16-Core Ryzen 9 3000 Series ES Sample spotted & Zen 2 is a Memory OC Beast, DDR4-5000 Possible! (Updated)

[porina]

5 hours ago, Ithanul said:

Hopefully they can do it.  Though, what I read, even these newer ones will set the IF to a 1/2 divider mode if trying to go higher on RAM speed.

Even if these speeds come to pass, we don't know about the width. For current Zen my understanding is the IF bandwidth is equal to the ram bandwidth. If they made it half as fast, but twice as wide, that could remain the case. It will be interesting to hear more about the IO die.

 

5 hours ago, Ithanul said:

I been following 1usmus for some time since I found his Ryzen RAM calculator.

Have you used that? I saw a new version come out recently. been meaning to try it some time as I've not had much luck with manual OC of B-die on Ryzen, topping out around 3400.

 

34 minutes ago, leadeater said:

Be a sad day when your memory out clocks your CPU lol.

I'd be happy with that, assuming consumer sticks to 2 channels for the foreseeable future. It means I might be able to feed it with enough data so it doesn't just sit around twiddling its thumbs. For a dual channel ram coupled with Intel quad core, I'd estimate the "practically unlimited" bandwidth point occurs with a ram speed somewhere above core clock speed. Basically anything with more than 4 cores and a good FPU is limited AF right now (for my uses).

[cj09beira]

5 hours ago, Stefan Payne said:

I mean think about it, DDR4-5000??? Why? For what??

Or do they mean that its half way to DDR-5??

Here Wikipedia:

https://en.wikipedia.org/wiki/DDR5_SDRAM

 

THAT sounds more likely that there is some DDR-5 SDRAM Stuff inside than "DDR4-5000".

i would read that as during benchmarking you might be able to post at 5000 mhz ram clock, not more than that 

4 hours ago, imreloadin said:

You mean to say that these are actually in the realm of realism whereas all the 5 GHz crowd is pretty much in bat shit crazy land

 

There will NOT be a Ryzen 3000 series CPU that turbos up to 5 GHz, anyone believing this is just setting themselves up for disappointment...

though people shouldn't expect 5ghz, its really not out of the question, just look how far people have managed to clock vega 20, they have been able to push it to 2200mhz, from vega 10's what? 1800 on a good day, which is a huge 22%

add the fact that binning will probably be more rigorous (bad for us that like to Oc the lower end parts) and its really quite reasonable, though myself i am expecting 4.8 thats what i would be happy with

 

[RejZoR]

2 hours ago, leadeater said:

Be a sad day when your memory out clocks your CPU lol.

It's sill good to see that Ryzen can now pump out such clocks on memory, knowing how limited it was on that regard in its first iteration. I wish they could do the same on the CPU clocks side too. Coz when they manage to do that, they'll devastate Intel whose only real benefit at the moment is the higher CPU clock.

[MeatFeastMan]

4 hours ago, RejZoR said:

So, they meant 5GHz on the memory, not on the CPU cores XD

5000mhz is insane. Can't be true surely.  If so, my 3466 dual rank 2x16gb kit I've just paid silly money for is golden. All I've heard about the memory on zen2 is that it's a small improvement. (On that article about it having a 4.5ghz boost and 15% IPC)

 

And for all those doubting they'll be more than 8 cores at launch, remember what Lisa said in her interview?

 

 

She said to 'expect' more than 8 cores. That didn't include the words 'at launch' but the way she said it hinted to the same launch as the rest.

 

 

 

[Neftex]

7 hours ago, imreloadin said:

You mean to say that these are actually in the realm of realism whereas all the 5 GHz crowd is pretty much in bat shit crazy land

 

There will NOT be a Ryzen 3000 series CPU that turbos up to 5 GHz, anyone believing this is just setting themselves up for disappointment...

"hurr durr 5ghz not possible". meanwhile intel is on 5ghz+ on old node...

 

now, im not saying it will be 5ghz but you outright saying it wont seems just as stupid as saying it will

[leadeater]

9 minutes ago, Neftex said:

meanwhile intel is on 5ghz+ on old node...

Intel's nodes are specifically designed for high power and high performance, GloFo and TSMC don't target that as their biggest market by far is mobile and other low power chips. Old doesn't matter when it's made purpose for the task, node names mean nothing. You might as well just use animal names instead of numbers and nanometers.

[Trixanity]

1 minute ago, leadeater said:

Intel's nodes are specifically designed for high power and high performance, GloFo and TSMC don't target that as their biggest market by far is mobile and other low power chips. Old doesn't matter when it's made purpose for the task, node names mean nothing. You might as well just use animal names instead of numbers and nanometers.

