Wednesday, June 19th 2019

AMD E3 2019 Tech Day: All Slide Decks (Ryzen 3000 "Zen 2", Radeon RX 5000 "Navi", etc)

This continuously-updating article is a compilation of the 11 (eleven) slide-decks AMD pushed out that cover its upcoming Ryzen 3000 series processor- and Radeon RX 5000-series graphics card launches on July 7, at varying levels of technical insight to target different classes of audiences (consumers, editors, and industry partners such as OEMs and system-integrators). In the first segment, we post the Ryzen Deep Dive presentation targeting DIY consumers and SI (system integrators). Keep checking back on this article, you may find interesting details which we couldn't spot ourselves.
The slide-deck continues.


Deck 2 - AMD Matisse MCM Physical Design
This slide-deck is targeted at enthusiasts, engineers, and editors to help understand what's under the IHS of a Ryzen 3000 "Matisse" MCM.

Deck 3 - Zen 2 CPU Architecture
Also targeted at enthusiasts and editors, this deck describes the improvements and resulting performance gains of the Zen 2 microarchitecture.

Deck 4 - Zen 2 CPU Architecture (lite edition)
This slide deck is a slightly dumbed-down version of the one above, targeted at retail channel distributors, gamers/consumers, and system integrators.

Deck 5 - 3rd Gen Ryzen Main Slide Deck
This is the main slide deck for the 3rd gen Ryzen series, detailing its key features and SKUs.

Deck 6 - Graphics Architecture Update

Deck 7 - AMD Radeon RX 5000 Series

Deck 8 - AMD Radeon RX 5700 series

Deck 9 - AMD Graphics Architecture Update Deep Dive

Deck 10 - Radeon Software Adrenalin 2019 powering RX 5700 series

Deck 11 - Game Developers

Slides from the E3 2019 "Next Horizon" Event Keynote Not Found in the Above Decks
Add your own comment

48 Comments on AMD E3 2019 Tech Day: All Slide Decks (Ryzen 3000 "Zen 2", Radeon RX 5000 "Navi", etc)

#2
londiste
What is it with constant notes that OC, memory OC, PBO etc are not covered by warranty?
Posted on Reply
#3
kapone32
londiste, post: 4067124, member: 169790"
What is it with constant notes that OC, memory OC, PBO etc are not covered by warranty?
That has been official since CPUs became unlocked. It basically covers them from people breaking their parts by OC without understanding what they are really doing.
Posted on Reply
#4
cucker tarlson
Wavetrex, post: 4067119, member: 182738"
Did they ... just use: "It just works" ? :laugh:

https://www.techpowerup.com/img/dF4sjxFh6HNk7GXn.jpg

I smell a trend here...
didn't see that before but those latency numbers don't look great.70ns on 3200 c14.That's what 2nd gen can do on 3200 c16
https://www.techpowerup.com/forums/threads/share-your-aida-64-cache-and-memory-benchmark-here.186338/page-34#post-4064663

3733 limit is pretty meh either these days.
Posted on Reply
#5
Wavetrex
cucker tarlson, post: 4067137, member: 173472"
those latency numbers don't look great.70ns on 3200 c14.That's what 2nd gen can do on 3200 c16
That is the unfortunate effect of using a separate memory controller and I/O die.

The cores and their L2/L3 cache are now not directly connected to the IMC, but go through an IF link first, then the internal routing of the "North Bridge" die.
Latency will increase, there's no way around it...

The reason cache size was increased is to compensate for this design quirk.
I'm guessing it works fine... people game on their Threadrippers, and they have the same "issue" of having to jump on IF link to get to memory.
Posted on Reply
#6
cucker tarlson
Wavetrex, post: 4067141, member: 182738"
That is the unfortunate effect of using a separate memory controller and I/O die.

The cores and their L2/L3 cache are now not directly connected to the IMC, but go through an IF link first, then the internal routing of the "North Bridge" die.
Latency will increase, there's no way around it...

The reason cache size was increased is to compensate for this design quirk.
I'm guessing it works fine... people game on their Threadrippers, and they have the same "issue" of having to jump on IF link to get to memory.
well you can't have it all,but this is what we've got so far
-same core count for skus, r3 is 4c/8t,r5 is 6c/12t,r7 is 8c/16t.seems like the main point of this design was to introduce 12c for mainstream boards
-boost frequencies around 300mhz higher than 12nm
-same or worse latency
-higher cache size
-memory limitations pushed a bit higher
-higher price both for skus and for the boards
Posted on Reply
#7
fynxer
If you Goolge the barcode on the Ryzen boxes you will get a suprise
Posted on Reply
#8
Wavetrex
The only reason for chiplets is production cost.

Half the chip(let) size may 5 times lower production cost. Weird huh ? It's because the unavoidable defects in the silicon... and the smaller you get, the higher chance that you end up with a perfectly working chip.
So AMD can produce these cheaper than with a monolithic design and have more profit, which they seriously need...
The fact that it has some performance benefit is just a +

Oh, and they can shove more cores in due to that significant production cost reduction...

