Speculation: Ryzen 3000 series

[chrisjames61]

I don't see a 16/32 cpu in the mainstream. No reason for it. That is HEDT area.

 

[dlerious]

That's not a fair comparison as the Threadripper is using multiple dice, and a lot of the power consumption is coming not from the core.

The 2950X has a 180W TDP. That's for a much bigger chip. It's going to be harder trying to cool when you put that in a chip with a smaller surface area.

 

[whm1974]

The 2950X has a 180W TDP. That's for a much bigger chip. It's going to be harder trying to cool when you put that in a chip with a smaller surface area.

I certainly will not buy a desktop CPU with that high of a TDP. Something like that is best suited to the ThreadRipper platform.

 

[Markfw]

I certainly will not buy a desktop CPU with that high of a TDP. Something like that is best suited to the ThreadRipper platform.

That is a threaripper

 

[whm1974]

That is a threaripper

Oops sorry about that. Through I'll admit that a 12c/24t CPU will be suitable for a high end desktop with DDR5 memory.

 

[DarthKyrie]

I don't see a 16/32 cpu in the mainstream. No reason for it. That is HEDT area.

That's what people said when it was leaked that Ryzen 1000 series was going to be 8c/16t. Intel disagreed with you then and they are disagreeing again with you now if you buy the latest rumor about them making a 10c CPU for the mainstream. Hell, I even seem to remember people saying the same thing when AMD made the first dual-core CPU. If Intel hadn't bribed OEM's to not use AMD back in the day we would have had 8c/16t mainstream CPUs back around 2010.

 

[whm1974]

That's what people said when it was leaked that Ryzen 1000 series was going to be 8c/16t. Intel disagreed with you then and they are disagreeing again with you now if you buy the latest rumor about them making a 10c CPU for the mainstream. Hell, I even seem to remember people saying the same thing when AMD made the first dual-core CPU. If Intel hadn't bribed OEM's to not use AMD back in the day we would have had 8c/16t mainstream CPUs back around 2010.

I was arguing that dual and quad core CPUs would be great for heavy multitasking.

 

[beginner99]

Consoles might have much lower clock speeds which means 8c chiplets that bin poorly (below ryzen 3 spec) might be suitible for use in consoles and selling those rather than throwing them away is of benefit.

On top of that it is a guaranteed demand so it means they know they can order more wafers which may help them negotiate lower costs per wafer.

Servers with 64-cores also have low clock speed so the low-leaage low clockign dies are actually the best for server CPUs which at the same time can ask for the most money. An Epyc 7601 32-core is now roughly $4200. I would expect a 64-core to cost more but let's assume not. A simple calculation of $4200 / 9 (8 chiplets + 1 IO die) = $466.67 per chiplet. Selling that same chiplet in a console when there is server demand would be a huge waste of money. And let's not forget. Servers move slowly. Till the Rome 64-core CPU is widely available and bough tit will be at least 1 rather 1.5 years from now, just look at current epyc which is only getting some traction now.

Only reason to sell the same chiplet for less than 1/4 in a console is if you expect not to have much demand in the server space.

Are they really that expensive? And as far as I know, Zen2 is designed specifically for TSMC 7nm HPC, so any 12/14nm would have to be Zen+, right? And PS5 will use Navi which is also TSMC 7nm HPC with larger die size than just 72mm². Sure PS5 could use a 7nm monolithic die with one 4C CCX but then if a critical defect hits the CCX the whole SoC would be unusable for PS5 and would be wasted. That's the beauty of the chiplet design, using the same 8C chiplet for vastly different markets.

Yes, that is what would make sense in a console. A low (2.5-3 ghz) clocked zen+ mixed with a 7nm GPU. Why? large volume of 14/12nm to fulfill the WSA. And yes I agree the GPU part must be 7nm else it will get huge for a console and because Navi was made partially with sony and is 7nm so yeah this seems given.

Of course as someone else pointed out Sony could simply take a rather large loss on the hardware upfront but I thought does days were over.

Because it might not be either one or the other but both at the same time, or are you saying that they can't get enough chiplets so they have to ration? Profit is profit and in a totally different market. We could also use the same argument against the desktop Ryzen as Rome has much higher prices and margins.

