Thursday, May 9th 2019

AMD Ryzen 9 3000 is a 16-core Socket AM4 Beast

AMD is giving finishing touches to its 3rd generation Ryzen socket AM4 processor family which is slated for a Computex 2019 unveiling, followed by a possible E3 market availability. Based on the "Matisse" multi-chip module that combines up to two 8-core "Zen 2" chiplets with a 14 nm I/O controller die, these processors see a 50-100 percent increase in core-counts over the current generation. The Ryzen 5 series now includes 8-core/16-thread parts, the Ryzen 7 series chips are 12-core/24-thread, while the newly created Ryzen 9 series (designed to rival Intel Core i9 LGA115x), will include 16-core/32-thread chips.

Thai PC enthusiast TUM_APISAK confirmed the existence of the Ryzen 9 series having landed himself with an engineering sample of the 16-core/32-thread chip that ticks at 3.30 GHz with 4.30 GHz Precision Boost frequency. The infamous Adored TV leaks that drew the skeleton of AMD's 3rd generation Ryzen roadmap, referenced two desktop Ryzen 9 parts, the Ryzen 9 3800X and Ryzen 9 3850X. The 3800X is supposed to be clocked at 3.90 GHz with 4.70 GHz boost, with a TDP rating of 125W, while the 3850X tops the charts at 4.30 GHz base and a staggering 5.10 GHz boost. The rated TDP has shot up to 135W. We can now imagine why some motherboard vendors are selective with BIOS updates on some of their lower-end boards. AMD is probably maximizing the clock-speed headroom of these chips out of the box, to preempt Intel's "Comet Lake" 10-core/20-thread processor.
Sources: TUM_Apisak, Tom's Hardware
Add your own comment

197 Comments on AMD Ryzen 9 3000 is a 16-core Socket AM4 Beast

#101
bug
EarthDog said:

It's comical, the true fanboys don't get called out, yet, people that have at least a half a clue (in this case more, a full clue) are called fanboys. TPU members are AMAZING! o_O
It's not specific to TPU, but the rule of thumb on the Internet is it's ok the be a fanboy as long as you root for the underdog, not for the top dog. Do you get that EarthDog? :D

kapone32 said:

I really don't understand why some people think that this is not possible. I know it is different architecture but did not the same thing happen with the FX8150 4.5GHZ max vs the FX8350 over 5 GHZ OC possible. in terms of clock speed increase. Even the 1700 could not go past 4 GHZ but the 2700 goes to 4.3? Do you not think a node shrink of that size would not bring those types of gains. The Vega 64 was 1247-1546 base and the Vega 7 is 1400 - 1750 base. A CPU has way less cores than a GPU so does it not make sense that you would see bigger boost on the CPU side.
While I won't say it's impossible, here are the things that make me take this with a grain of salt:
1. So many cores, clocked so high and only 135W TDP
2. Feeding that many cores from only two RAM channels.
Posted on Reply
#102
Patriot
Hate to be the bubble burster (not really) clock speeds are going to start to decrease from here.
Stop expecting monster clock gains every gen, the clock wars are over. 7nm EUV and 5nm will probably have lower base than 7nm.
You are just setting yourself up for disappointment.

IPC can keep getting better with arch tweaks and Density will of course rise to a point, then die stacking will have to take over as physics wins.
Posted on Reply
#103
R0H1T
bug said:

While I won't say it's impossible, here are the things that make me take this with a grain of salt:
1. So many cores, clocked so high and only 135W TDP
2. Feeding that many cores from only two RAM channels.
Same reason why Intel can go from 4 cores to 8 with the same TDP.

IF2 will deliver over twice the bandwidth as compared to previous gen, pretty sure that counts for something.
https://www.techpowerup.com/253954/amd-ryzen-3000-zen-2-bios-analysis-reveals-new-options-for-overclocking-tweaking
Feeding 16 cores isn't easy but if you look at it discreetly, there aren't many consumer applications that are bandwidth starved at the speeds Zen2 will run it's memory.



https://www.tomshardware.com/reviews/best-ram-speed-x470-pinnacle-ridge,6064-6.html
Posted on Reply
#104
TheLostSwede
bug said:

While I won't say it's impossible, here are the things that make me take this with a grain of salt:
1. So many cores, clocked so high and only 135W TDP
2. Feeding that many cores from only two RAM channels.
No-one said memory performance was up, did they? If so, then I have missed that.
As for the TDP, I guess we'll see, but a Threadripper 1950X is 180W and has just as many cores, but at 14nm, so it doesn't seem impossible.
Salt is always good, just don't add too much, as that's not good for health... Personally I prefer ammonium chloride, although only in my liquorice.

