Return to Level1Techs.com

Ryzen 3000, GPU speculation

amd

#21

Except maybe to make use of partially defective dies. The 5GHz make me very suspicious though, that’s just too good on a new node at the end of what is possible on silicon.


#22

See, the NAVI predictions I don’t quite trust. We are basically talking 50% power draw for the 580lvl performance, and I don’t think 7nm is that powerful.
Vega 64+15% performance is entirely possible, but I expect basically the same power consumption, which will be painful for the Radeon branding… They had massive issues with power draw for a while now and a new GPU nowhere near 2080 with larger power draw will be pure misery, unless the price is right and it really is something like 250 dollars… Then we can talk…


#23

Yeah, if they design for up to 16 cores on AM4, they can make use of any combination of “defective” chiplets.

e.g., Ryzen 8 core may be 2 half defective dies. Ryzen 4 core, 2 mostly defective dies.

The beauty of what AMD is doing is being able to use dies that are massively defective in their lower end products like consoles, low core count AM4, etc., whilst the fully functional high bin parts go in threadripper or epyc. Their yields even on the brand new 7nm process are going to be far better than a monolithic design. Purely because so much of the die can be defective but still useful.

I’m personally expecting both the chiplets and IO die to have much larger than normal caches in them in order to somewhat mitigate the latency hit from moving the memory controller off-die… and also alleviate the bandwidth problems incurred by sticking so many cores in a socket.

In fact, check out this epyc photo…

See the central IO “die”. Looks like 2 dies to me, or am i seeing things? Maybe that bottom segment is HBM or similar high bandwidth local cache memory? Could just be a bad photo…


#24

I agree. and AMD had been underwhelming on GPU’s for a few years.


#25

What you are seeing is a reflection.


#26

i think those leaks are likely fake.

It would make them more expensive than zen1 series since it bases on
2/4 of the IO chip; and 2 chiplets (making the silicon of the die 2x the size of zen1) - The IO itself would be at the price of zen1 chip.
(it should be 1/4 io die, and single chiplet - if its not then its not profitable or price should be increased by at least 30%.)

  • rome; had 4/4 io die while having 8 chiplets. (so 2 chiplets per 1/4 io die.)

so… doubtful move from amd unless the release will be much later Q3-4 2019.
(the clock bumps are very likely tho)

then those GPU leaks… well shitty Likely false… I’m like 99% sure amd won’t change their naming scheme.


#27

to be fair, he does a good job of communicating that (and also which assumptions he is making)

Personally I like the few mostly technical videos the most, but I guess there is not enough audience for that.


#28

All seems plausible, even a few of us though of much of these in a basic form when they were showing off the last CPU that kicked off the tiny 7nn 8core chiplets.

So I would not be surprised at all to see this happen. Some of it is a bit speculative but that’s what it is. He does say pretty clearly that everyone should take it with salt and has explain this reasoning a bit.

Sounds about right or at least all of it is doable.


#29

I’d guess the 5.0GHz boost frequency is probably for only 1-2 cores. I’d expect the all core to be much lower, around 4.3-4.5 for a top sku. There’d be way too much heat otherwise for that many cores to hit that TDP target, unless AMD is pulling an Intel and lying about the actual TDP lol.


#30

You mean like AMD didn´t name their AM4 chipsets very similar to how Intel named theirs?
B360 - B350
H370 - X370
X299 - X399

Taking the piss and going for 3000-series before Nvidia does would fit that scheme perfectly.


#31

Yea, the 10 Ghz promise was announced at the end of 2000, but they said “in 5-10 years” so the earliest point for it to actually happen would be 2005.

There’s a lot of reasons it didn’t happen, but it’s one indicator of why this isn’t an impossible idea at this point. The clock race died down for many reasons, not all of which were for technical limitations, strictly speaking, so at any point we could see a return to higher clocks if either Intel or AMD think the conditions are right for it.


#32

With 7nm, we are at the point were silicon won´t really allow for further shrinks.
Then the clock race is on again. And Zambezi on LN2 could do 7GHz allready…


#33

AMD naming products is pure trolling. Which is a shame really. The products are so strong, they don’t need that.


