So it looks like consumer Vega in any quantity won't arrive until August. That's basically one year after Polaris.
Hasn't their gfx stuff slipped rather backwards in comparison?which, imm, makes their gains this last year all the more impressive (tech and finances/company health and efficiency). Those folks have been hitting on all cylinders (the ones that they have, anyway ) and punching well above their resources. It's really quite admirable.
The card Raja held up on stage was only 8+6 as some people pointed out, I went back and got a screenshot of it:
7nm is Navi and don't expect it to launch in 2018. I don't believe GF will be able to produce GPUs before the end of 2018 with launch early 2019.
I haven't seen anybody bring it up yet, but personally i'm hopeful that 8GB stacks will clock higher than the 16GB ones, and we might see 512 GB/s yet.Seeing 480 GB/s HBM2 is good and bad. Good: RIP 350 GB/s leaks. Bad: Still no sign of 512 GB/s. While surely not much of a difference in performance, from a brand strength standpoint it looks weak to have less than your predecessor. Consumer Vega, especially if it comes later could have faster.
GTX 1070 release date was June 10This is really sad.I am waiting for Vega since GTX1070 launch.I hate GTX1070 but AMD year later still dont have anything i can buy and probably next 3 months wont have(i want aftermarket air cooled vega.Not watercooled or with crap reference blower).This is just so sad.
Hasn't their gfx stuff slipped rather backwards in comparison?
Fury at least got close to the 980ti, this stuff seems likely to launch vs Volta, and the top card may well struggle to keep up with the 1170.
Doing quite well to be anywhere of course.
In tech markets it is a very difficult position to ask your customers to wait a year, then ask them to wait some more.GTX 1070 release date was June 10
And their market share had no hit due to the missing high end offering. Meanwhile they started spending a lot more in marketing. I see a pattern here...
As has been repeated innumerable times in this thread, adding a GDDR5(X) PHY to the die would massively affect die size. For sufficient performance, you'd need at least a 384-bit bus, if not 512 - that's 6 or 8 64-bit PHYs, each 3-4x the size of one of the current two HBM2 interfaces. If Vega without GDDR5(X) is 500-550mm2, it would easily be 600-650mm2 with GDDR5 support. Not only is it a complete waste for the highest end SKUs, but it would significantly affect yields and production costs.So again I'm wondering why AMD didn't provided an option with a GDDR5X implementation.
In tech markets it is a very difficult position to ask your customers to wait a year, then ask them to wait some more.
Launching a professional card won't help that.
simply because AMD gives away their Polaris with no profit in a high volume category. It's so huge issue for AMD that they decide to attack first the pro market with Vega. One slide says it all:So how do you explain this year of stable/slightly increasing market share?
More likely is that Frontier Edition is 8+6 pin connector, and the Nano is just 6 pin connector version.It could be 2x 8pin for 1.5-1.6GHz and 1x8pin + 1x 6pin for 1.2GHz I suppose.
Most likely the TDP for the Frontier Edition GPU is 250W.I could see it making sense if they had some chips that had disabled CUs or some chips that couldn't clock very high so AMD creates a bin for those and targets a 150W TDP or something like that and slots the card up against the 1070. Maybe it's just a minimal bin that sits right above the scrap heap to let AMD recoup some cost, but I can't see it being a money maker for them.
(slide cut, emphasis mine)simply because AMD gives away their Polaris with no profit in a high volume category. It's so huge issue for AMD that they decide to attack first the pro market with Vega. One slide says it all:
Is it realist to believe that a single high end Vega cpu will equal 2 RX480s in CF?
(slide cut, emphasis mine)
This is interesting, care to back this with some accurate data?
not difficult to find out. Google AMD quarterly financial report and you will see that the radeon division is loosing money...(slide cut, emphasis mine)
This is interesting, care to back this with some accurate data?
margin different than profit... see my last edited postyeah no profit...that's an interesting assertion that tends to reject the last 2-3 qtrs of ~30% margin claims. anyway, this is starting to derail.
You understand that Computing and graphics accounts for CPUs and GPUs?not difficult to find out. Google AMD quarterly financial report and you will see that the radeon division is loosing money...
edit: http://www.anandtech.com/show/11092/amd-announces-q4-2016-earnings
AMD Computing and Graphics
Q4'2016 Q3'2016 Q4'2015
Revenue $600M $472M $470M
Operating Income -$21M -$66M -$99M
if it's designed to be a data center card
But it does seem like it's almost a pride thing with AMD and HBM/2 at this stage. I'd guess after the Fury HBM1 shortage it was too late to do much about it, but doesn't HBCC imply they have at least somewhat abstracted memory?As has been repeated innumerable times in this thread, adding a GDDR5(X) PHY to the die would massively affect die size. For sufficient performance, you'd need at least a 384-bit bus, if not 512 - that's 6 or 8 64-bit PHYs, each 3-4x the size of one of the current two HBM2 interfaces. If Vega without GDDR5(X) is 500-550mm2, it would easily be 600-650mm2 with GDDR5 support. Not only is it a complete waste for the highest end SKUs, but it would significantly affect yields and production costs.