Patriot said:

Hate to be the bubble burster (not really) clock speeds are going to start to decrease from here.
Stop expecting monster clock gains every gen, the clock wars are over. 7nm EUV and 5nm will probably have lower base than 7nm.
You are just setting yourself up for disappointment.

IPC can keep getting better with arch tweaks and Density will of course rise to a point, then die stacking will have to take over as physics wins.
You're most likely right, this will be one of the final pushes in clock speeds for current technology. I doubt we'll see retail chips based on silicon running above 6GHz with typical PC cooling. This is obviously why we're going wider instead, at least for the time being, but it's a limit of how wide you can go until you hit diminish returns as well. At least until the software catches up and can make much better use of multiple threads.
Posted on Reply
#105
The Lighthouse
I have a dreaded feeling that AMD may have been just abandoned Threadripper, not just merely delaying it into 2020.

If any chance X570 have quad channels then it is safe to assume that AMD is no longer interested in Threadripper, instead focusing on Eypc and Ryzen.
Posted on Reply
#106
M2B
bug said:

Even if this is totally legit, feeding 16c/32t needs at least a quad-channel RAM setup. It's not going to make its way into the mainstream.
Me, I'm just curious what I can get in the $200-300 range to make me ditch my current CPU.
It's the matter of workload and also memory support will be improved.
Assuming that the X570 boards may be supporting up to 4000MHz of speed, that's still plenty of bandwidth for plenty of workloads.
And even if you're memory bandwidth bound in a specific application, it's not like it wouldn't perform any better than an 8 core, for example.
Posted on Reply
#107
efikkan
Darmok N Jalad said:

I still don’t see why 5.0+ isn’t possible. Zen 1 was a first try at a new architecture, and Zen 1.5 wasn’t a massive departure as much as it was some tweaks and a new node. Zen 2 is AMD’s first real refinement, and it’s on 7nm to boot. It sure seems like current Zen offerings have an architectural limit preventing higher clocks. Just a wild guess, but that may be one of the things that the chiplet design solves, as the IO, IMC, and IF components may be what was so sensitive to faster clocks. Won’t be long now to find out. I hope AMD discusses it when Zen 2 arrives.
I'm not saying 5 GHz is impossible, but as of now we have no evidence supporting it will happen, and it will be challenging to achieve with 7nm DUV. Surprisingly most people seem to think Zen 2 will hit 5 GHz with massive core counts, and it's all based on speculation falsely portrayed as "leaks".

Generally speaking, higher clocks require higher voltage, and rises rapidly once you get outside of the "sweetspot" of the node, and every 100 MHz above that gets exponentially harder. For TSMC 7nm DUV to easily hit 5 GHz, it would need to be better than expected.

Also, chips on this node will effectively be ~2x denser, which means that unless the power usage is more than cut in half, the heat per mm² will actually increase. CPUs like i9-9900K which hits 5 GHz is about 174mm²(including IGP), and it certainly runs into issues due to heat.

So just saying; manage your expectations.

M2B said:

Assuming that the X570 boards may be supporting up to 4000MHz of speed, that's still plenty of bandwidth for plenty of workloads.
And even if you're memory bandwidth bound in a specific application, it's not like it wouldn't perform any better than an 8 core, for example.
Last time I checked DDR4 supported up to 3200 MHz JEDEC speeds.
Posted on Reply
#108
TheLostSwede
The Lighthouse said:

I have a dreaded feeling that AMD may have been just abandoned Threadripper, not just merely delaying it into 2020.

If any chance X570 have quad channels then it is safe to assume that AMD is no longer interested in Threadripper, instead focusing on Eypc and Ryzen.
It's expected later this year.

X570 will be dual channel. Well, technically the X570 doesn't have memory channels, as they're in the CPU/SoC, but the boards will only support dual channel configurations.

efikkan said:

I'm not saying 5 GHz is impossible, but as of now we have no evidence supporting it will happen, and it will be challenging to achieve with 7nm DUV. Surprisingly most people seem to think Zen 2 will hit 5 GHz with massive core counts, and it's all based on speculation falsely portrayed as "leaks".

