Question Speculation: RDNA2 + CDNA Architectures thread

[uzzi38]

All die sizes are within 5mm^2. The poster here has been right on some things in the past afaik, and to his credit was the first to saying 505mm^2 for Navi21, which other people have backed up. Even still though, take the following with a pich of salt.

Navi21 - 505mm^2

Navi22 - 340mm^2

Navi23 - 240mm^2

Source is the following post: https://www.ptt.cc/bbs/PC_Shopping/M.1588075782.A.C1E.html

 

[Hitman928]

So why does the 3070 have a 256-bit bus and how many of Big Navi CPU's will have 256-bit instead of 384-bit.

We'll find out soon on your second question, but with regards to your first, the speed of the memory used is much higher than it was back then so you don't need as wide of a bus to get the same (or more) bandwidth. The 3070 probably could have used a wider bus or even faster memory, but that would have increased the cost of the card which Nvidia obviously didn't want to do.

 

[Stuka87]

For what it's worth. My AMD 7950 has a 384-bit bus. That is a circa 2012 card and we are in 2020.

Not comparable. Bus width is just one the memory characteristics to take into account. Also, Tahiti didn't really have any memory compression. Kepler at the same time had a 256bit bus with the 670/680.

Memory technology has changed a lot since then. GDDR5 vs GDDR6, memory compression, color compression, etc, all have an impact on GPU performance.

 

[JasonLD]

Exciting day tomorrow. I can finally decide if I want to hunt for 3080/90 or 6900xt(whatever they end up calling it).

 

[GodisanAtheist]

Some people chase Total Solar Eclipses, I chase GPU launches.

Oddly enough eclipsers get a chance at several eclipses a year, while I only get my fix every 2-3 years.

My wife suspects I've been hitting the rocks the way I'm tweaked out I'm anticipation...

 

[Glo.]

Considering that tomorrow there wont be any sub-300$ GPU launches, Im only interested in RDNA2 architecture deep dives.

 

[swilli89]

Considering that tomorrow there wont be any sub-300$ GPU launches, Im only interested in RDNA2 architecture deep dives.

You never know! Are we sure Navi 23 or Navi 22 aren't announcing tomorrow? I'm really, really hoping AMD does something special in the $300-400 range.

 

[Zoal]

You never know! Are we sure Navi 23 or Navi 22 aren't announcing tomorrow? I'm really, really hoping AMD does something special in the $300-400 range.

Wouldn't it be funny if the lower end options went the XSX approach and didn't employ the cache? An affordable yet competitive option

 

[linkgoron]

Exciting day tomorrow. I can finally decide if I want to hunt for 3080/90 or 6900xt(whatever they end up calling it).

If whatever AMD launches is competitive with the Nvidia equivalent(s), I'm expecting tons of DLSS and RTX shilling (I'm expecting AMD's RT solution to be slower than Nvidia's, especially at the beginning), and probably also down-playing more memory (and perf/watt if applicable).

 

[Veradun]

Considering that tomorrow there wont be any sub-300$ GPU launches, Im only interested in RDNA2 architecture deep dives.

Me too. Sadly I don't think tomorrow will bring such a boon

 

[mohit9206]

For those who want sub $300 gpu, you can buy a Series S

 

[Veradun]

For those who want sub $300 gpu, you can buy a Series S

I'm going XSX to also get a Zen2 since I'm still on a Zen+ :>

 

[kurosaki]

You never know! Are we sure Navi 23 or Navi 22 aren't announcing tomorrow? I'm really, really hoping AMD does something special in the $300-400 range.

Double 5700-performance. Is what I'm expecting economically for an upgrade. Realistically, 20% over 5600xt. It's like, nothing is happening in the perf per dollar sector anymore. You could just always just invest in top of the line för 1500usd and sell it for 1250 before next gen, to buy a new top of the line for 1500. In that sense, you only play 250 USD for top of the line perf. That and the hassle of having to sell on eBay or what not.

 

[Tup3x]

For those who want sub $300 gpu, you can buy a Series S

And kiss your PC game collection good bye....

 

[Veradun]

And kiss your PC game collection good bye....

Buying a console sets your pc to autodestruction?

 

[DeathReborn]

For those who want sub $300 gpu, you can buy a Series S

You mean the monstrosity that is the Microsoft iSpeaker?

