Radeon Instinct is not a graphics card, it is an accelerator. That is a different market and price bracket.
No, they are not. if the Frontier inferences faster and cost 45% less who is going to buy the MI25?
These two card a no where near as rigidly segregated as you are making them out to be. They aren't.
The Frontier does more and so probably will cost more. Does more and cost less. They are both in the "Professional" card bracket.
I don't see why Frontier cannot cost $999 when it is simpler than Duo, which compensates for the bigger chip. I think Duo has been obsoleted, multi GPU are fringe products.
It isn't simpler than than Duo in terms of yields. It isn't just that the chip is bigger ( so have process defect yield to get around) you also are clocked about 29% higher. So not only have to bin on defects but also on clock stability. That will push down yields. You have imposer and HBMv2 connections.
In contrast, the Duo is a smaller chip and likely on 2nd iteration of process design ( the tweaked Polaris for the RX 500 series); so defect yields are likely higher than 1st generation. Instead of clocking them to a higher clock, they leave it at the old WX 7100 rates. (so slower). Binning for slower clocks is easier so yields higher on both fronts.
Furthermore, The end user cost of the Frontier is the value it produces for the customer. If it is faster than previous generation inference case then it is worth something close to the prices charged of those products (after all those products had buyers.)
The Duo is pointed largely at substantively different workloads and/or contexts. Obsoleted 1-2 month(s) after released? Really? IF AMD thought they extremely overlapped that much and the next one was cheaper why would they released the Duo? I could see maybe if the Duo was a 6-7 month old product but it was just announced last month ( late April for May release ). The Duo probably has a lower TDP envelope which means can fit in more systems with more mainstream power supplies. It probably has a cheaper price too.
[ both of which are true relative to the previous Pro Duo that was based on Fiji.
Fiji 350W TDP 16 TFLOP peak total VRAM 8GB initial price $1,499
Polaris 250W TDP 11 TFLOP peak total VRAM 32GB initial price $999
it is not a drag race TFLOP targeted card.
]
Price it at $999 and an ecosystem will definitely grow. Average consumers will not buy it.
But will AMD grow? AMD could sell everything at cost.... I'm sure the customers would be happy while it lasted. Here is the real compute card market.
http://www.anandtech.com/show/11403/amd-unveils-the-radeon-vega-frontier-edition/2
Frontier Edition
is a compute card (that happens to have graphics output too) so it will probably be in the >$1,000 bracket. The Pro Duo isn't primarily targeting compute ( this Pro Duo walked backwards from the previous version) and it is just under the $1,000 mark. Probably in the range in the last 16TFLOP range card they had (even if changing focus from single to half precision).
The Frontier isn't the last Vega based card that AMD is going to do. That is way it is named Frontier. It is the first, "early access", bleeding edge Vega card. That isn't going to come cheap nor does it benefit AMD to sell it cheap. AMD needs to make some better margins. They can't keep borrowing money to keep the lights on.
P.S. Duo GPU cards are more more likely to evolve than disappear. Something like
Vegalike-GPU <--- Infinity Fabric ---> Vegalike-GPU <--- pci-e -->
wouldn't be very at all at some point. You can see the problem with Nvidia V100 in that they have reached limits of just how big can make the die. It is more cost effective to start to couple smaller dies together . You can say that a multiple chip module (MCM) GPU isn't a the fringe product market... but that board you are adding to a computer system is essentially a very large module. So no, not going to disappear. Just going to get smaller.
