorry for the confusion.

 

[igor_kavinski]

Naa, speculation implies understanding the limits of what's possible instead of going "512c sockets are viable rn because IPHONE PRO MAX you see".

Let people dream and soar to heights unknown!

 

[adroc_thurston]

Let people dream and soar to heights unknown!

they're known factors for known platforms.

 

[OneEng2]

512C totally doubt this with the meager density gains they have to make the package significantly larger 384C seems possible

I think it will come down to competition. If there is no reasonable competition to Zen 6 in DC, then I agree, they will not increase core count (maybe frequency, maybe a little more transistors per core).

Well there is speculation - Something based on known information.

As usual, your attempt to define the world according to your own definition is .... wrong.

Dictionary Definition

Speculation: "The forming of a theory or conjecture without firm evidence."

Not only are you wrong, you are 180 degrees wrong. Speculation, by its very definition, is a theory that has no proof or "known information".

Now, your "speculation" on why we won't see core counts go out the ceiling (RAM per core, or bandwidth per core) may have merit in some situations, but not all.

A productive speculation would be to outline what instances more cores without more RAM or more bandwidth is useful.

Also, as others have done, outlining the cost to AMD to increase core count on A14/A16 is an interesting topic for speculation .... and which node Zen 7 might be on in the future.

I don't consider 512 cores per socket to be out of the realm of possibility for AMD.... although as things sit today, it may well be outside the realm of financial good sense.

 

[igor_kavinski]

As always, competition to provide more value per dollar will force the hand of one or both players and core counts will HAVE to increase. Science or reason or logic be damned!

 

[511]

think it will come down to competition. If there is no reasonable competition to Zen 6 in DC, then I agree, they will not increase core count (maybe frequency, maybe a little more transistors per core).

It's not competition 512C just doesn't make sense with the density gains you have to significantly increase both your package size and need more Bandwidth to feed the beast.

 

[igor_kavinski]

need more Bandwidth to feed the beast.

Do they care? They can easily provide a starved beast to companies and they will buy it. Their consumer line-up is massively starved on bandwidth. We deserve something with the bandwidth of Strix Halo and the latency of DDR5-6000 CL28.

 

[adroc_thurston]

It's not competition 512C just doesn't make sense with the density gains you have to significantly increase both your package size and need more Bandwidth to feed the beast.

Package size is whatever, bandwidth you can manage.
How do I get enough DRAM per core for a 512c socket with 16ch 1DPC.

They can easily provide a starved beast to companies and they will buy it

this is server.

Their consumer line-up is massively starved on bandwidth

Oh no it ain't.

We deserve something with the bandwidth of Strix Halo and the latency of DDR5-6000 CL28.

You do understand that LPDDR is inherently higher latency.
It's either-or, and AMD's choice is that you get *neither*.

 

[511]

Package size is whatever, bandwidth you can manage.
How do I get enough DRAM per core for a 512c socket with 16ch 1DPC.

You can do 24ch 1DPC.