AMD: Navi Speculation, Rumours and Discussion [2017-2018]

[Anarchist4000]

That roadmap was a ROCm roadmap, so there was no reason for it to mention anything about gaming.

True, but in AMDs case they are often the same. The only FP64 part being Hawaii, so a specialized compute Vega on 7nm makes sense. Navi likely following the Zen MCM path and not as well tested for HPC as Vega would have been at the time. New node, but architecture would be established.

 

[Mat3]

I thought Powervr was the only architecture that is a true TBDR, even among other mobile GPUs. Does anyone else attempt to bin the entire scene before any rasterization?

 

[Anarchist4000]

I thought Powervr was the only architecture that is a true TBDR, even among other mobile GPUs. Does anyone else attempt to bin the entire scene before any rasterization?

Not that I'm aware, but consider ROVs and OIT for a second. All transparent fragments are Deferred and the architectures ideally process Tiles. The benefits of TBDR are understood and it seems a logical next step in light of Imagination's current status. A step further with GPU driven rendering would seem binning cubes. That would open up Ray tracing and more complete screen space reflections and the like. The concern is sufficient cache to store and process everything. Both Vega and Volta greatly increased cache, so it seems reasonable the IHVs may see it coming.

 

[Ryan Smith]

One thing to note though is that if you did have a true TBDR, the last thing you'd do is tell Imagination. It's a good way to get yourself taken to court to see if you violated any patents. It's similar to the reasoning behind why Qualcomm doesn't talk about its GPUs.

 

[fehu]

Qualcomm: what's a GPU?

 

[AlNom]

A miserable little pile of transistors.

 

[Grall]

It's similar to the reasoning behind why Qualcomm doesn't talk about its GPUs.

This strategy seems to skirt awfully close to the textbook definition of willful violation...

 

[Deleted member 13524]

One thing to note though is that if you did have a true TBDR, the last thing you'd do is tell Imagination. It's a good way to get yourself taken to court to see if you violated any patents. It's similar to the reasoning behind why Qualcomm doesn't talk about its GPUs.

So a honest and legal situation where IHVs simply pay IP royalties to Imagination is far from the realm of possibilities?

 

[Alexko]

I'm no expert but patent litigation is quite the rabbit hole: you can never know beforehand how deep it may go—or how much it may cost you, in this case. Erring on the side of caution seems preferable, at least to those who make the decisions.

 

[silent_guy]

So a honest and legal situation where IHVs simply pay IP royalties to Imagination is far from the realm of possibilities?

If Imagination stays a real company instead of converting to patent tr^H^Hlicensing company, then yes. After all, there is no way Imagination isn't stepping left and right on IVH as well.

 

[Ethatron]

One thing to note though is that if you did have a true TBDR, the last thing you'd do is tell Imagination. It's a good way to get yourself taken to court to see if you violated any patents. It's similar to the reasoning behind why Qualcomm doesn't talk about its GPUs.

You just brought Qualcomm into really really big trouble, Sir!

 

[Malo]

So a honest and legal situation where IHVs simply pay IP royalties to Imagination is far from the realm of possibilities?

Hasn't that been an option for GPU designers for decade or so already? It has IMG never allowed licensing if their tech?

 

[Anarchist4000]

So a honest and legal situation where IHVs simply pay IP royalties to Imagination is far from the realm of possibilities?

They certainly need the cash, and it does seem odd we haven't seen any legal maneuvering if that's the case. May be some simple technical maneuvering to avoid it.

If Imagination stays a real company instead of converting to patent tr^H^Hlicensing company, then yes. After all, there is no way Imagination isn't stepping left and right on IVH as well.

Even without trolling, you expect them to be pushing licensing as a means to increase share price with all these buyout rumors. Anyone in the business could consider a buyout at the current market cap of less than 500m and patents and existing business worth something. I could see Intel going that way if it avoided some of their current licensing issues through MAD.

 

[3dilettante]

One thing to note though is that if you did have a true TBDR, the last thing you'd do is tell Imagination. It's a good way to get yourself taken to court to see if you violated any patents. It's similar to the reasoning behind why Qualcomm doesn't talk about its GPUs.

It might take a little bit more pre-fishing than that before the court entertains a suit. For example, the WARF patent lawsuit with Apple in part required an engineer citing a University of Wisconsin alumnus' patent as prior art, which led to some documents indicating awareness. Intel also wound up paying for the same patent, although I believe the university or the foundation put in the work to create code that could detect the signature of predictive dynamic memory disambiguation that matched the patent.

Then again, I'm pretty sure there's a bunch of patents across various vendors concerning rendering pixels to rectangular regions on screen or in memory...

This strategy seems to skirt awfully close to the textbook definition of willful violation...

Only if you're looking, or accidentally write down that you did.
The usual way of dealing with it is to just not know if you infringe, and hope to license or be a big player who can infringe and settle with the bankrupted plaintiff later.

I'm no expert but patent litigation is quite the rabbit hole: you can never know beforehand how deep it may go—or how much it may cost you, in this case. Erring on the side of caution seems preferable, at least to those who make the decisions.

The inevitable counter-suits are likely enough to ruin a practicing entity looking at litigation as a last resort. It burns a lot of bridges, and can drag partners/customers into the litigation. That's why it often doesn't happen until a company is virtually certain it is done--but not always as, Rambus shows.

 

[bdmosky]

Now, I realize the patents are quite old, but didn't Gigapixel have TBDR patents which was acquired by 3dfx which then went to nVidia?

 

[Grall]

Wouldn't those patents be expired by now tho? 3dfx died an eternity ago, comparatively...

 

[BRiT]

My initial take: Not if they do tweaks/enhancements to the tech which then extends the patents.

Also, Utility patents generally last 20 years, while design patents last 14 years or 15 years for those filed on or after May 13, 2015.

 

[Magnum_Force]

Greenland lives?

http://electroiq.com/insights-from-...rformance-computing-chiplets-and-interposers/

 

[ImSpartacus]

Greenland lives?

http://electroiq.com/insights-from-...rformance-computing-chiplets-and-interposers/

That's a neat article, but I thought Greenland was the original "Arctic Islands" codename for Vega 10 in the same way that Baffin and Ellesmere were the original "Arctic Islands" codenames for Polaris 11 & 10, respectively. All three are islands in the northern hemisphere.

 

[Magnum_Force]

Could be, but I've only ever seen Greenland mentioned as part of an APU.

What I find most interesting in the article is the picture under the "thermal Issues" section. It mentions a "best mean configuration" of 320 CUs, 1000mhz and 3TBps, and another "best workload specific config" of 384 CUs, 700mhz and 5TBps.

These frequencies are lower than what Vega can handle, and it appears the lower heat produced by the lower clocked config produces higher bandwidth (unless I'm mistaken in thinking TBps is TeraBytes per second)