Generally speaking, higher clocks require higher voltage, and rises rapidly once you get outside of the "sweetspot" of the node, and every 100 MHz above that gets exponentially harder. For TSMC 7nm DUV to easily hit 5 GHz, it would need to be better than expected.

Also, chips on this node will effectively be ~2x denser, which means that unless the power usage is more than cut in half, the heat per mm² will actually increase. CPUs like i9-9900K which hits 5 GHz is about 174mm²(including IGP), and it certainly runs into issues due to heat.

So just saying; manage your expectations.
You are aware that some people actually work in the industry and gets this information ahead of product launches, right?
Just because you feel that it's speculation, doesn't mean that is the case.
If anything your post is a bunch of speculation about something you have zero insight into.
Posted on Reply
#109
Aerpoweron
I think we will see soon enough what AMD will put out with the Ryzen 3000 series. It will be interesting to see how well the chiplet design works considering memory latency.

In my opinion it is more and more relevant what you need the processor for. Examine your use-case and then buy the processor which fits it best.

But a disadvantage of a smaller production process is the heat generated in a smaller area. So heat transfer out of the hot zones will be quite challenging. There is even some research in boron-nitrite, which should have twice the heat transfer capability as copper. But even if it works as promised, it will take years until it can be useful in chip production.

I have to agree, that the core clocks get down the more cores you have (active) and it will be more challenging for programmers to utilize them properly. Nowerdays, at least in Germany, not many programmers can do multicore programming :( And then you have the scheduler from the OS which might trow threads around like crazy. So slowing it down the thread. It might be a good idea from a heat perspective. Can you imagine one core out of 16 running at 6GHz and drawing all the 135W. Thermal stresses might ruing the CPU quickly.
Posted on Reply
#111
bug
M2B said:

It's the matter of workload and also memory support will be improved.
Assuming that the X570 boards may be supporting up to 4000MHz of speed, that's still plenty of bandwidth for plenty of workloads.
And even if you're memory bandwidth bound in a specific application, it's not like it wouldn't perform any better than an 8 core, for example.
True, "shove many cores in there so no one will notice some are just waiting to be fed" is a viable strategy, too.
Posted on Reply
#113
efikkan
TheLostSwede said:

You are aware that some people actually work in the industry and gets this information ahead of product launches, right?
Just because you feel that it's speculation, doesn't mean that is the case.
If anything your post is a bunch of speculation about something you have zero insight into.
Please be serious.
There is nothing wrong in speculating when labeled as such, but what's wrong is portraying speculation as leaked facts, which many opinionators on YouTube, sites like Wccftech and Videocardz etc. does.

Final clocks and prices are not set until the final stepping has gone through qualification, which usually happens weeks before launch. Anyone who claims to know these details >6 months ahead is lying, regardless of who they claim to be. You can speculate all you want, but you can never know a fact before it exists.

TheLostSwede said:

So what's this then, fake news? https://www.corsair.com/us/en/Categories/Products/Memory/vengeance-lpx-black/p/CMK16GX4M2K4700C19
Ever heard of XMP? That's the overclocked speed. I'm sorry, but I thought this was common knowledge.
Posted on Reply
#114
TheLostSwede
Aerpoweron said:

Check out the SPD speed of the memory, which is DDR4 2133

https://www.corsair.com/us/en/Categories/Products/Memory/vengeance-lpx-black/p/CMK16GX4M2K4700C19#tab-tech-specs

The 4700MHz refers to XMP profiles, which is not a JEDEC specified speed.
And this is relevant how? Does this mean that the memory can't run faster than 2133MHz?

efikkan said:

Please be serious.
There is nothing wrong in speculating when labeled as such, but what's wrong is portraying speculation as leaked facts, which many opinionators on YouTube, sites like Wccftech and Videocardz etc. does.

Final clocks and prices are not set until the final stepping has gone through qualification, which usually happens weeks before launch. Anyone who claims to know these details >6 months ahead is lying, regardless of who they claim to be. You can speculate all you want, but you can never know a fact before it exists.


