Speculation: Ryzen 3000 series

[Yotsugi]

they have a great chance to do a single 8 core chiplet with high clocks and market it for gaming

But instead they will clock the 6/8c parts conservatively and go gangbusters with 12/16c ones for them sweet sweet ASPs.

 

[maddie]

7nm does neither of those. The numbers might make you think that but the are deceiving. Look at the TSMC 7FF vs 16FF+ numbers. They are the most comparable to GloFo 14/12nm. Apparantly, AMD themselves has said they can get We know that AMD has said it can deliver 25 percent more performance on 7nm within the same power envelope. That's good, but certainly not 1/2. All we can do is wait and see, though.

Power is approx a cubic function, so 25% more perf aka clocks is the same as 1/2 power at same clocks. 1/1.25^3 = 0.5

 

[maddie]

Fair enough, but the point still stands. Heat won't be distributed evenly over the new, smaller surface area. It can lead to hotspots.

True, but hotspots always exist, and what I'm saying is that for a given group of transistors doing any specific computation, we will have 1/2 the power generated in 1/2 the silicon area. I see the heat flux as similar and have not seen any analysis to dispute this. All of this assumes similar clocks.

If you start to do more intensive work in a limited area then this can lead to an increase in local heat generation. AVX512 is a good example of this.

 

[maddie]

One chiplet has less energy to dissipate. Two chiplets increases the energy needed to be dissipated. When you have 12 active cores, you have less energy to dissipate vs 16. So, going up 4 more cores increases the amount of energy that would need to be dissipated without an increase in the surface area of the heat spreader. Further, the whole package now has increased power usage as there are more cores doing work over things like the IF.

In my opinion, they have a great chance to do a single 8 core chiplet with high clocks and market it for gaming. They would sell a lot more of those chips than a cpu with two chiplets in a single package that costs much more.

Agree with the high clocked 8C part. That was one I predict to be the highest clocking part, but (2) 4C chiplets binned for the highest clocked cores and with the greatest area of silicon to dissipate heat.

A 50th anniversary special edition?

 

[Yotsugi]

That was one I predict to be the highest clocking part

Bad take.

A 50th anniversary special edition?

That would be a 16c product.

 

[maddie]

Bad take.

That would be a 16c product.

If you can bin chiplets for power and frequency, can't you do the same within a chiplet by disabling the weaker cores? This allows a better performing 4C chiplet even within the premium 8C chiplet bin as you will still have some variation within any full chiplet.

Guess we'll have to wait and see.

 

[Yotsugi]

This allows a better performing 4C chiplet

That will go into ludicrously expensive per-C licensing Rome SKUs, not into cheapo desktop parts.

Guess we'll have to wait and see

That's the spirit.
You don't even have to wait that long anymore.

 

[Veradun]

570 boards will be mostly an AMD in house design. Something to do with the previous vendor & PCI 4.0
I’m just hoping for better memory choices

I'm currently using 32GB (2x16) 3466 JEDEC on a Ryzen 1600 (and x470), plug'n'play.

If you use standards you shouldn't have any problem and still have lots of available choices :>

 

[Thala]

Power is approx a cubic function, so 25% more perf aka clocks is the same as 1/2 power at same clocks. 1/1.25^3 = 0.5

Indeed, or corollary, if you use 50% power at iso clock, you can get +25% performance at iso power (0.5 (@iso clock) * 1.25^3 = 1 (iso power)).

 

[DrMrLordX]

True, but hotspots always exist

They get worse on smaller nodes. That's why there were posts on this forum years ago about the same subject (and the challenges of making nodes like 7nm, 5nm and 3nm functional). Consider this:

https://semiengineering.com/power-challenges-at-10nm-and-below/

I guess what I was really reaching for here was current density.

 

[maddie]

They get worse on smaller nodes. That's why there were posts on this forum years ago about the same subject (and the challenges of making nodes like 7nm, 5nm and 3nm functional). Consider this:

https://semiengineering.com/power-challenges-at-10nm-and-below/

I guess what I was really reaching for here was current density.

