AMD Zen 2 (Ryzen 3000) - *** NO COMPETITOR HINTING ***

[LePhuronn]

Apparently Ryzen 2 will launch with 12C/24T and AMD will only release a 16C/32T model when Intel responds...

I see the logic but that particular approach makes no sense. If CES proves to be accurate and we get core-to-core performance parity with Coffee Lake Refresh, then AMD already have a counter to a 10-core Comet Lake in the shape of a 12 core Ryzen.

And exactly what can Intel respond with? How much of a per-core performance uplift are we expecting with Comet Lake? And other responses have been laughable: apparently 28 core Cascade Lake-X is a counter to a 32 core Threadripper, Cascade Lake AP similarly has a 8-16 core deficit to not-even-released EPYC Rome, and apparently there are some new single socket Cascade Lake Xeon Golds coming soon which Intel are going to sell at half the price of EPYC Naples?

I'm not sure rehashing 14nm+++++++++++++ again with a lower core count to what you're responding to is any form of response at all. If it were me, I'd release a 16 core part purely to rub it in Intel's face.

 

[michty_me]

Ah, I didnt realise the rumoured 9 would be a new level release. I'm really behind as I haven't read much into processors since buying my i5-2500k.
I guess we wait for some concrete news.

 

[LePhuronn]

Ah, I didnt realise the rumoured 9 would be a new level release. I'm really behind as I haven't read much into processors since buying my i5-2500k.
I guess we wait for some concrete news.

Aye. It's a case of AMD using similar naming schemes to their competitors to give the general consumer an idea of where they roughly match. So Intel have used i3, i5 and i7 for a long time, so the introduction of Ryzen went Ryzen 3, Ryzen 5 and Ryzen 7. Then Intel introduced 8 core parts into an i9 bracket, in which AMD will mirror with Ryzen 9 for their 16 core parts (if the chart proves accurate).

 

[CAT-THE-FIFTH]

I see the logic but that particular approach makes no sense. If CES proves to be accurate and we get core-to-core performance parity with Coffee Lake Refresh, then AMD already have a counter to a 10-core Comet Lake in the shape of a 12 core Ryzen.

And exactly what can Intel respond with? How much of a per-core performance uplift are we expecting with Comet Lake? And other responses have been laughable: apparently 28 core Cascade Lake-X is a counter to a 32 core Threadripper, Cascade Lake AP similarly has a 8-16 core deficit to not-even-released EPYC Rome, and apparently there are some new single socket Cascade Lake Xeon Golds coming soon which Intel are going to sell at half the price of EPYC Naples?

I'm not sure rehashing 14nm+++++++++++++ again with a lower core count to what you're responding to is any form of response at all. If it were me, I'd release a 16 core part purely to rub it in Intel's face.

The 8C chiplets will be in demand for Epyc and Threadripper which are higher profit margin products. AMD is using TSMC 7NM which is relatively less mature than GF 14NM was at the time Ryzen was launched,so yield is lower.
Consumer products sell for less money than the server and HEDT products,so will get the runts.

So if Intel makes a 10C 14NM CPU,you need to consider it will be over 200MM2 in size,and if AMD gets close in single threaded performance and latency,they will be matching or beating the top yields of the 10C die,with a pair of salvaged chiplets.

Now,if Intel launches on 10NM with say 12 cores,its going to be a relatively large monolithic chip for early on an unproven process,against salvaged chiplets. Also by the time Intel moves to 10NM,yields should be up on 7NM so AMD can find it easier to supply consumer parts with fully enabled 8C chips.

For me I am more concerned about single threaded performance and latency.

This is where Intel wins in when it comes to benchmarks. 12C or 16C is not going to make as much of a difference especially since AMD has its own HEDT platform.

Most consumer software still finds it difficult to use 8C effectively especially since we have gone from 4C to at least 12C in the span of less than 3 years on a consumer socket.

 

[4K8KW10]

The 8C chiplets will be in demand for Epyc and Threadripper which are higher profit margin products. AMD is using TSMC 7NM which is relatively less mature than GF 14NM was at the time Ryzen was launched,so yield is lower.
Consumer products sell for less money than the server and HEDT products,so will get the runts.

So if Intel makes a 10C 14NM CPU,you need to consider it will be over 200MM2 in size,and if AMD gets close in single threaded performance and latency,they will be matching or beating the top yields of the 10C die,with a pair of salvaged chiplets.

Now,if Intel launches on 10NM with say 12 cores,its going to be a relatively large monolithic chip for earlu on an unproven process,against salvaged chiplets. Also by the time Intel moves to 10NM,yields should be up on 7NM so AMD can find it easier to supply consumer parts with fully enabled 8C chips.

The problem for the supply will be the wafer starts, not the yield, mind you!

 

[CAT-THE-FIFTH]

The problem for the supply will be the wafer starts, not the yield, mind you!

