Ryzen 3000 Leaks

[WereCat]

36 minutes ago, VegetableStu said:

on mobile now, the thumbnail for the adored video mentioned something about navi. anyone knows what it's about?

For the NAVI APU you really have to watch his previous video as well to get a better picture, it's 1h long though. 

 

At the end of the video regarding this Zen2 leak he also mentions a bit about NAVI 12 dedicatef GPU leak. 

 

A 75W TDP only PCIe powered RX580 performance like NAVI GPU for around $130. 

Vega 56 like performance 120W TDP for around $200, etc.. 

[mariushm]

If the rumors are true, it seems like they're gonna make 2 dies  one with 2 CCX (4c/8t each) and graphics and one with 4 CCX (up to 16 cores /32 threads) and no graphics

The 3600g would have 2 fully working CCX units and 3300g would have one core and its thread disabled on each CCX. 

 

I'm wondering if AMD doesn't plan to go the same route as with Epyc and do "chiplets" on the 8+ core parts  ... basically have up to 4 smaller dies with just the cores (4c/8t each) and cache which all connect to a center " I/O die") that has the memory controller, pci-e fabric and SOC part (usb and sata)

Would be interesting but I suspect it would add to the cost of making the final chip, more steps, more testing etc  but like they plan on EPYC, it would allow them to make that I/O chip at a cheaper 12nm or 14nm process and they could change to DDR5/DDR6 controller for a future AM4+ socket or something like that, without remaking the Zen2 dies.

 

// and think about ... they could quickly mix these chiplets with gddr5 or gddr6 and you have the next  xbox or ps5 , if they're not going to customize the die used for ryzen + navi for such purposes.

 

It's what they're doing on EPYC like I said, since they have the know-how now, I don't see it being hard to bring it to AM4 (except cost reasons):

 

the epyc rome below , 64 cores , 4 channel, 128 pci-e lanes...  the i/o die is in center and made on cheaper 14nm , the chiplets ( 8 dies x 8 cores / 16 threads) are 7nm 

the i/o die could be much smaller on Ryzen as it would have only 2 DDR4 channels and only 24 pci-e lanes (16 video + 4 m.2 + 4 chipset)

 

 

 

 

[Taf the Ghost]

20 minutes ago, GoldenLag said:

another interesting thing is that AMD should have a 5% IPC advantage if the IPC gains are remotely true.

 

im still a bit skeptical about the clockspeed of these chips, among other things, though it seems to check out

It depends on the instruction types where the IPC gains happen.

[GoldenLag]

1 minute ago, Taf the Ghost said:

It depends on the instruction types where the IPC gains happen.

very true, we have no idea where they focused their IPC gains. as regular consumers we might be getting nothing, or a minimal amount. though 5-10% seem likely at the very least.

[Bananasplit_00]

i WANT this to be true, i doubt its 100% right but i hope it is

[Taf the Ghost]

@mariushm Both Sony & MS will have customized parts, but that doesn't mean the "framework" won't be similar. It's more noticeable with GPUs, but AMD can spin up customized versions of their GPUs on pretty short order, as far as silicon goes. They've clearly designed their systems to handle that approach now. So we'll see parts that are similar but not the same.

 

Main differences are likely to be certain instructions/features that are specific to their platform that they can leverage in a console situation.

[Taf the Ghost]

14 minutes ago, GoldenLag said:

very true, we have no idea where they focused their IPC gains. as regular consumers we might be getting nothing, or a minimal amount. though 5-10% seem likely at the very least.

On most normal tasks, AMD is level with Intel or actually slightly ahead. (There's a reason why those Render results are so good.) On the AVX2 and beyond stuff and the Front-End way they're all handled, Intel has an advantage in Desktop and a brutal advantage on Server/HEDT. The Stilts had the clock for clock IPC as between 25-30% on Intel for leveragable Workstation tasks.

