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

In this case we have no idea how many are there and how where they made.

 

Thread reply bans are being handed out to anyone who insists on turning this into a comparison thread or bringing Nvidia into the mix after more than enough warnings.

 

I replaced upper midrange with performance.

 

If AMD decided to go 256-bit and large cache route instead of wider bus, they will have a good reason. Question is even if this is biggest navi which i doubt, there is no way its just RTX3070 +10% when its tied to RTX3080 at 4K. 3080 shoud be 30-35% faster than 3070 at 4K so this RX6000 showed is at least 25% faster than 3070/2080Ti at 4K and prettly impressive alone if its below 300W TDP, beeing it AMDs best, 256-bit, 16-14gbps ot whatever...

i think its not HBM neither 512-bit, so should be 256-bit for 16GB sweet spot. If its not biggest navi, even more impressive and well they could definitely take performance crown this gen. RTX3080/3090 better relative performance at 4k is at least in part due to high bandwidth, and its almost sure RX6000 will have less bandwidth, so RX6000 performance at 1080 and 1440p should be even better. I guess when talking about performance crown, its not just 4k.
Come on, AMD was just showing what they can do at 4K, more than 60fps, and confirmed that they´re here really shooting for highend this time, not sure if enough for performance crown but at least quite close to 3080 for sure. Will they "disrupt" 4K gaming as promised? All points to RX6000 performing better at lower res, but not bad performance at 4K either, so lets wait and see in a couple weeks, maybe some "disrupting" prices coming? Seems difficult considering RTX3080 already disrupted 4k gaming prices compared to 2080 Ti...

 

Let me be the pessimistic one and say those numbers are from their high-end offering. Have we forgotten "Poor Volta" already?

And that performance is more than good enough for the majority. If it's akin to RV770 then it will win based on the price alone. The Radeon HD 4850 and 4870 were dirt cheap for the performance they offered.

RDNA2 supposedly has a less complex PCB (only 256-bit memory bandwidth) and uses GDDR6 (cheaper).

For some reason $700 graphic cards have been normalised and talked about as being midrange.

 

how possible it is for amd to make a similiar setup for rdna 2.0?

 

Question doesn't make sense without more specifics? Existing GPU design paradigms don't have a split L3 cache (or any L3 cache for that matter) which would "benefit" from unification. L2 cache is also shared and not split either. If yu mean more cache in general that's generally beneficial but not

 

One question I have is if AMD is really going with 128MB L1 and L1 coherency among all shader processors, would they still need L2? At least if they used L2 as victim cache after L1, it would make little sense.
Then again, in RDNA1 the ROP were clients of the L2 IIRC.

 

In theory with a 384 bit bus, you can get to 16 GB, with 8x 1GB and 4x 2 GB GDDR6 chips.
Though would not have much benefit over 12 GB, as the extra 4 GB is on a 128 bit bus.
Still very intrigued why they would go with 256 bit, where the XBX has 320 bit.

 

It’s not gonna be cheap. Not when AMD can sell two 70mm2 CPU chiplets for 800 USD instead.

 

At the moment, Ryzen has a mindshare that Radeon does not.

Ryzen isn't considered a "cheaper alternative to Intel" anymore, as it seems to have taken over the DIY market by storm.


OTOH, Radeon's marketshare on DIY is still lagging behind.
I agree that AMD should try to make a lot of money with their GPUs, but I don't think this is the time for that.

 

If we follow that logic, there's no reason for AMD to produce anything except EPYCs and Threadrippers as the frequencies there are much more closer to the V/F sweetspot (variation in silicon quality is less pronounced), margin is far greater and so on. There's also the "professional GPU" market - former Quadro/FireGL (or how are they called now) cards are usually getting sold for thousands of dollars or even 10k+ dollars (as in case of RTX 8000).

 

I'm genuinely interested in the TAM of the Pro cards. How many CAD/ modeling workstations truly require these pricey cards (ehm, drivers)? Compared to gaming it seems rather small. HPC/AI cards are present at every uni but Pros?

 

The Pros are mainly for video editing IIRC. The need could be rather large if you're doing lots of video encoding/decoding, a if you worked at a place that was constantly producing 4K youtube videos weekly.

 

AMD, like most companies, want to have multiple revenue streams. There’s not a lot of room for them to grow in DIY gaming PCs. They have much greater growth potential in discrete GPUs and laptop APUs. Apple moving to 5nm for their iPhones should mean that TSMC 7nm capacity should be sufficient for anything they’d like to produce,

 

It can't be L1 since you won't even get an integer number by dividing 128 by the number of CUs.
Moreover, L1 has enormous requirements for bandwidth and latency (less for GPUs though), it will cost an arm and leg in this case.

If AMD went for large L2 or L3 cache, I can imagine they did it to be competitive at least in some cases with Ray-Tracing.
As long as BVH fits in the cache, they should be good. Though, for long ranged rays and disperse rays, this might be PITA.
They will cut to the minimum BVH for games with reflections == fewer objects in reflections, smaller draw distances for reflections will be standard optimizations in AMD games.
I am expecting there will be quite a good performance when something fits nicely in the cache and in case of blending.
Although, there will be a lot of performance holes, so perf will vary wildly from case to case and title to title.

 

I think this is mixing different concepts.
The clustered L1 concept was intended to not increase the L1 at all, and it didn't make the L1 generally coherent since groups didn't span more than a subset of the CUs.
The version discussed in the paper wouldn't have made them acceptably coherent because it was being applied to a Kepler-like GPU.

The 128MB rumor has multiple versions, none citing the L1, and one version stating it's outside the existing cache levels.

 

There's nothing that prevents you from using a Quadro card for computational tasks if your algos are not reliant on fp64 that much and are severely bw-limited. So i'd say it's far greater than the classic "workstation" applications