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

Yeah, both in client and DC.
Arcturus is whopping 8 SEs.

 

The logical CU count is 2*Navi10 => 80 CUs with both frontend and backend scaled proportionally. Also finally a truly historical moment - more than 64 ROPs for AMD!

Can the xGMI be a simple copy&paste from compute cards aimed to produce a gimmick like "Cross Fire Reloaded"? Since you know, Navi 21 is a consumer oriented chip.

 

They can do weird shit too.

True that.

No.
mGPU is dead outside of proviz, and AMD isn't exactly very interested in proviz due to that market being pennies.

Not entirely, no.

 

Yep, 30 000M Pennies and probably with margins to die for.

 

Something in the consumer gaming PC space will need to compensate for the embedded hardware decompressors in next-gent consoles.

I.e. we should all want GPUs with NVMe slots in 2021. It might be the best case scenario.

 

Just throw more cores.
Also idunno how that relates to N21 having xGMI.

 

Throwing more cores is worthless.
Decompression isn't infinitely parallel. Kraken is two-threaded and Oodle makes a big deal about it.

When Microsoft says their decompression unit "is doing as much work as 4 Zen2 cores" what they actually mean is "it's doing the same work as a 15GHz Zen2 core". Which no one can produce at the moment.

 

NVMe slots on GPU's won't get around the need to decompress the output from the SSD in realtime. Either shader based texture decompression or less compression with bigger installs are the answer to that.

On the other hand if I/O throughput is the primary concern then you can look to more RAM to mitigate the need, or as above, shader based texture decompression can boost the already very fast capabilities of PCIe4 drives. Or just wait for PCIe5 to blow everything away.

 

Yeah but we can get plenty.

Not coming in client for a god damn while due to PCB material reqs.

 

Lower end of the Radeon Instant product stack for running inference, perhaps. Parity in xGMI across the lineup probably enables direct mGPU interconnect, as MI60 already does but only with its own kind.

The pipe dream would be enabling fine-grained system SVM (that HSA APUs were meant to deliver), but that had been publicly slated for "3rd Gen Infinity Architecture" and CDNA 2.

 

You can just get a MI100, it's fairly cheap should you be annoying enough.

 

Of course the PCIe 4.0 4x connection isn't enough (to reach PS5 I/O throughput speeds). The GPU would need an embedded decompression block, which both consoles already have. That was my suggestion.
In fact, given how much more dynamic the GPU market is compared to the motherboard, CPU and storage ones, I think this is the most likely to come to fruition at the moment.

Besides, we've had GPUs with dedicated NVMe slots for years.

If you're talking about plenty of threads for decompression, then we really can't..

I think shader based decompression is more of a method to offload the CPU than to achieve higher performance (due to low thread counts). Bigger installs alone don't solve the I/O throughput limit.
Getting very large installs with decompressed data with long initial load screens that then make the game occupy a truckload of RAM with decompressed data is a more plausible solution, if we're not getting dedicated decompression hardware on the PC anytime soon.

 

Alleged benchmarks of MI100 shows it beating A100 in FP32 by a large margin, while losing in FP64, FP16, and mixed precisions by a large margin.




https://adoredtv.com/exclusive-cdna-and-mi100-presentation-slides-leak/

 

Rumor: “Big Navi” not so Big?

https://coreteks.tech/articles/index.php/2020/07/29/big-navi-not-so-big/

 

So it's Vega and 5700xt all over again ?

 

It's all so tiresome, yes.
Or at least pretend it's that before launch.

 

That rumor sounds kinda silly. Even a modest 64 CU 2.2GHz RDNA2 GPU should easily beat 2080Ti. And some people here are hinting its actually bigger than that. Yeah I don't buy it for a second.

I'm more concerned about TSMC capacity and what the availability will be like.

 

Of course, they're just generating clicks and covering all the bases.
N21 perf now ranges from 2080ti to 2080ti+50%.
No tier left behind!

This, Rome/Milan orders are big enough for AMD to pretend dGPUs don't exist for at least the next 2Q.
Maybe even into Q1'21.

 

if the price is right it wont matter much.

 

Yikes!