 

[Tup3x]

Buying a console sets your pc to autodestruction?

Well, no, but buying a console doesn't do anything. Your current games would still be the same dia show and you wouldn't have any games on that new console.

 

[Veradun]

Well, no, but buying a console doesn't do anything. Your current games would still be the same dia show and you wouldn't have any games on that new console.

I don't have any game running as dia show atm @ 1440p on a 5700, i can buy new games for pc or xbx depending on the kind of game :>

 

[hkultala2]

They could also use on-die eDRAM (what Ibm uses for L3 cache on it's CPUs). Or it could be a separate die alltogether. Interesting times!

No, they could not.

eDRAM scales badly with new mfg processes and is not supported by TMSCs "7nm" process at all.

Those IBMs chips wil eDRAM cache are made with old mfg processes.

SRAM scales very well with new processes and is only about 1mm2 per 1 MiB on TMSCs "7nm"

 

[moinmoin]

Considering that tomorrow there wont be any sub-300$ GPU launches, Im only interested in RDNA2 architecture deep dives.

Maybe Lisa can do a small parody, pull out a GPU card from an oven and talk about anything but the card itself, before starting the proper event.
Edit: Taking the new RDNA2 out from a fridge may possibly be more apt though.

 

[hkultala2]

Is it possible to do unequal transistor densities across a die (maybe it's actually normal)? As in, can AMD throw in a huge cache but pattern that portion of the die to be extremely dense (60Mtr) while the remainder of the die that actually houses logic is less dense (40Mtr)?

Or does that already happen and the 40Mtr density that's thrown around an average of the entire die with large variations in density between different portions of the chip?

Its not just "possible", its how EVERYTHING has been done for the last 25 years.

SRAM(cache) is about 4 times more dense than logic in average. And there is also big difference between logic made to be fast and logic made to be dense. Comparing transistor counts do not make much sense.

Transistor Count: A Flawed Metric

Transistor count and transistor density are often portrayed as technical achievements and milestones. Many vendors brag about the complexity of their design, as measured by transistor count. In reality, transistor count and density varies considerably based on the type of chip and especially the...

www.realworldtech.com

 

[Tuna-Fish]

Is it possible to do unequal transistor densities across a die (maybe it's actually normal)? As in, can AMD throw in a huge cache but pattern that portion of the die to be extremely dense (60Mtr) while the remainder of the die that actually houses logic is less dense (40Mtr)?

Or does that already happen and the 40Mtr density that's thrown around an average of the entire die with large variations in density between different portions of the chip?

Yes, this is absolutely normal. Chips are not designed out of individual transistors anymore, they are designed out of small separately optimized "macros" that provide some fixed functionality and contain anywhere from a few to low thousands of transistors, and the macros for SRAM are generally way more dense than the macros for logic, so a chip that has more SRAM is generally denser.

They could also use on-die eDRAM (what Ibm uses for L3 cache on it's CPUs). Or it could be a separate die alltogether. Interesting times!

There is no on-die eDRAM available for any TSMC 7nm process.

For what it's worth. My AMD 7950 has a 384-bit bus. That is a circa 2012 card and we are in 2020.

Note that wider busses get harder with every new GDDR generation, as signal quality requirements get stricter.

 

[Glo.]

I'm really, really hoping AMD does something special in the $300-400 range.

Yeah, special offers for RDNA1 based GPUs!

 

[Qwertilot]

They apparently still (as of early October anyway) haven't actually stopped making the 5700/XT yet, so it would be odd if there was anything immediate.

Patience in order The GPU division might be rather better funded now but both Zen3 & the consoles are genuinely absolutely massive launches.

 

[Ajay]

Yay! https://www.anandtech.com/show/1620...eon-live-blog-starts-at-noon-eastern-1600-utc

And, crumbs, I'll be out of the house for a couple hours then

 

[zinfamous]

Yeah, special offers for RDNA1 based GPUs!

For now, likely.

....I mean, everyone at least assumes we'll get a full stack announcement (card names, some limited performance numbers and architecture details), right, but with release date (and earliest) for the larger cards? I'm guessing only the 6800 and "6900" cards will have release information today ("Novecember"), with the rest of the stack = "Q1 2021."