flash it into an MI50!

 

I'm sure someone will figure out how to flash the Mi50 bios on it. The question is if it's a software cap or a hardware cap.

 

the 3rd one should be "cheaper cards" imho, just to rub on Nvidia's RTX pricing.

but to answer his question seriously, i think they should look into promoting SLI NVLink more.
edit: on the other hand, i was under the impression that they were from Nvidia and talking about RTX... derp.

 

How about SRIOV?

AMD same as Nvidia cuts out all the "cool" features that enthusiasts would find interesting and leaves them to workstation/server cards only and asks even 10x price for those at times.

The US government business manipulation/wars are endless. Too scared and paranoid anyone else will be better by a hair. Well guess what, they already are, everything is made in China thanks to western governments not protecting local investments into manufacturing. China right now is on dark side of the moon, literally. Meanwhile US's space program is dead very dead, taken over by private sector. Forcing these blockades, tarifs and what not only hurts the one who issues them. Manufacturing of electronics will not move back to US because there is no "back", it was never made there (modern electronics, phones, PC parts, ...), they penalize imports from China... well they move to Vietnam, Taiwan, ... any nearby country that has a decent logistics connection so that they can keep using the same distribution channels and partners in south east Asia.
The technological fear of leaking anything from US is unreal, blocking specific technologies and even access to them in US itself if you want to work on them.

 

ironically the US has one of the largest landmass to contribute for productions, taiwan in comparison isn't even as large as one state of america's.

 

They should start making chips big again, stop this endless downsizing!

 

Bring on the V100 gaming cards! I'm sure the price will be reasonable.

 

When I was trying to get a figure how to build simple and cheap machine for bruteforcing I used 4x 7970s 100 dollars each, 1000Glops Doubleprecision each. Single 7979 is capable of 10k password attempts on locked ZIP file, and when i knew it was password with 12 digits, it broke it in 1 minute.

6000gflops in Double Precision would do that in 10 seconds, 4 cards will do that in 2.5...

It would be way to cheap to have that kind of bruteforce sold to random customers in hundrets of cards...

 

Well, damn. Got me all excited here for a moment.
FP64-enabled Radeon VII would have been a kick-ass replacement for my quad-Titan Black HPC rig.
Oh well.

 

AMD, crushing your dreams since 2006.

 

So true [emoji3]

 

https://techgage.com/news/radeon-vii-caps-fp64-performance/

I knew it! Enjoy your overpriced Vega VII deep learning crippled Mi50

FP64 has little to do with deep learning or gaming and capping it on these parts is a completely reasonable way to segment the market.

Most deep learning work is standard or half-precision, with the trend going to even lower precision. This is why NVIDIA's Tensor cores don't have native FP64 at all, but have significant FP16, INT8, and INT4 performance.

 

There are many deep learning frameworks that require FP64 just because researchers don't use it since they are focused on time constraints doesn't mean FP64 isn't used. It makes sense why they've segmented their market and I expected them to do so.

 

https://techgage.com/news/radeon-vii-caps-fp64-performance/

I knew it! Enjoy your overpriced Vega VII deep learning crippled Mi50

The article must have been updated since your post. In now says:

That means instead of delivering 6.7 TFLOPS of FP64 like the MI50, Radeon VII will be closer to ~862 GFLOPS (it’s 1:16 with single-precision like RX Vega).

The edit is correct, both Polaris and Vega are 1/16.

 

The article must have been updated since your post. In now says:


The edit is correct, both Polaris and Vega are 1/16.

Thanks, updated the OP with the change.

 

Impending $699 disappointment.

 

I still would like to see if the card can undervolt. If the card performs the same as a GTX 1080 Ti using 180w it's still decent. I suspect it being the same as Vega with the default set to the power saving 220w power limit. With the power target set to max it will be 300w.

I suspect a 2Ghz version in the future with a CLC for $500 more just like the Frontier Edition with a $500 price for the CLC.

 

I'm disappointed now, if it was even 1/8 I would think about getting one. But now, I don't think I want to spend that much money on something I will probably never attach to a display.

 

I understand that, but i guess none of us want to see chinese hackers try 60k password attempts in a second with single card.

 

Do you know the price of MI60/MI50?

 

Too much for normal people. Like 6-7-8k$. And probably not available outside of DC deals.

 

And now...

we have a rumor by TweakTown which states that the AMD Radeon Vega VII graphics card will have less than 5000 units made during its production cycle and each card is going to be sold at a loss considering these are just repurposed Instinct MI50 parts that could’ve been sold for much higher prices to the HPC sector.

Not a surprise, really, given its big die with lots of HBM.

And no AIB models, too.

 

Alright, I finally have the Radeon VII FP64 performance matter sorted out with AMD. Contrary to earlier statements, it is being throttled. Radeon VII's rate will be 1:8, versus Vega 20's native 1:2 rate. Notably, this is still twice the native FP64 rate of all other Vegas

https://twitter.com/RyanSmithAT/status/1085680805802733568?s=19

Oh well
Not much to this after all.

 

Hmmm....If it is 1:8, it's back on my radar. I might get one...hopefully at a good price.

 

Actually 1:8 FP64 isn't that bad.
It certainly could have been worse, but still pales compared to HD 7970/R9 280X.

 

HD 7970 is 1 TFLOP FP64


1/8th of a 12 or 13 TFLOP FP32 is still at least 60% faster.


However, the GTX TITAN Black (Kepler) was 1/3 FP64 and it resulted in higher than this prospective VEGA 7nm.
It's also slower than the Firepro W8100 (~2 TFLOPs FP64).