AMD RDNA 3 Is Chiplet Based | RDNA 4 Confirmed | RTX 30 Specs Show Up to 24GB RAM!

Article for this video -

RDNA 3 is almost certainly chiplet based at this point, with several leaks hinting that this is the direction AMD will go with future GPUs. We also have 'confirmation' that AMD has started Navi 41 (RDNA 4) design too. This shows AMD's Graphics Cards will be a lot more consistent in releases than before, and most likely the company will be a lot more aggressive attempting to tackle Nvidia.

Speaking of Nvidia, the companies RTX Ampere specs have partly leaked, and the RTX 3080 Ti is apparently going to feature up to 24GB of RAM.

Times

0:00 - Intro
14:05 - RTX 30
17:00 - The Medium Xbox Ray Tracing

SOURCES

AMAZON AFFILIATE LINKS

GreenManGaming Affiliate link -

SOCIAL MEDIA

[Sephyroth]

I don't know if it counts as an intro if it occupies 73% of the video but hey, who am I to judge.

[bhill]

RDNA3 HBM2e chiplet, etc. Monster apu on the way as well.

[samlebon]

So basically we're going to have an rDNA mutation every year. good work AMD.

[HDLTHEDR]

Hi Paul. Yea, Developers cleared up any confusion and false rumors of The Medium not having RT feature. Glad they did.

Let's hope MS and SONY finally announce pricing, dates and preorders this month. Tired of waiting and speculations!

[lagathy]

24gbs of memories....???....christ, don't even think i've had that many..

[gsydaz]

24Gb of VRAM, I certainly don't need that much for 2k UW gaming.
But maybe they are giving some future thought to playable framerates at 8k resolutions.

[HablaConOwens]

Id like to see zen 3 with navi cores. I think we are getting close to APUs replacing mid-to-low range cards (2060, 5500 xt).
Maybe another year or 2.

[ericjohns]

The question is. How good is AMD's first gen ray tracing compared to nVidia's second gen ray tracing.

[Void-ujjd]

Wonder what the drivers will be like. 5700XT took some months to recover.

[matsv]

The gpu chiplets have bee rumors in engineering circles if over half a decade now. Early grapics processors pretty much worked that way. But when AA was introduced scaling got a lit worse.

This can peobobly be solved if post processing is done in a third chip. This is not that processing intensive, but very very memory intensive. This would in that case be done in a die with a bit of buffety memory.

This is not a new idea. This is pretty much how xbox 360 works.

Having a center die with buffer memory, like xbox esram, interface chip as well as the post processing wos solve that problem and simultaneously make it very easy for chiplets.

A high end chiplet based gpu would need 5 die, rather than 3 like the ryzen.

The issue is that the die would need to be a maximum of just over 100mm2... they could probobly push it to 150mm2, but not really much larger.

Navi2 is 505mm2.. presumably 80CU. Say 400mm2 is used for CU, that would give a 100mm die only 20CU. But with 5nm and say 120-130mm2, it could have a theoretical 40CU, with a theoretical 4 chiplets, making a total of 160CU.

There is one more benefit. If all the screen space information and maybe even of screen reflection is stored in the cache that would drastically decrease the amount of memory access needed, cutting down the bandwidth constraints, pretty much making HBM not even needed and maybe even just needing 256bit GDDR6 and for something like a apu running just fine of DDr5.

[VoldoronGaming]

For a chiplet based gpu the stream/cuda cores have to get either much much smaller or incredibly more efficient than they are now. An 80 cu die is roughly 500mm/squared. At 3nm node that 80 cu die is still 250 mm/squared roughly. So i do not think we will see chiplet designs until navi4 which is where you could get a 80 cu x 2(160 cus)designs. With that being said a chiplet design will give you more flexibility in cu combinations: a 40 cu + 40 cu or 60 cu + 40 cu.. or 40 cu with rt cores or 80 cu with rt cores....chiplet designs open up all kinds of configurations...
Now onto ampere....the amount of vram tells you how expensive ampere is going to be....1200 or more for 3080ti since navi will fall between 3080ti and 3080.

[faustianblur]

For RDNA based chiplets it might make sense to have each Shader Engine be on a separate die, a separate die for memory IO and another for command queues and work distribution.

[walterthomas]

The Navi 21, 23 and the refresh of Navi 10 GPU's will all run at 2020, 1835 and 1725 MHZ.
The mid range Navi 23 is very interesting, one card has 52 CU's and runs at 1835 MHZ and has 12.2 TFLOPS, this is the upcoming RX 6800, this card matches the specs of the Xbx.

The equivalent GPU of the ps5 is the RX 6600 which has 40 CU's and runs at 2020 MHZ and 10.3 TFLOPS, this GPU uses 230 watts compared to the Ps5's 175 watts.
The GPU of the Ps5 compares well with the RX 6700xt which has 48 CU's and runs at 1725 MHZ and 10.59 TFLOPS.

[xthedetail]

Love the ads combined with terrible audio, nice touch!

[jwhews]

He has the deepcool captain AIO. I love that AIO. Best $90 I ever spent.

[mixk]

9600GT user here, glad I skipped the current GPU generation.

[blueeyednick]

Why is Graphe Dracula talking about hardware lol

[denim]

It's time to look at some sound dampening.

[coladict]

"A small update on the Medium"
I see what you did there!

[synthdude]

I like your channel and the way you deliver information. Very good