That article has some weird writing. For example.

"First, dynamic current density increases because transistors are physically closer to each other in every direction, trapping heat between the fins. And second, below 16/14nm, current leakage begins increasing again, which also shows up as heat.

Trapping heat between the fins is a very, very strange explanation. I could see it being used for non engineering types as a metaphor, but this is an engineering site.

What I was saying is that the (TSMC) power consumption figure by itself already accounts for all of the design factors. You don't need to say why the silicon is heating up, by using current density, leakage or anything else. For a simple adder circuit as an example, moving from 16nm to 7nm has 1/2 the silicon needed for the transistors, but generates 1/2 the heat (power), so the thermal flux is identical.

Once you go for higher clocks however, then you start to see higher heat flux.

 

[darkswordsman17]

570 boards will be mostly an AMD in house design. Something to do with the previous vendor & PCI 4.0
I’m just hoping for better memory choices

Ah, that could be nice. Which didn't they use someone else's memory controller in Zen 1?

I don't think there's anything concrete about the 500 series yet. Some rumours i've seen mention it's no longer ASMEDIA making the pch, adding a lot more additional IO lanes to make them more comparable to intel's more powerful chips and also negative ones where there's delay on the production of the parts causing launch delays including pushing back the cpu launches.

I couldn't remember who did the chipset they'd been using, but I seem to recall there being talk that they were using a pretty outdated process for the chipset (so its not like they'd need to be doing 7nm on it but just using a newer one would help them cut some power use) and that the chipset wasn't especially impressive (but with how much is on CPUs these days its not that big of a deal, like the PCIe lanes being mostly dictated by the CPU). Which wasn't the chipset also the source of those security issues?

Sounds like the 500 series stuff would be worth waiting for (especially for Zen 2), but bummer if its causing delays. Hopefully nothing too bad.

On the other hand, do we realize that Zen3 is probably coming 1 year later. Not a whole lot of time to relax and space out your SKUs.
Got some negative reactions for simply saying that a 6C/12T CPU might not be the best for gaming in a few years as the consoles would be 8C/16T CPUs and establish a new acceptable minimum. Seems some folks just can't accept the big changes happening and more importantly, continuing to happen.
Throw out all of your biases built up over the last several years. We're in uncharted waters here, for the times they are a-changin.

That still surprises me. I was expecting something like the 2000 series. Which that might be the case but they're just calling it Zen 3 (kinda like how Intel is calling every new year the _th Gen, even when there's little difference and now there's a whole variance of cores being used even), but then I feel like they should sync the x000 series with the Zen _ or do something

 

[DrMrLordX]

Trapping heat between the fins is a very, very strange explanation. I could see it being used for non engineering types as a metaphor, but this is an engineering site.

I thought so too, on second examination. Perhaps this is related?

https://semiengineering.com/will-self-heating-stop-finfets/

 

[maddie]

That still surprises me. I was expecting something like the 2000 series. Which that might be the case but they're just calling it Zen 3 (kinda like how Intel is calling every new year the _th Gen, even when there's little difference and now there's a whole variance of cores being used even), but then I feel like they should sync the x000 series with the Zen _ or do something

Well it first went Zen on both 14nm and 14nm+, Zen2 and Zen3. Now we have Zen, Zen+, Zen2 and Zen3 as 14nm+ became 12nm. Zen3 is a different beast to Zen2, no + moniker.

 

[Yotsugi]

Which didn't they use someone else's memory controller in Zen 1?

PCH has nothing to do with the IMC.
AMD uses their own IMC, DDR PHYs are Synopsys.

That still surprises me

Yet you're probably posting that from a phone, a world where vendors live and die on yearly cadence.
Ever wondered how does that work?

Which that might be the case but they're just calling it Zen 3

It's very much a new core.

 

[maddie]

I thought so too, on second examination. Perhaps this is related?

https://semiengineering.com/will-self-heating-stop-finfets/

That looks to be a general situation for all finFETs. Heat from itself added to heat from fins on both sides versus the outside fins that only get external heat from one side, with a large spacing to the adjacent transistor end fin.