Yes if server takes of, the will certainly have a supply issue and yeah that will certainly also then reflect in Ryzen desktop prices. I outlined a basic calculation above in same post. For Ryzen AMD can take the higher leakage higher clocking parts which are less useful for server so there is less overlap to a console. And if a 16-core Ryzen sells for $500, it will still net a lot more profit than in a console (unless sony takes a loss on the hardware).

 

[ericlp]

I have the build it...and They shall come attitude. It use the be, that programmers were always waiting for hardware to catch up, now I think that the hardware is waiting for the software to catch up. Windows core management is a mess. Even worse 64/128 gigs of ram allocation is left to be desired by the OS. The more high end core CPUs coming out the better the software will be written to utilize them. When the OS can juggle multi core/threads well. Apps and games will speed up as a result. When that happens, pretty much everyone will be at 8/16 core CPU. Or else you'll be stuck running the older version of an OS that isn't updated well. Core up or be left out.... it will be that simple.

 

[alcoholbob]

According to Kyle Bennett at [H], most of what AdoredTV said was true. Assume he's under NDA and can't say more than that.

 

[PotatoWithEarsOnSide]

If the bottom tier 6c CPU is made from defective dies, it's cost of production is negligible. Its reasonable to assume that in factoring their wafer costs per die that AMD will base their entire cost per die solely on the fully functional dies. For instance, a $15k wafer yielding 600 fully functional dies, and 200 defective dies, would work out at $25 per fully functional die, with the 200 harvest able defective dies having zero cost. The costs getting attributed to dies that end up in the higher margin SKUs, which makes perfect sense.
Now, the IO die also has costs associated with it, so these would need to be factored in too, but overall you can still see a potential profit margin in those 6c defective dies sold at $99. There'd be 3 fully functional dies for every defective die, so as long as they have enough demand for 8c+ then they'll also have enough dies for the 6c segment.
It is a step change in our core needs, but one I think could be worthwhile if Zen 2 turns out to be pretty phenomenal. A less than phenomenal Zen 2 could go badly wrong for AMD's future prospects.
A big gamble for sure.

 

[Timorous]

Servers with 64-cores also have low clock speed so the low-leaage low clockign dies are actually the best for server CPUs which at the same time can ask for the most money. An Epyc 7601 32-core is now roughly $4200. I would expect a 64-core to cost more but let's assume not. A simple calculation of $4200 / 9 (8 chiplets + 1 IO die) = $466.67 per chiplet. Selling that same chiplet in a console when there is server demand would be a huge waste of money. And let's not forget. Servers move slowly. Till the Rome 64-core CPU is widely available and bough tit will be at least 1 rather 1.5 years from now, just look at current epyc which is only getting some traction now.

Only reason to sell the same chiplet for less than 1/4 in a console is if you expect not to have much demand in the server space.

Well another reason is because some chiplets will be below the bottom bin. If consoles represent an even lower bin it means there is less waste. It also means another 100,000 ish wafers (based on 20 million console sales a year) which gives AMD the ability to negotiate a lower cost per wafer and therefor reduce the cost per chip across their entire line up.

Yes, that is what would make sense in a console. A low (2.5-3 ghz) clocked zen+ mixed with a 7nm GPU. Why? large volume of 14/12nm to fulfill the WSA. And yes I agree the GPU part must be 7nm else it will get huge for a console and because Navi was made partially with sony and is 7nm so yeah this seems given.

Of course as someone else pointed out Sony could simply take a rather large loss on the hardware upfront but I thought does days were over.

Zen+ is already nearly the size of the current console APUs and so is raven ridge.

Sony and MS cannot get a generational leap over current gen by using anything other than 7nm because the die size and power requirements are just too much. I do not see next gen until 2020 BTW.

Yes if server takes of, the will certainly have a supply issue and yeah that will certainly also then reflect in Ryzen desktop prices. I outlined a basic calculation above in same post. For Ryzen AMD can take the higher leakage higher clocking parts which are less useful for server so there is less overlap to a console. And if a 16-core Ryzen sells for $500, it will still net a lot more profit than in a console (unless sony takes a loss on the hardware).