That's why it was disappointing GloFo shut down its 7nm development. They had a high power node for IBM in the works that specifically targeted 5+ GHz.

[Neftex]

1 minute ago, leadeater said:

Intel's nodes are specifically designed for high power and high performance, GloFo and TSMC don't target that as their biggest market by far is mobile and other low power chips. Old doesn't matter when it's made purpose for the task, node names mean nothing. You might as well just use animal names instead of numbers and nanometers.

doesnt change anything, we see 5ghz cpus are possible. and noone here knows the process and architecture that well to outright say it cant happen on zen2

[leadeater]

5 minutes ago, Neftex said:

doesnt change anything, we see 5ghz cpus are possible. and noone here knows the process and architecture that well to outright say it cant happen on zen2

Neither can anyone say it will happen. We know more about what TSMC does do than we know about what they are going to in conjunction with AMD and 7nm, there are also architecture limits to how a CPU will clock and fab nodes don't negate that. We know TSMC targets power efficiency and we know Zen/Zen+ have very rapid and hard clock limits and the CPU core architecture is not significantly changing other than the FPU which will actually cause lower clocks re: Intel AVX offset.

[Drak3]

1 minute ago, leadeater said:

not significantly changing other than the FPU which will actually cause lower clocks re: Intel AVX offset.

The offset that only applies for AVX workloads, meaning that AMD can still achieve higher clocks for everything that isn't AVX.

[leadeater]

1 minute ago, Drak3 said:

The offset that only applies for AVX workloads, meaning that AMD can still achieve higher clocks for everything that isn't AVX.

That's also if AMD implement, or even need to, an offset. Beefing up the core with larger caches, wider FPUs etc means more power and lower clocks. Zen is already very strong architecturally, performance wise, and it doesn't actually need to hit 5Ghz clocks like many want or expect.

 

Either way if it were my money and I was placing a bet I would be putting it on less than 5Ghz. I'm happy to lose a bet that benefits me and everyone else than lose one and the outcome is worse.

[cj09beira]

18 minutes ago, Trixanity said:

That's why it was disappointing GloFo shut down its 7nm development. They had a high power node for IBM in the works that specifically targeted 5+ GHz.

that node was so sexy, though i think samsung will continue it as they had a joined research venture 

9 minutes ago, Drak3 said:

The offset that only applies for AVX workloads, meaning that AMD can still achieve higher clocks for everything that isn't AVX.

amd confirmed already that they wont do avx offsets, at least not in the epyc side.

on the consumer side my guess would be that normal xfr would apply which means it would be power and temp based (which we can increase  )

[Drak3]

1 minute ago, leadeater said:

it doesn't actually need to hit 5Ghz clocks like many want or expect.

It absolutely doesn't. IMO, the performance sweet spot, before hitting diminished returns, is between 4 and 4.2GHz, where many Ryzen 2 chips are already capable of hitting 4.2GHz on a manual overclock and a few cores can hit 4.3 on XFR.

 

 

Personally, I'd much rather see 28 PCIe lanes coming off the chip, and one of the two following:

Latency improvements on IF

Some type of setting in Windows to keep programs running on only on CCX/Die if it doesn't benefit from more cores/threads than the CCX/Die provides.

[leadeater]

18 minutes ago, cj09beira said:

that node was so sexy, though i think samsung will continue it as they had a joined research venture 

Sadly not, that was being developed for IBM over on the fab facilities that GloFo got in the merger/buyout of IBM fabs.

 

Quote

To ensure that GlobalFoundries remains competitive against Samsung Foundry and TSMC in the long run, Sunjay Jha obtained IP and development teams from IBM (along with two fabs and a lot of obligations), and poured in billions of dollars in development of the 7LP fabrication technology platform. The latter would include three generations of manufacturing processes and, possibly, a custom high-performance technology available exclusively to IBM. While everything appeared to proceed smoothly with the 1st Gen 7 nm process (DUV only), the 2nd Gen 7 nm process (with EUV used for non-critical layers, such as padding) still needed some additional development investments, and the 3rd Gen 7 nm (with intensive usage of EUV) required even more money for development and further could require installation of additional EUV equipment. Meanwhile, there were two things to consider.

Only chance we'll see it is if IBM can buy the rights to it, or has it already, and is allowed to give it to Samsung. But Samsung had no involvement with this particular 7nm.

[Hifihedgehog]

9 hours ago, Stefan Payne said:

Depends on the Workload and how much Bandwith you need.

If you need the bandwith (OR I/O) Threadripper is till better.

 

And they probably won't kill Threadripper, it will just get wider -> up to 32 Cores, maybe 64 Cores?? But I doubt 64 COres because of clockrates...