 

Now, Xeons have more AVX units, which is why they do what they do so much better. Once Zen2 widens the AVX front-end to 256bit, a lot of that Workstation advantage is going to disappear. We'll see how much is left, but if the one workload, in that one slide, was 29% better, it's probably going to be near parity. Even IPC and a lot more cores, that's what AMD is going to be rolling out with. It's going to be fascinating to watch.

[GoldenLag]

2 minutes ago, Taf the Ghost said:

Even IPC and a lot more cores, that's what AMD is going to be rolling out with. It's going to be fascinating to watch.

im mostly interested how intel is going to react with this. rumoured Comet lake can be on its way early next year though i struggle to see how it is going to change anything with the exception of a couple of workloads. 

 

this is truly a nice time to be shopping for hardware (the time being about 10 months from now)

[ravenshrike]

2 hours ago, mxk. said:

ughhh I really want zen 2, but this mediocre PC I have right now just isn't cutting it. I've waited 4 years and I don't want to wait any longer, even if these leaks aren't true. 

A 2600+B350 is currently 210 dollars in the US. If you can make that up over the next 6-7 months(?) pull the trigger on the PC you want to build now with those two components in place of the Ryzen 3000 parts.

[Taf the Ghost]

4 minutes ago, GoldenLag said:

im mostly interested how intel is going to react with this. rumoured Comet lake can be on its way early next year though i struggle to see how it is going to change anything with the exception of a couple of workloads. 

 

this is truly a nice time to be shopping for hardware (the time being about 10 months from now)

They'll roll out a 250w 10c Desktop part. It's all they can do for a while. Icelake will eventually happen, but the schedule is still being pushed back. Earliest we'll see the Server parts announced is H2 2019. We should see Icelake in 2020, but there's a very real chance AMD will be on the second generation of Zen2 desktop parts before Intel is on 10nm. It's going to be ugly for Intel for a couple of years.

 

Intel's one saving grace? Intel is the world's largest Fab and by a massive amount. Most of the market has to buy Intel, otherwise there are no CPUs to be had, which would drive AMD's prices up. However, AMD can take a massive amount of market share in Server, which is high Wafer Volume, but not really high CPU # volume, by comparison. 

[Speakerator]

7 hours ago, Shreyas1 said:

I just got my 2600 darn it

Don't feel so bad - I just got my 2600 (k)  

[GoldenLag]

2 minutes ago, Taf the Ghost said:

They'll roll out a 250w 10c Desktop part

dont you mean 95 watt? /s

2 minutes ago, Taf the Ghost said:

Intel's one saving grace? Intel is the world's largest Fab and by a massive amount. Most of the market has to buy Intel, otherwise there are no CPUs to be had, which would drive AMD's prices up. However, AMD can take a massive amount of market share in Server, which is high Wafer Volume, but not really high CPU # volume, by comparison. 

manufacturing is one of those segments AMD will struggle to keep up. even if the entire world wants their CPUs they need someone to produce them. AMD will probably be buying any wafer they can from TSMC just to be certain they can push high volumes and that is the biggest weakness AMD has. they simply cant output what intel can.

 

AMD could be playing with close to 100% yield when binning CPUs, but even that is probably not going to be enough

[Taf the Ghost]

1 minute ago, GoldenLag said:

dont you mean 95 watt? /s

manufacturing is one of those segments AMD will struggle to keep up. even if the entire world wants their CPUs they need someone to produce them. AMD will probably be buying any wafer they can from TSMC just to be certain they can push high volumes and that is the biggest weakness AMD has. they simply cant output what intel can.

 

AMD could be playing with close to 100% yield when binning CPUs, but even that is probably not going to be enough

TSMC has a lot of capacity, though the vast majority goes to Apple. AMD could probably get to 35% in non-Server market share with what they can actually get from TSMC, but their prices would also go up to do it. If AMD could get to 25% in Server, they'd be able to buy expansion from TSMC on processes to keep them in the game. And we're also ignoring that AMD works with Samsung as well. (We forget AMD via GloFo used to be a big Flash player, and AMD still is a big Memory IP player.) AMD can find capacity to a pretty hefty amount, but they need the products and time to get there.

