New Upcoming ATI/AMD GPU's Thread: Leaks, Hopes & Aftermarket GPU's

AMD’s next-gen Navi 2X GPU IDs confirmed for three RDNA 2 dies

Some alleged GPU IDs of the entire next-gen Navi 2x lineup have been reportedly spotted by _rogame and @KOMACHI_ENSAKA on Twitter (via Videocardz).
These entries confirm that we have at least three different GPU variants of the next-gen RDNA2 architecture.

We have 3 RDNA 2 based dies, namely:

• Navi 21: gfx1030
• Navi 22: gfx1031
• Navi 23: gfx1032

The ID GFX10** is often found in the description of CompuBench and drivers, and if you know it, you can infer to some extent which generation the GPU is. GFX101x is the first RDNA.
And GFX103x is currently assigned to the GPU that is estimated to be the 2nd generation of RDNA.

Out of these three spotted entries, only the full-fledged “Navi 21” die will likely form the BIG Navi GPU which will also likely end up in the high-end flagship gaming card (e.g. the RX 6900/XT ?).
Navi 21 also known as ‘Big Navi’ has the GFX ID of 1030. We are more interested in this card.

Navi 21 is estimated to be a high-end GPU of the 2nd generation RDNA2 arch. On the other hand, “Navi 23” is presumed to be a subordinate model of Navi 21, but it is not very clear where this card will fit, but it is presumed to replace Navi 14.
Which means Navi 23 is basically an entry-level product. This Navi chip is now listed in between Navi 22 and Van Gogh as GFX1032.

AMD next-gen Navi 2X RDNA2 GPU IDs



Navi 23 does not seem to have a direct predecessor, because today AMD uses Navi 10 and 14 chips on its cards. It could be a new GPU with RDNA2 (or still RDNA, who knows), but belonging to a more powerful sector than Navi 14 for sure.
Some previous leaks mention that the die is about the same size as the Navi 10, so it doesn’t seem like a low-end chip to me though.

Lastly, the Navi 22 could then replace the Navi 12 for Apple computers. This chip is a custom GPU for the Apple MacBook lineup. “Navi 22” in the middle seems to be the successor to Navi 12 that appeared as Radeon Pro 5600M few days ago.
In other words, it seems that it will be an Apple model again. As was the case with the Vega 12, will the last digit ‘2’ be an Apple-only model once again?

We also have a couple of Navi based APUs. These are “Van Gogh” and “Van Gogh Lite”. These most likely won’t be for the AM4 platform.
It’s unclear what Van Gogh refers to (whether it’s an APU for PC or some custom chip), but it’s presumed to be something that belongs to RDNA2 as long as it’s GFX103x.

THX to Metal Messiah

 

This is the keyboard guy. I really like his keyboard reviews

 

Hype Train will leave the station when 3Dmark Leak finally hit us, e.g. 50k-60k in FS

 

Yea, i agree. Tho, i dont think AMD can charge what nvidia want. Im afraid nvidia has messed up the marked for us consumers considering today's prices..
AMD cant screw up, it dosent have nvidia's "mindshare" and it finally looks like more consumers are informed enough to see what nvidia has been doing since amd more or less dropped out of the gpu race. I hope we consumers vote with our wallet, to show that its not cool to overcharge a product just because they can. Maybe im wrong, and nvidia's 3000 series will go up even more in price that current gen. And people will buy the cards because of good marketing. Soon every tier (low end.mid and high) has gone up 2x or more in price.
Low end card cost about the same as high did not that long ago, mid end chips are sold as "high" end and high end cards cost about the same or more as all the hardware in my pc. I really dont like this development, and its not good for us, the consumers.
Dont take me wrong, what i wrote above goes for AMD to. And i think amd should be a bit careful, and not start overcharging their products like intel and nvidia.

 

Developing all this new tech on the new fab sizes, and the fabs building all new production machinery isn't cheap. Innovation costs money. That cost gets projected to the consumers. Essentially, the cost to produce all these new chips has risen vs older tech, ergo the base cost rose, and thus pricing will reflect this.

 

I doubt it can justify the almost x3 increase of top tier cards though.

 

You want to fix this? Get EVERYONE to not buy anything more than a specific price point. Good luck.

 

This thing presents you 14GB card with 18Gbps GDDR6 memories for 1TB/s total memory bandwidth. (Same can be achieved with widely available 16Gbps GDDR6 and 16GB of VRAM.)
When 5700(XT) was made, both 14Gbps and 16Gbps GDDR6 were widely available and card would benefit to good extent from more bandwidth. Yet AMD went with cheaper 14Gbps GRRD6. Likely because they thought that price bump was not worth performance bump.
I have feeling that 18Gbps memories will come at more than 14% higher price than 16Gbps, therefore giving 2 extra GB VRAM would be cost efficient. (And more appealing to all users due to extra 2 GB of VRAM.)
And with that, there could be "lite" version of card then, with just 8GB of VRAM. (I would consider that since I am on 1080p.)

