Return to Level1Techs.com

Ryzen 3000, GPU speculation

amd

#1

Almost too good to be true. 5.0Ghz chip from AMD. Who actually will make a 28 core 5Ghz CPU now ?


#2

IF and I stress this… IF that 3850X is in fact a 16 core/32 thread @ 5GHZ for 500 dollars is in fact true that could very well be this generations 2600k. A chip so far ahead of its time you could run it for 8 to 10 years and still be able to play modern games.


#3

Id agree my bull shit meter is up there. But 7nm is new ground.


#4

My educated guess is none of this is real.

Also adored


#5

Looking at Adored`s history of accurate predictions, I would take 10% salt with the information given.

I doubt the 6c chiplet, I think it is an 8c chiplet with 2 dead cores.
I/O die makes sense (look at Epyc Rome, 8x 8c chiplets + I/O die)

The numbers presented make sense as 7nm is still a massive black box (apart from the PR material from TSMC). So AMD could in fact have 5GHz chips that make sense in early production.
I don´t like the Ryzen 9 stuff. 16 cores on threadripper are starved for memory in some workloads allready. On AM4, unless AMD pulled some magic in 2017, it will be worse.


I am curious to see how Navi will look. The I/O die would make sense as it would make the GPUs much cheaper to manufacture.
Another benefit would that it could talk the same language as the CPU I/O die making integration into APUs easier (=cheaper).


Seeing all of it and without semiAccurate, canardPC, videocardz or WCCFtech pushing the rumor mill, I have my doubts.


#6

The numbers are close based on what 7nm has purported to be, bit high side… possibly. As far as rumor monger sites don’t really fuck around with videocardz since the whole bullshit from earlier this year where they supposedly saw invoices for tech youtubers along with another jackass who named people directly only to have that guy take the fall. And WCCFTech… yeah, I just don’t buy them, at all.

Certainly some good ideas… and good news is if the rumors are true we only have to wait a month.


#7

It is about having multiple sources. I don´t care what sort of BS a site publishes, I care about another source for rumors. It is a paradox in itself, but so far, it worked (kind of).

Patience is key.


#8

No idea if this is it but been imagining that HBM collar being used similar to RTX cards, but then with just GPU chiplets instead combined with some sort of neural net

:thinking:


#9

Honestly why wouldn’t it be? I mean we already have 32c/64t 4+Ghz threadripper. If the die shrink is used purely for speed bump and binned well enough - why wouldn’t there be 5Ghz parts?
Anyways I doubt we will see 32cores on desktop AM4 right now… Not with those motherboards and TDPs.


#10

The data looks like wild guesses completely based on the data seen at the rome launch.

From what I’ve seen with the zen arch, and TSCM 7nm, a 5GHz part would surprise me, and so would anything beyond 8 cores on AM4.
The chiplets AMD are getting from TSMC now are 8 core parts, and they are expensive AF. On top of that TSMC has limited capacity on the 7NM node, and Apple is their #1 customer, nvidia their second, so if they were to go beyond 8 cores on AM4, they would need two chiplets on the package, and that doesn’t make sense from an economic point of view.


#11

The general idea of a 5GHz part isn’t really anything that surprising to me, it’s only the rest of the context would make it unlikely.

Intel was promising 10GHz processors as soon as 2005 (based on transistor technologies they announced at the end of 2000 and used in the Pentium 4), but that was before they stopped the clock race, which was expensive for power and heat (and thus again power for cooling) and switched to things like cores and other ways of advancing speeds. I seem to recall more recently in Scott Mueller’s Upgrading and Repairing PCs 21st edition, he mentioned an already long known manufacturing process that CPU makers could (and presumably would) switch to which would allow pushing clock speeds to around 15Ghz over it’s usage, before switching to another already mostly understood process that could go higher and do other things. The only reason it seemed this process wasn’t already used seemed to be because it wasn’t yet necessary and they could just keep pushing things through other means first, but he expected around 2015 the switch would start.

We’re past that point now, but we also know Intel slowed down (for many reasons, including lack of competition from AMD) and probably found some unexpected advantages in the meantime allowing them delay that transition. Soldered CPUs which are allowing Intel processors to push clocks higher with some extreme cooling allowing overclockers to hit 6.9 GHz, but I can’t remember at the moment if that was the tech that would allow things to push much higher.