Ever heard of XMP? That's the overclocked speed. I'm sorry, but I thought this was common knowledge.
I am being serious. I also know things you don't know, but clearly everyone else is wrong, but you.
You're so stubborn you can't even listen to other people or try to read between the lines.
Let me spell it out for you, YOU ARE WRONG.
Posted on Reply
#115
bug
TheLostSwede said:

I am being serious. I also know things you don't know, but clearly everyone else is wrong, but you.
You're so stubborn you can't even listen to other people or try to read between the lines.
Let me spell it out for you, YOU ARE WRONG.
Look, all he's saying is while you may have additional information, the rest of us don't. We only know that some guy posted something, somewhere. He's been right before and so has the horoscope. What should we make of this?
Posted on Reply
#116
TheLostSwede
bug said:

Look, all he's saying is while you may have additional information, the rest of us don't. We only know that some guy posted something, somewhere.
Exactly. But who then gives you the right to go and call that person a liar and a cheat? Feel free to take it with a boatload of salt, say that you don't believe it's real, but you can't call someone a liar if you don't have the facts to prove them wrong. This is in part with what's wrong with the internet.
I'm sorry I can't share the information I have, but I've been trying to tell all the naysayers in this thread that he's not way off, in fact the details he released in December are very close to what will launch. I don't understand why that is so impossible to comprehend.
Posted on Reply
#117
storm-chaser
I'm sure AMD has a couple tricks up its sleeve.

I for one would be excited to see another 5.0Ghz piece from AMD.

They've done it before, so it's not like this isn't without precedent.
Posted on Reply
#118
Aerpoweron
TheLostSwede said:

And this is relevant how? Does this mean that the memory can't run faster than 2133MHz?
The 2133 MHz is the JEDEC specification it will run guaranteed with this speed if the CPU officially supports it.
The XMP settings are overclocking settings. It might likely work, but it is not guaranteed.

I have a i7 4790k which can handle one DDR3 2400MHz memory, but not another one. Strange crashes, even corrupt data appears. If you want to use such high speed memory in production environment, you have to to excessive testing.

What i find a little strange about the Corsair memory, they don't support the JEDEC 2666MHz specification, which i would assume should be ok at 4700MHz overclocked.

For JEDEC specified Memory is the highest 2666MHz DDR4 what i have seen for consumer computers. For servers there is already 2933MHz JEDEC specified available.
Posted on Reply
#119
bug
TheLostSwede said:

Exactly. But who then gives you the right to go and call that person a liar and a cheat?
Was the internet invented so we can call people names?
TheLostSwede said:
I'm sorry I can't share the information I have, but I've been trying to tell all the naysayers in this thread that he's not way off, in fact the details he released in December are very close to what will launch. I don't understand why that is so impossible to comprehend.
You have to realize that, to us, you're still just a guy posting on the internet. Right as you may be, we can't verify your claims anymore than we can verify the supposed leak. I, for one, believe you. But that's my personal choice, don't be surprised when others don't.

Aerpoweron said:

The 2133 MHz is the JEDEC specification it will run guarantied with this speed if the CPU officially supports it.
The XMP settings are overclocking settings. It might likely work, but it is not guarantied.

I have a i7 4790k which can handle one DDR3 2400MHz memory, but not another one. Strange crashes, even corrupt data appears. If you want to use such high speed memory in production environment, you have to to excessive testing.

What i find a little strange about the Corsair memory, they don't support the JEDEC 2666MHz specification, which i would assume should be ok at 4700MHz overclocked.

For JEDEC specified Memory is the highest 2666MHz DDR4 what i have seen for consumer computers. For servers there is already 2933MHz JEDEC specified available.
Oh, XMP will work, that much is guaranteed. You just need to pair the sticks with a mobo that can also run them at those speeds.
Posted on Reply
#120
TheLostSwede
Aerpoweron said:

The 2133 MHz is the JEDEC specification it will run guarantied with this speed if the CPU officially supports it.
The XMP settings are overclocking settings. It might likely work, but it is not guarantied.

I have a i7 4790k which can handle one DDR3 2400MHz memory, but not another one. Strange crashes, even corrupt data appears. If you want to use such high speed memory in production environment, you have to to excessive testing.

What i find a little strange about the Corsair memory, they don't support the JEDEC 2666MHz specification, which i would assume should be ok at 4700MHz overclocked.

For JEDEC specified Memory is the highest 2666MHz DDR4 what i have seen for consumer computers. For servers there is already 2933MHz JEDEC specified available.
Again, what does have to do with anything being discussed here?
We all know AMD has had memory controller issues and I don't know of they've resolved them or not.
What I do know is that we can expect support for faster memory and it's outside of JEDEC spec.