If slides were not fake. And actual game comparison looks well "faked" considering that some of those games have certain limited performance scaling due to CPU limitations...
I would easily price that card at $800. And I would expect bit bigger DX-R performance loss, but thing is that AMD's design for rays processing is really better than nVidia's. (That's unless nVidia decides to go with a lot more RT cores than they give now... like 4 times.)

 

The Radeon VII's a pretty good example of the memory bottleneck isn't it? Some other improvements but additional bandwidth and faster memory through the extra stacks and also a bit more memory for work applications though nowadays games are also closing in on 10 GB or more so 8 GB GPU models can be a bit limiting although dropping settings can give some nice performance gains and reduced memory or video memory usage with a pretty small image quality drop due to the whole thing around the smaller gains once you get to the highest settings in many games.

Testing of the 5700's not the easiest as error control (ECC) can be what's happening due to sensitivity around overclocking or possibly the memory controller itself but the 5600's after flashing when that works is a bit better in showing some gains and then the editing of memory timings or the straps although there's a risk of longer term degradation or instability for a 2 - 3% gain but that's still a increase.

For my own testing the card shows gains but doesn't scale very well and might need additional voltage or better cooling to also keep the GPU core up around the 2 Ghz range at which point wattage starts spiking whereas a undervolt and mild underclock can give a at most -5% performance for a sizeable reduction in power draw and operating temperature both junction and the general surface or overall temps.


Curious to see how Navi20 handles this, Navi10 already improves on some of the limitations on Vega but it is a more mid-range card in terms of hardware so true scaling with additional compute units and a more high-end product will be very interesting.
~250w - 300w whatever core voltage AMD sets on it and pushing up into the 2.1 Ghz or above clock speeds whatever these extra compute units can take without requiring a ton of power and how it scales once past the current 40 something on Navi10 here.

Plus other changes in the hardware and all that stuff with RDNA2 I suppose or what it will be called.
I don't really mind if it's a 2x 8-pin but it's going to look a bit problematic if Ampere on NVIDIA's side can scale well with less plus AMD getting over that power usage hurdle and voltage defaults on the GPU core and memory for whatever gains they can get would be nice but we'll see once there's actually a product available and some info about it.

Sure the GPU can take it but it also looks a bit silly when you trade 5% performance for more than halving power draw and reducing temps by nearly one third.


So a bit better scaling would be nice but I can't fault AMD for pushing the cards higher by default as long as they're designed to handle it either.
And I do expect they'll want this for the desktop consumer versions of the Navi20 GPU's too.


EDIT:
But perhaps the curve won't be a full upwards spike right at the end for these last small gains this time or perhaps it will if there's any headroom at all for additional overclocking ha ha.
Suppose a big waterblock would work well as usual if the GPU core or components like the VRM's is still primarily heat limited before it throttles first and then power limited after that.



EDIT: Well that's secondary stuff I suppose, first we'll have to see what Navi20 actually is for these high, mid and low range SKU's I suppose from 20 to 23 or how it went again and then what they'll perform like by default and the usual custom designs and coolers both air and water maybe even blower designs unless AMD has something else or if they are leaving it to third parties immediately.
(Suppose not if they're doing another limited type model again themselves. 5700XTX / 5700 50th Anniversary Edition type of thing.)

Then overclocking and undervolting once more.

 

Spoiler: RDNA2 CU (QuadCU) consists of block with 16x TMU 256 SP...

RDNA1: DualCU which has basically a CU that has 2x amount of shaders than CU in GCN. (With 2x "2xWave32 on 2xSIMD32" instead of GCN's 1x "Wave64 on 4xSIMD16".)
RDNA2: QuadCU which has basically a CU that has 4x amount of shaders than CU in GCN.(With 4x "2xWave32 on 2xSIMD32" instead of RDNA1's 2x "2xWave32 on 2xSIMD32")
And that would mean 16TMUs per CU too.

This would mean that RDNA2 GPUs could come in following SP/TMU numbers:
512/32; 1024/64; 1536/96, ...
5120/320 (For 80 "CUs" = 20 QuadCUs)

Why I think this is the case? Because there is penalty to communication from one DualCU to another. And exceeding what can fit into one DualCU/QuadCU should lower performance.
RDNA1 has shared cache within DualCU = same cache up to group size 128.

 

Oh, I'm aware... I was talking about justifying it, not changing it.
I have the feeling the polarisation is setting a firm foot on PC land as well. As long as we have ADHD youtube 'reviewers', I'll never change people's mind.