All I say for now is… if that 3700X is legit with those specs and no downsides in games … I’m gonna switch to that from my 2700X the second it is available.

That is a five year chip on a solid platform with mid range prices.


#34

Didn’t it do 8 on LN2? Or are you talking about multi core?


#35

My mass response to all the things that interest me.


If you look, he kinda clarified that. He never outright said “6 core chiplet” so much as “single CPU die with 6 active cores” being how I heard it.

I’m a bit skeptical of more than 5GHz unless they’ve made some architecture changes. The combination of power savings on top of clock increases just seems too good to be true. 5GHz alone seems reasonable enough though.

Two words: IO Die.

The memory DIMMs are wired directly to the CPU, right? So your memory channels are dicatated by the memory controller on the CPU. Meaning you can have 4 dimm, 4 channel memory on AM4, theoretically. (I’m making a fair bit of assumption here though) I know quad channel still isn’t quite enough, but I’m guessing we’ll see memory controller maturity coming up as well, so we may see 3800mhz being not unrealistic for the highest end systems if people are willing to shell out for DRAM that’s capable.

Here’s my stance: I try not to hype over it, but the problem I have is that we’re looking at someone who’s pretty much been right on everything since the launch of first gen Ryzen. The biggest thing he got was the 8x 7nm chiplets with a 12/14nm IO die. This guy has his finger on the industry’s pulse. To say otherwise would be flat out ignorance. If you believe his leaks, that’s another story entirely. I, personally, have no reason to not believe them, but if anything, this simply makes me want to hold off on buying AMD until CES, when we hear about third gen Ryzen.

What makes you say that? The limiting factor on frequency was not, and is not the Zen architecture, even so, 3rd gen ryzen is Zen 2, not Zen+. Additionally, 7nm was claiming 30% frequency increases over 16nm. let’s take the 2700x boost clock of 4200MHz, multiply that by 1.25 (because the 2600x was 12FF, not 16nm) and get 5250MHz. I’d say it’s absolutely realistic to expect somewhere near 5GHz. 5.1? Maybe not. 4.8? Absolutely. I’ll be disappointed if not. Keep in mind that these frequencies are all boost clocks.

You can get 4 IO chips from one IO die, not to mention that with the right architecture, even on 12FF, you’d get 99.5% yield.

I’d agree. Not to stay in spec. that said I’m fairly certain, we’ll be able to run it out of spec on 4.8-5.0 on all cores.

I think AMD is having a great time fucking with everyone elses product line naming schemes.

Nobody I know thinks it’s bad and anyone who noticed thinks it’s awesome. I don’t think it detracts from it. And honestly, when I think of “the 300 series chipset” I think AMD, not Intel, so if that’s their goal, it’s worked.


#36

You do now. :wink: I would even go as far as saying there is not a lot of ways they could have done it worse.

Why Ryzen 3, 5 and 7? Why not Ryzen 4, 6 and 8, the number of cores?
X marks multithreading and taking a page out of the book of GPU boardpartners call the higher clocked version OC. A good naming scheme, gamery, descriptive, easy to read.

I don’t care if it worked or not. It is modeled after Intel’s product palette while it could have been it’s own identity.


#37

It could have been but to the average consumer who doesn’t know any better it is much easier. Especially since AMD has been absent for so long in this space. It was a good call to have a direct comparison to Intel’s line. Though I do think the 4/6/8 would have been good too. Though once they started adding more cores it would have been a problem.

Also Kyle from HardOCP on the leaks


#38

Zen may very well scale beyond the capabilities of 7nm (it does outscale 12nm).

Definetly.


#39

That’s a fair argument and I agree there.

Well, that’s the thing. 7nm is Zen2. It’s a different uArch.

We could see better, we could see worse. I think they’re definitely leaving some performance on the table, but I think we’re going to see Ryzen run into heat problems on the high end (if believing leaks, I’ll call that the 12 core processors) and be severely limited by cooling, whereas I could run my 1700 at 3.9GHz on the stock cooler.


#40

Splitting IO and having two chiplets running six cores each should spread the heat quite a bit. That can have dramatic effects. Think about what those relatively simple noctuas were able to pull off on TR4 just because they had a bigger base plate.