If AMD are ordering an extra 30,000 wafers for CPU chiplets alone it gives them more of everything meaning the very worst silicon still goes to the consoles (even if it is far above the minimum bin) leaving better silicon available for the rest of their product stack.

 

[Zapetu]

Selling that same chiplet in a console when there is server demand would be a huge waste of money.

Only reason to sell the same chiplet for less than 1/4 in a console is if you expect not to have much demand in the server space.

AdoredTV has talked about this in many of his videos and while he doesn't always get everything right (nobody does), I think he has a very good and valid point about chiplets and speed binning. Only the top 5% (*) (high leakage, high clocking) or so will be EPYCs and ThreadRippers and another top 5% (*) or more (low leakage, high performance/watt, lower clocks) (higher end) EPYCs an the more you sell the same 8C chiplets overall the more top end chiplets you will have. Sure some of those top chiplets have different characters (high clock speed or high perf/W) but nevertheless there are only about 5-10% of those really good chips/chiplets. Then you have 90-95% of those more avarage chips/chiplets and high volume console sales would be ideal for AMD to get rid of the most low end and below avarage ones. If AMD is planning to expand their server market share then they will need more of those top binned chips. And it's not only the top 5%, the top 1% would have even better chips. It's the avarage selling price (ASP) that counts for each wafer.

Yes, that is what would make sense in a console. A low (2.5-3 ghz) clocked zen+ mixed with a 7nm GPU. Why? large volume of 14/12nm to fulfill the WSA. And yes I agree the GPU part must be 7nm else it will get huge for a console and because Navi was made partially with sony and is 7nm so yeah this seems given.

If AMD doesn't for some reason want to give Zen2 to Sony/MS then sure, they could make a small 12nm zen+ chiplet with maybe only one 4C CCX. That would be a possibility. Still 7nm 8C chiplet would consume about half the power of a similar 14nm one (same performance, same number of cores) as AMD has stated. For the long run and future proofing the PS5, Sony really should go with the full 8C chiplet if possible all things considered. But i agree, that 12nm 4C CCX would be a possible choice if Sony and AMD have disgreements about manufacturing arrangements or some other corporate stuff (I hope not).

Of course as someone else pointed out Sony could simply take a rather large loss on the hardware upfront but I thought does days were over.

Or if the harware is really good and MS is late then Sony could ask some more money from the early adopters. Or they could go directly for market share and suck up the losses. It's really up to them.

And if a 16-core Ryzen sells for $500, it will still net a lot more profit than in a console (unless sony takes a loss on the hardware).

16-core Ryzen 3000 would still use top binned chiplets. Maybe not just the 5%, as Vattila pointed out here, but maybe still the top 10-15% of chiplets. There's still a lot of those avarage and below avarage chiplets that need to go somewhere.

Addition: (*)

1. As beginner99 points later out, more than 5% of the chiplets are used for different EPYC SKUs.
2. As H T C (and AdoredTV in is video) pointed out, only few chiplets are usable for the highest end SKUs per wafer. Therefore to have any volume for really top binned chips (much less than 5%) you need a large base volume of wafers.
3. As itsmydamnation points out even later, EPYCs and ThreadRippers come from different bins of chips (high perf/W or high clocking).

 

[darkswordsman17]

The 2950X has a 180W TDP. That's for a much bigger chip. It's going to be harder trying to cool when you put that in a chip with a smaller surface area.

The person I was responding to was claiming that it wouldn't be possible to provide the power for such a chip through the consumer Ryzen AM4 socket. My point was, saying its not possible then citing Threadripper power doesn't apply as so much of its power use was not directly related to the core and some of it won't carryover to a consumer AM4 chip.

As for TDP, again, it won't be a 180W chip, it'll be a ~100W chip. That's won't be that difficult to cool (especially for the overbuilt HSFs these days). Plus if they sandwich an I/O module between them (or alongside) it will likely end up not that much smaller than a single older Ryzen module. Plus, I doubt AMD suddenly goes "hey, let's not solder this stuff, let's put paste on it". I agree that I think they will need to start thinking about some new ideas though to get heat off these tiny crazy dense chips (if I remember, people have found patents about stuff like that by AMD, one was to have special dummy modules to space out the heat), but I don't think we'll be at the point where it'll be a major issues just yet. And I think we'll see aggressive thermal/clock management. Plus there's money to be made by setting up tiers, with special top-binned chips being paired with AIO watercooling or something.