Common sense is dying out, I tell you

/politics.

 

macOS 11 Big Sur beta – AMD Navi31, MI100, MI200 & Cezanne

-> https://hardwareleaks.com/2020/06/27/macos-11-big-sur-beta-amd-navi31-mi100-mi200-cezanne/

 

SK hynix Starts Mass-Production of High-Speed DRAM, ”HBM2E”

Seoul, July 02, 2020 – SK hynix Inc. (or ‘the Company,’ www.skhynix.com) announced that it has started the full-scale mass-production of high-speed DRAM, ‘HBM2E’,
only ten months after the Company announced the development of the new product in August last year.

SK hynix’s HBM2E supports over 460GB (Gigabyte) per second with 1,024 I/Os (Inputs/Outputs) based on the 3.6Gbps (gigabits-per-second) speed performance per pin.
It is the fastest DRAM solution in the industry, being able to transmit 124 FHD (full-HD) movies (3.7GB each) per second. The density is 16GB by vertically stacking eight 16Gb chips through TSV (Through Silicon Via) technology, and it is more than doubled from the previous generation (HBM2).

HBM2E boasts high-speed, high-capacity, and low-power characteristics; it is an optimal memory solution for the next-generation AI (Artificial Intelligence) systems including Deep Learning Accelerator and High-Performance Computing, which all require high-level computing performance. Furthermore, it is expected to be applied to the Exascale supercomputer – a high-performance computing system which can perform calculations a quintillion times per second – that will lead the research of next-generation basic and applied science, such as climate changes, bio-medics, and space exploration.

“SK hynix has been in the forefront of technology innovation that contributes to human civilization with achievements including the world’s first development of HBM products,” said Jonghoon Oh, Executive Vice President and Chief Marketing Officer (CMO) at SK hynix.
“With the full-scale mass-production of HBM2E, we will continue to strengthen our presence in the premium memory market and lead the fourth industrial revolution.”

Annotation
HBM (High Bandwidth Memory)

• High performance, high bandwidth memory products that adopt TSV technology to dramatically accelerate data processing speed over traditional DRAMs.

TSV (Through Silicon Via)

• An interconnecting technology that connects the upper and lower chips through thousands of fine holes on DRAM chip.
• Delivers data, commands, and currents through column-shaped paths that penetrate the entire silicon wafer thickness after stacking multiple DRAM chips on the buffer chip.
• Up to 30% decrease in size and up to 50% decrease in power consumption over existing packaging methods.

Standards for data process speed conversion

• 1GB = 8Gb
• 3.6Gbps per pin with 1024 data I/Os (Inputs/Outputs) = 3686.4Gbps
• 3686.4Gbps / 8 = 460.8GB/s (Gb -> GB conversion)

 

Another and bigger Navi chip? First indications of a possible multi-chip package!

It’s no secret that the well-known Twitter-Leaker _rogame now runs his own website with hardwareleaks.com that I also visit there from time to time, certainly not.
Based on the latest news about a new navi entry in Apple’s drivers for the upcoming macOS 11 “Big Sur”, which was announced at the WWDC conference recently and is supposed to set the course towards own chips,
Rogame has now published a news item about this, which I have followed up first.



Since AMD is the official GPU supplier for Apple’s MacBook Pros, iMacs, Mac Pros, AMD has to provide its GPU drivers for every MacOS release, which logically includes upcoming products that have not yet been launched.
Especially beta versions for upcoming macOS releases contain support for unreleased AMD hardware from time to time, and this macOS 11 beta version is no exception.

The fragments contain the confirmation for Navi 21, 22 and 23, which I myself had already mentioned in the nomenclature some time ago and which can be found again with it.
Especially since my own sources, especially Navi23, have now confirmed that I’m not in the Apple universe. It remains to be seen whether this is a rebrand of the current generation of navigation systems,
whose models are one performance class lower, but that too remains pure speculation for the time being, based only on leaks.

The following graphic shows the publication of Rogame, which by the way also reaches as far as Cezanne, but the APU would be a different topic again. Rogame concentrates primarily on Navi 31,
the actually much more important part, namely the MI100 and MI200 only mentioned in passing. And this is where my own sources come in.

More -> https://www.igorslab.de/en/even-bigger-navi-chip-first-indication-of-a-possible-multi-chip-package/

 

Who needs Big Navi, when we can have Ultimate navi.

 

And then we get a Big Ultimate Navi.

 

I just need good GPU for 100-144FPS at 144p HDR (Ultra + RT)
... at max 699-750$

 

And I 60 fps @ 5K with some reshade

 

60 FPS Ultra @4K in current most demanding games, with a little extra juice for the future is what I "need/want". I'd be very happy to get it for $700, but I'm worried it will be closer to $1K