Unless you have 3600 MHz or faster RAM which helps fill the dual-channel bandwidth gap for Ryzen.

[cj09beira]

12 minutes ago, leadeater said:

Sadly not, that was being developed for IBM over on the fabs facilities than GloFo got in the merger/buyout of IBM fabs.

 

Only chance we'll see it is if IBM can buy the rights to it, or has it already, and is allowed to give it to Samsung. But Samsung had no involvement with this particular 7nm.

its such a shame, it was looking so good

[Stefan Payne]

18 minutes ago, Hifihedgehog said:

Unless you have 3600 MHz or faster RAM which helps fill the dual-channel bandwidth gap for Ryzen.

...wich you can also use for Threadripper.

ANd the reason for RAM giving so much performance on Ryzen might actually be the Infinity Fabric colocks wich is syncronous with the memory clock. Higher memory Clock -> higher IF Clocks...

Wich is the next point: With Zen2 that changes and Memory Clock is independant from the IF Clock.

[Doobeedoo]

16-core part and being ES too. Also we don't know how good power delivery will higher end AM4 X570 boards have yet opposed to TR4 ones. As for memory controller and faster memory speeds improvements that's great, how much we'll see getting higher speeds easily with kits without extra tinkering I wonder.

Though I'm mainly interested to see a lesser core count frequencies like 8-core part. Can definitely expect higher clocks there.

[Stefan Payne]

1 hour ago, leadeater said:

Either way if it were my money and I was placing a bet I would be putting it on less than 5Ghz. I'm happy to lose a bet that benefits me and everyone else than lose one and the outcome is worse.

I agree.

You should hope for the lower outcome wich is 5% "IPC" (except AVX) and 10% Clock tops.

 

15% Clocks and more than 5% IPC would be nice (and might be possible, as I disagree that they didn't work on the cores. THe team might have started on it as soon as Zen was finished and a major shrink isn't a small feast. That is basically a rework of the core!)

 

 

It might also be possible that both is true at the same time:
It doesn't come with 5GHz but the CPU can work at 5GHz.

 

But in 2 Weeks and 2 Days we know more...

 

[S w a t s o n]

1 hour ago, Trixanity said:

That's why it was disappointing GloFo shut down its 7nm development. They had a high power node for IBM in the works that specifically targeted 5+ GHz.

True but TSMC also has HP 7nm, everyone in here acting like TSMC only makes phone SoC's

[leadeater]

27 minutes ago, Stefan Payne said:

as I disagree that they didn't work on the cores

Oh I don't mean they didn't but the architecture is largely the same, things like pipeline stages. There should be cache improvements in size, performance/latency and logic (cache hit ratio) across each layer in some combination of those 3 aspects. But because it's the same architecture the clock behavior won't significantly change, it'll be an iterative process much like Intel over the years with larger gains coming from joint node improvements and required optimizations to accommodate that.

 

And by architecture I mean that much like engineering building schematics. You can take the same plans and scale the size of it however due to material properties you must make specific changes due to that size change but it's still the same building, just of a different size.

[LAwLz]

I often see "is often right" and "have been right in the past" mentioned regarding these types of leakers, but it would be interesting to see how often they are right.

A lot of times it seems like it's some Twitter account which just posts like 10 "leaks" a week and a few of them happens to be correct. It would be fun if someone wrote down exactly what they "leaked" and then compared how many things were right, and how many things were wrong.

[leadeater]

Just now, LAwLz said:

I often see "is often right" and "have been right in the past" mentioned regarding these types of leakers, but it would be interesting to see how often they are right.

A lot of times it seems like it's some Twitter account which just posts like 10 "leaks" a week and a few of them happens to be correct. It would be fun if someone wrote down exactly what they "leaked" and then compared how many things were right, and how many things were wrong.

And what is right? Many times I've seen 'close enough' taken as right, then also what is close enough? Plus we're more likely going to remember the times someone was right more than wrong for this type of situation, reinforcement bias etc.

[LAwLz]

1 hour ago, leadeater said:

And what is right? Many times I've seen 'close enough' taken as right, then also what is close enough? Plus we're more likely going to remember the times someone was right more than wrong for this type of situation, reinforcement bias etc.

I thought of tracking all the rumors for an upcoming product launch before. Write down who said what and when, and then after the product has launched, go back and check how much was correct. 

 

Maybe I'll do it for the next big product launch. 

[Trixanity]

53 minutes ago, LAwLz said:

I thought of tracking all the rumors for an upcoming product launch before. Write down who said what and when, and then after the product has launched, go back and check how much was correct. 

 

Maybe I'll do it for the next big product launch. 

Sounds like a lot of work with virtually no pay off.