But let's not forget that they've been keeping large console APUs cool despite those being rather large power hungry chips in a small form factor. I think Microsoft is using a vapor chamber on the One X so that's possible (I think one of the patents people found related to a heatspreader that served as a vapor chamber?).

 

[beginner99]

Only the top 5% or so will be EPYCs and ThreadRippers and the more you sell those same chiplets overall the more top end chiplets you will have.

My limited researched found that intel sells roughly 24 million server CPUs per year, rising. So if AMD takes 10% market share (including TR) that means roughly 2.5 million units per year. Each unit uses 8 chiplets so 20 Mio chiplets for Epyc and TR.

If only the 5% of chiplets are used for epyc then, that would mean they produce 20 Mio * 20 = 400 Mio chiplets per year. if we asume both ryzen and console all take 2 chiplets (consoles would probably only take one, so this is in your favor) that would mean they need to sell 200 Mio Ryzen and Consoles combined. Even if we combine all consoles including switch "only" 40 Mio units were sold.

And even if we half the number (5% server market share) it still is obviously not realistic to only use the top 5% for Epyc. So that "5%" can be easily debunked with trivial math.

If we keep the 10% share assumption (= 20 mio chiplets) and assume 30 mio console sales (xbox + ps) and more realistic assumption consoles only take 1 chiplet, consoles use only 1/3 more chiplets than epyc. As far as I could tell intel sells about 5 times for client cpus than servers. I would assume that is less for AMD because of the laptop market. let's assume 4xtimes more that would mean 10 Mio ryzens per year (it's probably less, didn't find any hard numbers). If each one takes 2 chiplets that is again 20 mio chiplets. So we have total of 70 mio chiplets of which 20 mio are used for epyc and TR which would be
28% of all chiplets, very far from the 5% suggested by Adored.

I suspect their main logical issue was that Epyc uses many more chiplets than consoles or Ryzen.

 

[Zapetu]

So we have total of 70 mio chiplets of which 20 mio are used for epyc and TR which would be 28% of all chiplets, very far from the 5% suggested by Adored.

Please note that not all EPYCs will be 64 core versions with 8 chiplets, only some of them will be that. There will likely be EPYC SKUs with 4 chiplets or even less, we don't know what the requirements for the IO die are and what configurations are supported. It may even be that all configurations from 1 to 8 chiplets are supported and that would drastically change your calculations.

I know that this is just speculation but kokhua have thought about different chiplet configurations. Sure that may not be exactly right but at least the 4 chiplet configuration looks very plausible.

But sure you are right that not all current EPYC SKUs use only the top 5% of chips and it could be maybe top 10-30% of the top chips. But while AdoredTV doesn't get every detail right, his main point is still valid on speed binning and a large pool of wafers to choose from.

Addition: I will give your post a like because it's good to correct and point out that not all EPYCs are only the top 5% chips.

 

[guskline]

Jumping from an 8 core/16 thread 2700x to a 16 core/32 thread 3850x or what ever AMD will call the next series of the highest end desktop chip seems like quite a leap. Especially since this is the next series of desktop Zen.

My uneducated guess is a 7nm chiplet-14nm IO-7nm chiplet combo with top speed of 4.7 in some cores and perhaps as many as 12c/24 threads.

 

[H T C]

Please note that not all EPYCs will be 64 core versions with 8 chiplets, only some of them will be that. There will likely be EPYC SKUs with 4 chiplets or even less, we don't know what the requirements for the IO die are and what configurations are supported. It may even be that all configurations from 1 to 8 chiplets are supported and that would drastically change your calculations.

I know that this is just speculation but kokhua have thought about different chiplet configurations. Sure that may not be exactly right but at least the 4 chiplet configuration looks very plausible.

But sure you are right that not all current EPYC SKUs use only the top 5% of chips and it could be maybe top 10-30% of the top chips. But while AdoredTV doesn't get every detail right, his main point is still valid on speed binning and a large pool of wafers to choose from.

