AMD: Navi Speculation, Rumours and Discussion [2019-2020]

Status
Not open for further replies.
CarstenS said:
Maybe you could share the actual numbers including source?
Click to expand...
When it comes to contract pricing, does anyone truly know outside the two parties signing the contract? It's not like it would show up in any quarterly report, so it would have to be due to someone leaking it. Thus - no official source.
For something as volatile as spot market memory pricing, and specific as HBM modules, estimating contract pricing with any precision is going to be difficult.
 
Entropy said:
When it comes to contract pricing, does anyone truly know outside the two parties signing the contract? It's not like it would show up in any quarterly report, so it would have to be due to someone leaking it. Thus - no official source.
For something as volatile as spot market memory pricing, and specific as HBM modules, estimating contract pricing with any precision is going to be difficult.
Click to expand...
Not to mention they're suggesting that HBM2 price is exactly the same it was 1½ years++ ago. They didn't have any sources back then either.
 
I would imagine, something as HBM gen2 would be manufactured to contract and not traded at spot markets. Spot market would imply a vast surplus of unused products from failed/miscalculated contracts.

Now, I don't have exact numbers as well, I just remember Gamersnexus writing about it in 2017 (18 months ago), quoting David Kanter "The HBM2 memory is probably around $150, and the interposer and packaging should be $25." and others on estimated HBM2 pricing.

AFAIK, HBM2 is being used mostly in high margin products like data center accelerators, FPGAs and network switches. All of which seem to be in a steadily high demand. Also, you have only two manufacturers for HBM2: Samsung and Hynix, right?

There are also various and continuing reports that HBM2 is in very high demand. SK Hynix going on record middle of 2017 stating "I can tell you that the customers are willing to pay at least double or, let's say, 2.5 times of the HBM." and Samsung apparently saying it ISC2018 (mid-2018, multiple sources here's one), "even if the company doubles the HBM2 production, they won’t be able to fulfill the market’s demands." They continue to say, that "This is another factor that has resulted in HBM2 being too expensive and a product with a high-profit margin for Samsung and SK Hynix."

Now I wonder, if competition on the manufacturing side demand is that high and for high-margin products nevertheless, why would prices drop drastically? I don't know if $320 is the right number, but I would not discount it as being something conjured up from fairyland -- especially if I had only my gut feeling that it might be on the high side of estimates.
 
Bondrewd said:
Three.
Micron is now also there.
Click to expand...
Strange, I was under the impression Micron expressed mid-2018 it's commitment to working on also having HBM products, but none of it has surfaced for now. At least as I knew.
 
ToTTenTranz said:
One thing is for sure: someone at AMD does like the guy.
I remember noticing how his reach was relatively tiny when Vega 64 launched and yet he was one of the very few who got a Liquid Cooled edition for review. I think his channel had less than 50K subscribers back then.
Click to expand...
It goes pretty much to the top, Lisa Su her self commented on one of his videos. I don't think I have ever seen a situation like that in the hardware industry.

https://twitter.com/x/status/1089715760895852545
 
AMD is patenting Variable Rate Shading
February 28, 2019
AMD has apparently filed a patent application that also describes Variable Rate Shading. The patent was already filed on 25 May 2017, but this was only published now.

"A technique for performing rasterization and pixel shading with decoupled resolution is provided." The technique involves performing rasterization as normal to generate rasterization data and a quadrature quads are generated from the quads in the tile buffering on a shading rate to be combined. The positions of the coarse pixels of the coarse quads are set on the positions of the corresponding fine pixels.The coarse quads are shaded normally and the resulting shaded coarse quads are modified based on the fine rasterization data to generate shaded fine quads. "
Click to expand...
https://www.hardwareluxx.de/index.p...atentiert-sich-ein-variable-rate-shading.html
 
So in other words, AMD is finally catching up to existing Nvidia tech. Cool. Let's hope Navi's 7nm can also match Turing's 12nm in perf/Watt.
 
"A technique for performing rasterization and pixel shading with decoupled resolution is provided herein. The technique involves performing rasterization as normal to generate fine rasterization data and a set of (fine) quads. The quads are accumulated into a tile buffer and coarse quads are generated from the quads in the tile buffer based on a shading rate. The shading rate determines how many pixels of the fine quads are combined to generate coarse pixels of the coarse quads. Combination of fine pixels involves generating a single coarse pixel for each such fine pixel to be combined. Each coarse pixel has coverage data corresponding to the coverage of the corresponding fine pixels (i.e., the fine pixels combined to form the coarse pixel). If the combination of sample coverage data from fine pixels would exceed the number of bits assigned to coverage in a pixel, then some of the coverage data from the fine pixels are removed, leaving a reduced set of coverage data in the coarse pixel. The positions of the coarse pixels of the coarse quads are set based on the positions of the corresponding fine pixels.

In one example, the position of each pixel of a coarse quad is set to a center point of the area of the render target covered by the corresponding fine pixels. The coarse quads are shaded normally and the resulting shaded coarse quads are modified based on the fine rasterization data to generate shaded fine quads. This modification involves upsampling (cloning) each pixel of the coarse quads to generate a number of upsampled cloned pixels each having the color of the corresponding coarse pixel and then applying the fine rasterization data to each upsampled cloned pixels."

-from the patent.
 
So I guess variable rate shading is going to be a standard now?

I'm really not sure it's that useful. Similar results are already obtained for alpha particles, and have been around for ages. They look like crap but have been used as a necessary performance saver. As for temporal downsampling, temporal upsampling and AA and etc. already does this very well. Ultimately I'm not sure this is an entirely necessary hardware feature except for a few very specific use cases. The results from Wolfenstein certainly aren't encouraging. But hell maybe some developer will prove me wrong.
 
Status
Not open for further replies.
Back
Top