Interesting though as one wonders if the hot spot temps in Radeon VII is from the whole transistor area. If yes, that means that the internal fins might be very much hotter (40% hotter?).

 

[darkswordsman17]

PCH has nothing to do with the IMC.
AMD uses their own IMC, DDR PHYs are Synopsys.

Yet you're probably posting that from a phone, a world where vendors live and die on yearly cadence.
Ever wondered how does that work?

It's very much a new core.

I never said it did?

I could swear I recall people discussing AMD using someone else's memory controller in Zen 1? Perhaps it was an oversimplification or misinformation or misunderstanding.

Sorry, not on a phone. Have absolutely no clue what your point there is (fairly substantially different markets that have tended to operate in fairly substantially different manners for decades), as it is my understanding that it was pretty odd move for AMD. I'm not sure it was bad, but I recall people remarking about it. I would wager its better for AMD to do their own stuff. Arguably that was why they thought they'd do an ARM core back when they were considering what to do and wanting to co-develop K12 and Zen, they had a healthy amount of their own IP to utilize and could also enhance the standard ARM cores.

Are you saying Zen 3 will be all new core and not a revision of Zen or just saying what AMD has stated that its "Zen 3 architecture" which doesn't tell me anything other than what AMD is calling it. As I pointed out, Intel has called chips using largely the same architecture different things (mostly I think just to simplify marketing so they can tout it as the next Gen version of Core so that the customer sees it as better/newer). I'm personally skeptical that it'll be much of a change from Zen 2 (especially since it'll be one of if not the first AMD chip on 7nm+ and supposedly using EUV which I think is where most of the improvements it'll offer will come from; I think AMD will keep it fairly simple to focus on the substantial changes that EUV will bring to production). And considering how long its taking them to get Zen 2 and Navi out, I really am doubting that they'll have an all new core in Zen 3 out next year (unless its like Zen 2 where they had silicon back on it in 2018 but its not actually out in any products til this year - and even then sounds like Rome/EPYC based Zen 2 might struggle to make this year even though it was I believe the lead platform for Zen 2).

 

[DrMrLordX]

I had thought the IMC from Bristol Ridge at least was not in-house. If The Stilt were still here, he could probably add some information.

 

[jpiniero]

There is going to be a 50th anniversary edition product from AMD... but it's just a 2700X with a special edition cooler?

Says ETA is the 24th.

http://www.shopblt.com/cgi-bin/shop...011003000502_B6FR626P.shtml&order_id=!ORDERID

 

[Markfw]

The retail 2700x comes with a better cooler ??? Or do I have them mixed up ?

 

[Fanatical Meat]

The retail 2700x comes with a better cooler ??? Or do I have them mixed up ?

Yeah, looks like a regular 2700x with a $100 price premium

Maybe its like those golden sample things where its hand picked to run the highest clock speeds?
Maybe its some sort of BS scam like it comes with some AMD 50 coin or useless case badge. AMD isn't too good at managing their supply partners.

 

[maddie]

There is going to be a 50th anniversary edition product from AMD... but it's just a 2700X with a special edition cooler?

Says ETA is the 24th.

http://www.shopblt.com/cgi-bin/shop...011003000502_B6FR626P.shtml&order_id=!ORDERID

Extremely unimpressive.

 

[Markfw]

Extremely unimpressive.

They say not to order it anymore. Maybe it was a late Aprils fool joke ?

 

[moinmoin]

Yeah, looks like a regular 2700x with a $100 price premium

Maybe its like those golden sample things where its hand picked to run the highest clock speeds?
Maybe its some sort of BS scam like it comes with some AMD 50 coin or useless case badge. AMD isn't too good at managing their supply partners.

If it were a gold sample they should have gone calling it 2800x.

 

[HurleyBird]

If it were a gold sample they should have gone calling it 2800x.

Or at least, you know, a 2750x.