 

 

[paddy-stone]

If this turns out to be even close to true then my upgrade for next year is sorted. Seems a rather large jump in clock speeds, but really hope it's true.... I might also snag a Ryzen 3 for another server/NAS build.

[Mihle]

I have planned upgrading when Zen 2 come for quite a while.

If this so true, upgrading from my 3570k to Ryzen 3700 will be quite an upgrade.... Hell, even if I did 3600x it would basically be more than dubble the CPU power.

[GoldenLag]

6 minutes ago, Taf the Ghost said:

TSMC has a lot of capacity, though the vast majority goes to Apple

One of the more interesting things would be if AMD could partner with Apple to get access to some of their manufacturing capability. Though i doubt Apple will give up anything. 

[jiyeon]

9 hours ago, Deus Voltage said:

Ryzen 5 3600X: 8C/16T //  Base: 4.0 GHz //  Boost: 4.8 GHz // 95 W TDP // Price: 229$

I'm speechless if this is true.

[Taf the Ghost]

1 minute ago, GoldenLag said:

One of the more interesting things would be if AMD could partner with Apple to get access to some of their manufacturing capability. Though i doubt Apple will give up anything. 

Apple is paying TSMC something between 10-20 billion USD per year, for pretty much the new nodes are for Apple.

 

Though, the GPU capacity could easily end up as part of Apple's. Vega was pretty much for Apple, and they'll take on Navi at some point as well. (Or a Navi SKU with HBM. That'd be interesting.)

[Mihle]

1 minute ago, GoldenLag said:

One of the more interesting things would be if AMD could partner with Apple to get access to some of their manufacturing capability. Though i doubt Apple will give up anything. 

Is it unrealistic that AMD manage to pay more to TSMC per chip than Apple do?

[jiyeon]

1 minute ago, valdyrgramr said:

Gunna have to wait for CES next month to see if Lisa makes your jaw hit the floor.

I'll never look at my i5-8600K the same way again if Lisa does the thing.

[porina]

8 hours ago, Billy Pilgrim said:

However, I think it will be less of a problem than with the memory bottlenecking on the 2990wx. Now each ccx can directly access each stick of ram instead of having to do so over the infinity fabric

I'm not thinking so much about latency as bandwidth. That's a lot of cores competing for the similar bandwidth we had since dual core.

[leadeater]

21 minutes ago, Mihle said:

Is it unrealistic that AMD manage to pay more to TSMC per chip than Apple do?

Apple has what is called "F U Money", AMD doesn't have that . Apple gets what Apple wants even if you don't want to give it to them, because "F U Money" 

[GoldenLag]

10 minutes ago, Taf the Ghost said:

Apple is paying TSMC something between 10-20 billion USD per year, for pretty much the new nodes are for Apple.

 

Though, the GPU capacity could easily end up as part of Apple's. Vega was pretty much for Apple, and they'll take on Navi at some point as well. (Or a Navi SKU with HBM. That'd be interesting.)

considering the Buzz around AMD joining up with Microsoft for future Surface products it wouldnt be all that surprising if Microsoft decided to put some cash down for AMD aswell. how does the Glofo deal work when AMD isnt the one buying the Wafer, but instead is receiving them from someone like Apple or Microsoft for contract manufacturing. 

[axido]

so is this real, then? 

the performance numbers are looking amazing and I don't want to get hyped up over nothing

[GoldenLag]

1 minute ago, axido said:

so is this real, then? 

the performance numbers are looking amazing and I don't want to get hyped up over nothing

its at least within what one would expect to be comming, but this could very much be some clever guesswork from some random. give me like 6 hours and i could probably have made something similar.

 

though a bit too much seemed right for it too be 100% fake