That was just a random stick of RAM I pulled to prove my point, that memory speed isn't limited to JEDEC spec and hasn't been for years.
Shit I remember testing the first PC133 memory on a VIA board and it didn't work for shit. It wasn't JEDEC spec at the time.
Posted on Reply
#121
M2B
efikkan said:

Ever heard of XMP? That's the overclocked speed. I'm sorry, but I thought this was common knowledge.
Who the hell cares about the JEDEC standards?
Overclocking the memory is as easy as drinking water.
Posted on Reply
#122
TheLostSwede
bug said:

Was the internet invented so we can call people names?

You have to realize that, to us, you're still just a guy posting on the internet. Right as you may be, we can't verify your claims anymore than we can verify the supposed leak. I, for one, believe you. But that's my personal choice, don't be surprised when others don't.


Oh, XMP will work, that much is guaranteed. You just need to pair the sticks with a mobo that can also run them at those speeds.
Some days it feels like it.

Oh, I have, trust me. Look at when I joined here though. I've known the owner of this site since before it started. In fact, I told him I thought the name was a bit pants...

Isn't that kind of the point of leaks, they can't really be verified. Again, not surprised, but I don't get the continuous counter arguments when someone is trying to let you know that they know something you don't and maybe you should just sit this one out.
I'm trying my best to share some details here, without getting people in trouble. At least I know what I will be spending my money on later this year.
Posted on Reply
#123
yakk
Dizzying back&forth here... o_O

Just a guess, but if even the next Ryzen generation will be a chipset design, then brand differentiation with Threadripper becomes just pci lanes & memory channels. Not the CPU design itself.

I can see AMD repositioning their successful "Threadripper" branding differently for maximum impact, which might cause it to disappear, at least for a while until needed again.
Posted on Reply
#124
Aerpoweron
TheLostSwede said:

Again, what does have to do with anything being discussed here?
We all know AMD has had memory controller issues and I don't know of they've resolved them or not.
What I do know is that we can expect support for faster memory and it's outside of JEDEC spec.

That was just a random stick of RAM I pulled to prove my point, that memory speed isn't limited to JEDEC spec and hasn't been for years.
Shit I remember testing the first PC133 memory on a VIA board and it didn't work for shit. It wasn't JEDEC spec at the time.
My apologies, misunderstanding on my side.

Back to the discussion then. Are there official memory supported speeds for the Ryzen 3000 series? As far as i have seen AMD was always a little ahead of Intel concerning officially supported memory speeds. My guess would be DDR4 3200 support since the Ryzen 2000s have 2933 support.
Posted on Reply
#125
efikkan
TheLostSwede said:

Exactly. But who then gives you the right to go and call that person a liar and a cheat? Feel free to take it with a boatload of salt, say that you don't believe it's real, but you can't call someone a liar if you don't have the facts to prove them wrong. This is in part with what's wrong with the internet.

I'm sorry I can't share the information I have, but I've been trying to tell all the naysayers in this thread that he's not way off, in fact the details he released in December are very close to what will launch. I don't understand why that is so impossible to comprehend.
I'm going to give you a last chance to read my posts and apologize for your adolescent behavior. Even if I were wrong, there is no excuse for behaving like this, and if I were, any grown up should be able to make a serious argument for their case.

We're not going to descend into a philosophical discussion about truth really is here, but any reasonable person understands the distinction between speculation and fact. Just like no-one can know the outcome of a soccer match before it's played, but do a qualified guess, and even be right, it still doesn't change the fact that the result was not known in advance.

Please keep this discussion on topic from now on.

Until the final specs are actually set, AMD, Nvidia and Intel operate with targets, and these are ranges, not specific clock speeds and exact prices, since this depends on the final yields. This is set after the final stepping has gone through qualification. There are many leaks that can contain true information, like bigger architectural changes etc., but no-one can know the final clocks/price etc. until they are set, and you can use this knowledge to discredit many false "leaks". False leaks will accelerate disappointment when they eventually come crushing down.

bug said:

Oh, XMP will work, that much is guaranteed. You just need to pair the sticks with a mobo that can also run them at those speeds.
The sticks are guaranteed, but not the motherboard or the memory controller. Many Skylake/Kaby Lake CPUs struggle to remain over 3000 MHz over time.
Posted on Reply
Add your own comment