General Discussions

[ajlueke]

Looks like we are getting the 5700 XT for $449 and the 5700 for $379.

Too me, the result is a little underwhelming, and more of the graphics market stagnation we have seen over the past few years.  The AMD 5700 XT is essentially on parity with the current NVidia RTX 2070, for a $50 reduced MSRP, but higher board power (225W vs 185W). 

NVidia started that trend nearly 9 months ago with the launch of Turing.  Turing, ultimately, performed exactly the same in current titles for the dollar as the now 3 year old GTX series.  The only thing additional for the money was the inclusion of RTX cores which still have extremely limited use.

AMD now, has followed that trend,  dropping the 5700 XT in at $449.  Vega 64 when it launched 22 months ago, was $499 MSRP.  So in 2019, AMD is offering you about 14% more performance for $50 less, with the added benefit of a 50W reduction in TBP.   Not exactly earth shattering. 

But Navi, clearly isn't the high end RDNA part.   The unified shaders, texture mapping units, and compute units are all about 60% of the previous Vega cards.  In fact, the standard 5700 has virtually identical shader, TMU, and CU numbers as the Polaris based RX 480/580/590.  But AMD has chosen to price them at Vega levels as opposed to Polaris levels.  That means the Vega price point will now be occupied by Navi, and when AMD releases a Navi variant with Vega levels of shaders/TMUs/CUs it will be priced likely at Radeon VII levels or beyond.  Meanwhile, the Polaris price point will be occupied by...well Polaris I guess.  It is possible that AMD may release cut down Navi variants at lower prices, but they will likely perform similar to the Polaris parts.   

Form AMDs perspective, a Navi GPU is a much smaller piece of silicon than a Vega GPU, the actual die size is almost 50% of the Vega die size (251 mm^2 vs 486 mm^2).  They can get twice the number of GPUs per wafer than Vega, and by pairing it with GDDR6 vs HBM2 they save some money in that regard too.    That small GPU performs slightly faster than Vega, so by pricing it at those levels which is still competitive with NVidia, and eliminating Vega, AMD makes vastly more money on every GPU it sells.  That will also be true as AMD moves Navi down the stack.  But for us end users, it seems like we have identical performance for the money, but for AMD each sale is vastly more profitable. 

https://www.anandtech.com/show/14528/amd-announces-radeon-rx-5700-xt-rx-5700-series

[ajlueke]

On the plus side though, Vega prices are pretty low right now as vendors clear the inventory.  So you can get a card that performs almost as well as a brand new Navi for a substantial discount. 

[jdrobinson314]

I think it's pretty exciting - the introductory Navi cards strategically targeted at a wide audience for a price point which favors introducing a highly tweaked-to-the-point-of-being-new architecture to as many mid range Guinea-pigs as possible. And for good reason - this accessibility directly translates into data which can be used for further tweaking to the RDNA architecture and it's done so in such a way as to target that larger market without actually needing to gain market share at the extreme high end of the market.

So if they had overwhelmed us with a $800 msrp or more than it wouldn't be as grand of a flop if it did fail to impress ( even with say, 64 or 80 CUs )...it wouldn't be a catastrophic loss or setback. Well played. Learning from the competition too.

I'm looking forward to the board partner cards, and all the rest of it. Particularly the lower end integrations and the Pro series as well.

If I were going to overwhelm you with a card, it would be 80 CU card with HBM3 - in this type of configuration I would use lower capacity 7nm hbm3 stacks ( more numerous chips ) would provide very high bandwidth and possibly have better silicon yields on account of this size to silicon platter cost ratio, so lower cost because of this as well. Not entirely infeasible since they're working with Samsung on RDNA as well there could be an IP for Silicon/IP as a potentially at cost trade between AMD and Samsung where AMD gets HBM3 chips for licensure of RDNA while Samsung fabs the HBM3 and hands some of it off to AMD at a discount and Samsung integrates RDNA into their mobile chipsets ( possible via Samsung's own 7nm EUV process ).

So a dual 40 cu with 4 stacks of 4gb HBM3 chiplets all on an infinity fabric interposer would have the same physical footprint and layout as a dual 40 cu, with 4 stacks of 8 or 16 or 24 GB HBM3 ( think pro or instinct cards ).

So AMD may or may not be locked into a production contract as it was with GloFlo, but it still has the ability to license its IP to Samsung for producing RDNA custom chips specifically for Samsung's use in the same manner as it has with Microsoft and Sony for their consoles...its a tad different but not too far a stretch as an example. 

[zasxas]

For me, the prices are horrible. What is AMD thinking? I know what, they just don't care. Customers managed to first ignore the good products from AMD, and they instead bought Nvidia. Nvidia right now has set the prices so high, and since AMD saw that people don't care about their good valued products they followed Nvidias path. The market is going horribly wrong. I want to buy a new gpu for my system right now, but since the prices are so high i might don't even buy. I hate to say this, but i hope Intel will help getting the gpu market prices down...Horrible times.