I'm guessing it's literally cheaper for them to make 1x 16-core 3950X than it was to make 1x 8-core 1800X, but they are selling it with 50% higher launch price.
And still, for the consumers, the price per core is lower. Magic !
Posted on Reply
#9
Xuper
hmm compare to Intel/Nvidia sliders , AMD has better virtual quality.That font on slider is cool.what's font ?
Posted on Reply
#10
juiseman
fynxer, post: 4067147, member: 103789"
oolge the barcode on the Ryzen boxes you will get a suprise
???
Posted on Reply
#11
Fatalfury
AMD was always good at making Slides... xD

btw .. thats a lottttttttt of SLIDES!!!!!!!!!!!
Posted on Reply
#13
junglist724
Wavetrex, post: 4067153, member: 182738"
The only reason for chiplets is production cost.
It also makes the CPUs really modular. They can just add a Navi chiplet for the APU, they can make an IO die with video outputs for APUs and LPDDR4X support for mobile, or they can eliminate stuff like the onboard ethernet controller that was on every Zen 1 die on platforms that don't need it.
Posted on Reply
#14
Midland Dog
intel igps support freesync so thats 1 slide gone, how many others have bs or extended truths
Posted on Reply
#15
Easo
londiste, post: 4067124, member: 169790"
What is it with constant notes that OC, memory OC, PBO etc are not covered by warranty?
Does not really matter ir EU at least.
Posted on Reply
#16
dicktracy
[IMG width="663px"]https://www.techpowerup.com/img/dERNb4Bo4cap1Nvw.jpg[/IMG]

yeah using a 2015 CPU with a 2016 GPU versus the 3700x and Navi is downright deceiving. Who are they kidding... Zen 2 is virtually Skylake 2.0 LMAO pathetic marketeers. Pair that outdated 6700k with 5700XT and it'll deliver the same FPS.
Posted on Reply
#18
cucker tarlson
TheGuruStud, post: 4067249, member: 42692"
You can bet the subtimings are disastrous.
you mean on the presentation numbers?
anyway,I think doubling the cache and the freqency increase will more than make up for that.

dicktracy, post: 4067242, member: 173119"


yeah using a 2015 CPU with a 2016 GPU versus the 3700x and Navi is downright deceiving. Who are they kidding... Zen 2 is virtually Skylake 2.0 LMAO pathetic marketeers. Pair that outdated 6700k with 5700XT and it'll deliver the same FPS.
lol,this is rich.
3700x is faster than 6700k and 1070 when using a 2070 equivalent,that is surprising.
also,the avg. shows 24% over gtx 1070 for 5700xt while tpu has 2070 38% faster than 1070.40% is best case scenario here,more often you see 1070Ti-2060 performance.
Posted on Reply
#19
Mephis
dicktracy, post: 4067242, member: 173119"
[IMG width="663px"]https://www.techpowerup.com/img/dERNb4Bo4cap1Nvw.jpg[/IMG]

yeah using a 2015 CPU with a 2016 GPU versus the 3700x and Navi is downright deceiving. Who are they kidding... Zen 2 is virtually Skylake 2.0 LMAO pathetic marketeers. Pair that outdated 6700k with 5700XT and it'll deliver the same FPS.
Yeah, that one was kind of weird. The rest of their comparisons are legitimate.
Posted on Reply
#20
Xuper
Mephis, post: 4067257, member: 186806"
Yeah, that one was kind of weird. The rest of their comparisons are legitimate.
Did you see "It's time to upgrade" ? It's not weird. I wonder did you expect 3700X + RTX 2070? there are some people who sold their 6700K and bought 2700X.
Posted on Reply
#21
delshay
Wavetrex, post: 4067141, member: 182738"
That is the unfortunate effect of using a separate memory controller and I/O die.

The cores and their L2/L3 cache are now not directly connected to the IMC, but go through an IF link first, then the internal routing of the "North Bridge" die.
Latency will increase, there's no way around it...

The reason cache size was increased is to compensate for this design quirk.
I'm guessing it works fine... people game on their Threadrippers, and they have the same "issue" of having to jump on IF link to get to memory.
There is, when they start stacking. This will reduce latency as it will be much closer together.
Posted on Reply
#22
moproblems99
Xuper, post: 4067266, member: 83814"
Did you see "It's time to upgrade" ? It's not weird. I wonder did you expect 3700X + RTX 2070? there are some people who sold their 6700K and bought 2700X.
Don't people upgrade every year? Who actually has a 6700k?
Posted on Reply
#23
ixi
moproblems99, post: 4067327, member: 155919"
Don't people upgrade every year? Who actually has a 6700k?
I'm using still 6700k. And what do you want to say with that?
Posted on Reply
#24
olymind1
moproblems99, post: 4067327, member: 155919"
Don't people upgrade every year? Who actually has a 6700k?
No, i'm still on i5-3550s, i will buy a 3600 and for a few years that would be enough for me. :)
Posted on Reply
#25
Vayra86
A Missy Elliott tune comes to mind

Posted on Reply
Add your own comment