Yield is less than 14NM Ryzen was getting according to recent reports. So we don't know how many decent 8C chiplets will be available that can do 4GHZ to 5GHZ and if you have less wafers it only makes the problem worse. For things like servers it's less important when you are targeting 2GHZ to 3GHZ clockspeeds.

Also for consumer software which is years behind CPU development in what it can actually use,single threaded performance and latency is where I am more interested in seeing where AMD has made improvements.

I mean AMD might launch a limited production run 16C FX CPU at launch,so that could happen.

 

[LePhuronn]

But CAT, we've discussed this before, the chiplet approach means binning and allocation is not that black-and-white any more. AMD aren't going to waste a 5GHz-capable chip on EPYC.

All of the supremely power efficient 6 and 8 core chiplets will go to EPYC.
All of the supremely fast chiplets will go to the top Ryzen SKUs.
Threadripper could get the fastest chiplets (2950X is the fastest Ryzen 2000 product for example) but could equally get comparative cruft and still be a monster; a 6 core chiplet only capable of 4.5GHz is pretty poor on its own, but 8 of those gives you a 48 core, all-core 4.5GHz monster.

And again, there is more money to be made on high volume, low margin sales. Yes, a HEDT part may carry 5 times the markup of a desktop part, but if the desktop parts outsell HEDT by 10 times then desktop makes more money. It would be utterly ridiculous for AMD to gimp their desktop products because of the promise of higher margins on HEDT and enterprise.

 

[chrcoluk]

They won't be too bad I'd guess. 8 core Ryzen can be done for 65W already, so 7nm will bring that lower by a good chunk depending on clocks.

Navi is said to be 75 to 120W on desktop, so can expect that to be trimmed on console. Under 200w then which is around One X territory.

2600X without XFR is pretty efficient.

1.08v at 3.6ghz 45-55w during cinebench.

Gaming doesnt draw cinebench levels of power, so they will be efficient. The chips in the consoles will be the non XFR variants very likely effectively similar to non X models.

The performance/power curve is pretty bad at XFR but very good before it, voltages skyrocket for that last 10-15% performance, but they wont chase that on the consoles.

I expect the chips will be consuming 20-35w on average in games.

The same way that in a intel XTU AVX test my 8600k will top 125w (if I let it), but in a game its comfortably under 50w so even overclocked is barely half of its TDP in gaming.

 

[chrcoluk]

But CAT, we've discussed this before, the chiplet approach means binning and allocation is not that black-and-white any more. AMD aren't going to waste a 5GHz-capable chip on EPYC.

All of the supremely power efficient 6 and 8 core chiplets will go to EPYC.
All of the supremely fast chiplets will go to the top Ryzen SKUs.
Threadripper could get the fastest chiplets (2950X is the fastest Ryzen 2000 product for example) but could equally get comparative cruft and still be a monster; a 6 core chiplet only capable of 4.5GHz is pretty poor on its own, but 8 of those gives you a 48 core, all-core 4.5GHz monster.

And again, there is more money to be made on high volume, low margin sales. Yes, a HEDT part may carry 5 times the markup of a desktop part, but if the desktop parts outsell HEDT by 10 times then desktop makes more money. It would be utterly ridiculous for AMD to gimp their desktop products because of the promise of higher margins on HEDT and enterprise.

Do the two go hand in hand tho?

To run with less power the chip does X clock speed at less voltage.
But the chips that can do X clock speed at less voltage will also typically have higher peak clock speed capability.

Maybe I am wrong and it isnt that simple tho.

 

[John24]

I wouldn't blame AMD for holding back a 16 core to counter Intels reaction. Could use a 16 core though to run multiple encodes in parallel so hope it's not too long and falls within my budget. Threadripper would be ideal but it's too rich for me. A competitevely priced Ryzen 16 core with decent clock speed would be a dream come true.

Whatever happens it's really good to see AMD back and pulling punches.

 

[Minstadave]

Do the two go hand in hand tho?

To run with less power the chip does X clock speed at less voltage.
But the chips that can do X clock speed at less voltage will also typically have higher peak clock speed capability.

Maybe I am wrong and it isnt that simple tho.

It used to be the opposite, the really high clockers were leaky and less efficient.

 

[4K8KW10]

Yield is less than 14NM Ryzen was getting according to recent reports. So we don't know how many decent 8C chiplets will be available that can do 4GHZ to 5GHZ and if you have less wafers it only makes the problem worse. For things like servers it's less important when you are targeting 2GHZ to 3GHZ clockspeeds.

Also for consumer software which is years behind CPU development in what it can actually use,single threaded performance and latency is where I am more interested in seeing where AMD has made improvements.

I mean AMD might launch a limited production run 16C FX CPU at launch,so that could happen.

Chiplet size is only 70-78 sq.mm.
Zen 1 size is 213 sq.mm.