Getting 5Ghz in a package with 12 Cores though, without a change elsewhere, that’s where it sounds difficult, particularly on existing AM4, unless they were already planning for this (which may be the case).

What’s more exciting is, if this all pans out, that may kick the development race back into gear. The highest end may become more expensive for a while, but we may see faster advancement due to a heated competition between Intel and AMD.


#12

It seems more likely the boost frequency would be referring a one or two core turbo, similar to how current XFR/PBO works now. With that it isn’t hard to believe they could fit x cores in a package and have them scale individually to fit a power target.


#13

My bet is on 3.8 ghz base, 4.5-4.6 ghz boost. Could they do 5 Ghz? Maybe. But the power efficiency and ease of cooling will be much better below 5 Ghz. Ryzen already has a lot of cores to feed, i don’t think AMD will push clocks quite as hard as intel because of this. Intel’s 5 Ghz i9 is an act of desperation more than anything else IMHO.

Not sure they’d integrate NAVI into Ryzen 3000 series, but i could be surprised.

My bet would be on Ryzen 3000 including Vega variants and Ryzen 4000 including onboard Navi in order to reduce risk incurred with changing both the CPU and GPU cores in the same generation (and also because Vega mobile is still super new and software support is still coming).

I could be wrong, but that’s how i’d play it. The integrated GPU market is not super performance sensitive, Vega already kicks arse there and it will be lower risk.

edit:
7nm may not provide the clock speed boosts some are expecting. I believe around 14nm we’ve reached the point where the concentration of heat in the die becomes a more significant issue (this is why sandy bridge or ivy bridge isn’t so far behind say, haswell or even skylake in terms of max clock) as opposed to just the total heat generation. My bet is AMD will focus on more cores in the more efficient clock speed region, and my bet is that will be somewhat below 5 ghz. If the rumours on their significant IPC improvements in Ryzen 2 (3000) are true, then they won’t need 5 Ghz to compete. Clock for clock, Ryzen 2000 is already pretty close to SkyKabCoffeeWhateverLake.


#14

We probably never will. Will be 16c at most. Ryzen 9 cpus are probably gonna have to be 5XX series or higher due to the higher draw


#15

Depends. DDR5 and/or a bigger local cache in (or connected directly to, on package) the IO die may be enough for 16 or more cores on AM4 to be actually usable for the workloads that AM4 is used for.

I’m just waiting to see AMD stick a couple of GB of HBM on the AM4 (or more likely, TR4) package as an L4 cache… i very much suspect that will happen eventually - HBM as yet another level of cache, because motherboard bandwidth isn’t scaling as fast as core counts are.


#16

I watched the fucking video :). There was mention of 12 core capping out the old AM4 Motherboards because power. Now that is all subject to speculation as well.


#17

Should be in the intro of every Adored video.


#18

Well if he was wrong a lot there would be meme’s upon meme’s.

Im not in for the CPU’s a 1700 is ok but a cheap vega 56 level navi GPU would be nice.


#19

He’s usually pretty clear it is mostly speculation, but his sources as of late have been pretty reliable, his reasoning is sound, etc. AdoredTV is speculation, but it is educated and well researched guessing and there’s a lot of actual research behind what he does.

I’ll give it a watch later tonight, but credit where credit is due: he called the chiplet based arch with an IO controller for Epyc many months ago.


#20

Actually, 10 Ghz was on the roadmap back in 2001 (maybe even earlier - but yeah you’re probably right about the expected delivery date being 2005 or so). I remember because I was working in an ISP/IT shop, the P4 was new, clocks were ramping hard (the P4 was jumping from intel barely being able to crack 1 ghz on paper launch of the 1 ghz p3 to 1.6 and beyond) and i remember the employer i was at when i read about it and talked with other staff about it.

I quit that job in October 2001.

If the P4 (and derivatives) design could have actually hit 10 Ghz as intended it wouldn’t have been such a joke. Unfortunately for intel, it was designed for that sort of clock, but unforeseen silicon limitations getting there killed that idea.