Addition: I will give your post a like because it's good to correct and point out that not all EPYCs are only the top 5% chips.

Note that AMD doesn't have only the top "5%" of "perfect" 8c chiplets to choose from: it may also choose "5%"from 6c and maybe even 4c chiplets, depending on the total cores the Epyc chip has. Unless ofc the 12c / 24t or the 24c / 48t chips use solely 8c chiplets + a dummy die, which i highly doubt: would make more sense to use 6c chiplets to help distribute the heat more evenly and avoid any dummy dies, when possible. This isn't Threadripper, after all.

I used percentage within "" because, as you pointed out, that number may well be much higher, though i suspect it's actually lower.

 

[Zapetu]

It is a step change in our core needs, but one I think could be worthwhile if Zen 2 turns out to be pretty phenomenal. A less than phenomenal Zen 2 could go badly wrong for AMD's future prospects.

True, the chiplet design is a big change and will have both benefits and some disadvantages. But if AMD has chosen this path for desktop also then I'm sure that overall advantages are much greater than (some minor) drawbacks. People at AMD know x86 market very well and don't make hasty or misguided decisions.

Zen was already phenomenal compared to Excavator and AMD can't top that with Zen2. But sure Zen2 needs to be a little more than just good enough and I'm sure it will be.

Unless ofc the 12c / 24t or the 24c / 48t chips use solely 8c chiplets + a dummy die, which i highly doubt: would make more sense to use 6c chiplets to help distribute the heat more evenly and avoid any dummy dies, when possible. This isn't Threadripper, after all.

True, there's the point of heat distribution also. There are still many EPYC SKUs and not all of them need the highest performance possible (e.g. the 8 core ones). Still all IO needs to be enabled.

I used percentage within "" because, as you pointed out, that number may well be much higher, though i suspect it's actually lower.

It depends what EPYC SKUs are counted in. If you just take the most high end ones then sure it would be even lower.

 

[Atari2600]

My point was, saying its not possible then citing Threadripper power doesn't apply as so much of its power use was not directly related to the core and some of it won't carryover to a consumer AM4 chip.

Yes, *some* of it won't carry to an AM4 chip.

The PCIe lanes. The quad channel memory controller. Anything else? What would the wattage be there?


Feeding 16 cores with the right data in a timely manner still requires an awful lot of uncore power, that simply the nature of it. Perhaps the I/O controller of Zen2 will improve this. Perhaps Zen2 will be able to completely shut off modules to pull down idle and low thread power - perhaps not.

 

[Abwx]

The PCIe lanes. The quad channel memory controller. Anything else? What would the wattage be there?
.

With a 14nm I/O device that comprise all PCIE4 lanes and 8 memory controlers power is 0.5x at same CPU perf (about 0.9x the frequency) as stated by AMD, so we could imagine that it will be much less than this with a 2 channel I/O chip.

 

[DrMrLordX]

that 99$ 6c/12t would indeed really bring down ASPs. Even that's overkill for most people...

It would hit Intel so hard that Brian Krzanich would feel it in the unemployment line at CDK Global. Lisa Su be like f,d,df + HP SHORYUKEN. It would kinda make it hard for AMD to sell many of its older Zen/Zen+ chips as well. R3 1200? Pfft.

If that item represents an actual product with an actual on-the-street launch price, expect a). that product to come a few months after the high-price units go on sale (April? May?) and b). lots of great deals on Zen/Zen+ in the months prior to launch.

 

[Flayed]

I can't help thinking that $99 6c/12t will end up being more like $129.99 at launch, which is still $30 cheaper than what you can buy a 6c/12t cpu now.

 

[Shivansps]

The prices are probably wrong, it is also strange that there is not a 3200 or 3400/3500.

Thats more likely a $130 chip, with igp around $140. There is probably a 4C/8T or 6C/6T 3200 with IGP at $100.

 

[whm1974]

I can't help thinking that $99 6c/12t will end up being more like $129.99 at launch, which is still $30 cheaper than what you can buy a 6c/12t cpu now.

I highly doubt that at those price AMD will be sustainable, Unless the yields are very high.