From the same quantity of wafer starts, TSMC will produce more 7nm chiplets with 70% yields than 14nm 213 sq.mm chips with allegedly higher yields.

 

[LePhuronn]

chrcoluk: There's always going to be that super elite silicon that does high clocks at low power. Where they end up in the Zen 2 plan I have no idea, but I've long held the belief that the rumoured 3850X would be a limited edition part based on that elite silicon. Otherwise I can see AMD binning coarsely on power draw at EPYC-level clocks first; anything that hits that criterion gets put in the "probably EPYC" pile and the rest moved into a 2nd. And then it gets more granular after that to determine what goes into which product and SKU.

 

[CAT-THE-FIFTH]

But CAT, we've discussed this before, the chiplet approach means binning and allocation is not that black-and-white any more. AMD aren't going to waste a 5GHz-capable chip on EPYC.

All of the supremely power efficient 6 and 8 core chiplets will go to EPYC.
All of the supremely fast chiplets will go to the top Ryzen SKUs.
Threadripper could get the fastest chiplets (2950X is the fastest Ryzen 2000 product for example) but could equally get comparative cruft and still be a monster; a 6 core chiplet only capable of 4.5GHz is pretty poor on its own, but 8 of those gives you a 48 core, all-core 4.5GHz monster.

And again, there is more money to be made on high volume, low margin sales. Yes, a HEDT part may carry 5 times the markup of a desktop part, but if the desktop parts outsell HEDT by 10 times then desktop makes more money. It would be utterly ridiculous for AMD to gimp their desktop products because of the promise of higher margins on HEDT and enterprise.

Chiplet size is only 70-78 sq.mm.
Zen 1 size is 213 sq.mm.

From the same quantity of wafer starts, TSMC will produce more 7nm chiplets with 70% yields than 14nm 213 sq.mm chips with allegedly higher yields.

But again why would AMD prioritise 8C chiplets at launch when you have limited volume for low margin consumer desktop over say server/supercomputer or HEDT?? TSMC 7NM is much less mature than GF 14NM was when Ryzen was made. It costs less per wafer too.AMD also has to contend with other companies using a new node,so there is how much capacity they have too. Remember,Navi is also on 7NM too.

The server chips tend to run at under 3GHZ and sell for a lot more money than consumer chips. The consumer chips will need to run at 4GHZ~5GHZ to make any sense for people on here. So you are asking for perfect silicon for pocket money. Even if you have 70% yields,how many of those will do the high clockspeeds people expect??

What if AMD at launch pushes a run of special edition 16C Ryzen 2 chips but at well over £500?? Athlon 64 FX anyone??

I fear some of you are expecting way too much of the hype train and you are going to be disappointed if one of the things don't work out.

Edit!!

Also what is this magic Intel CPU AMD needs to compete with in June??

Another 14NM CPU with an extra 2 cores plonked on?? How is that going to work with the extra power consumption and cooling required??

If its 7NM/10NM then its a large monolithic core against salvaged chiplets??

When is that going to appear?? End of the year or early next year?? By then 7NM will be yielding even better and be cheaper to boot.

I mean I am not saying AMD won't launch mass production AM4 16C Ryzen 2 CPUs in June,but then again I don't see why they NEED to,and why launching 12C now makes no sense.

It makes more sense financially to have a 12C SKU which can match what Intel has now in single threaded performance with 12C, than a 16C SKU which has problems clocking high due to needing good silicon quality or a lack of cooling due to the early nature of the node.

I would be more concerned if AMD needed to compete with the 8C Core i9 9900K or the next 14NM refresh with 16C,when Comet Lake S has apparently only 10C,and will amplify the problems we saw with the 8C CPUs when pushed.

 

[Random Guy Esq.]

Do we know anything about RAM compatibility for Zen 2 yet? Any indications that it will be improved?

 

[Grim5]

10 core 14nm from Intel is going to be hilarious.

It’s going to have lower single core performance than a 9900k unless overclocked and because of its design overclocking such a chip would mean something crazy like 300w tdp or even worse. The tdp already nearly doubled from 6 core to 8 core

In all honesty I’d love to have a 5ghz 10 core cpu, but there is no way I could justify the electricity bill and massive cooling solution

 

[Perfect_Chaos]

Anyone thinking we'll get leaks this month? Proper leaks?

 

[Mercutio]

This thing on the 23rd says yes.

 

[Zartar117]

Do we know anything about RAM compatibility for Zen 2 yet? Any indications that it will be improved?

Rome supports 3200 MHz ECC memory compared to Naples supporting 2666 MHz. Which is something I'm extremely excited about. So I'm expecting Zen2 to at least match Coffee Lake when it comes to RAM compatibility and max frequencies.

 

[Andrei87]

I for one wouldn't mind just a regular 8 core chip but with a massive IPC & latency improvement. This should be possible, since 7nm is a lot denser than 12/14nm, so a big core evolution is in the realm of possibility.