Not bad for a 'cheap' card!
Not bad for a 'cheap' card!
Looks like AT has inadvertently confirmed the 1080MHZ base clockspeed of the RX480:
http://www.anandtech.com/show/10418/...-zen-or-jaguar
I think that's just derived backwards from the >5TFLOPS figure and shader count - 5TFLOPS would require a clock speed of 1080MHz.
if the process was good from the start .. could they not bring out a rx490 ? now that would be a shock to nvidia .. yes it would be an o/c version of the 480 .. but hey why not ..
What does it matter now if men believe or no?
What is to come will come. And soon you too will stand aside,
To murmur in pity that my words were true
(Cassandra, in Agamemnon by Aeschylus)
To see the wizard one must look behind the curtain ....
Actually (I think you'll find ) 1080MHz with 2304 shaders would be a shade less than 5TFlops (4.977, to be precise ). Looks like 1080MHz is probably the minimum base speed for the 4GB version of the 480 (as previous leaks have pointed to) - you'd need 1086MHz to hit > 5TFlops. Presumably - much like "representative" APRs on loans - it hits at least 1086MHz at least 50.1% of the time
Given the clock speeds we've seen leaked - 1080MHz/1266Mhz - I wouldn't be at all surprised if those turned out to be the base and boost clocks. Or perhaps AMD have gone back to having a performance mode with higher clocks/voltage/fan speeds. After all, they've got form for that kind of messing
I guess at this point it'd be nice to just have a proper review of the card - Don't understand why they're dragging the launch out for so long...
Hopefully so that the warehouses are full of well tuned custom cards.
OTOH, Nvidia seem to be doing fairly well out of their limited availability launch. Even planning a Titan version supposedly: http://www.fudzilla.com/news/graphic...an-coming-soon
Just how much would people pay for a graphics card?? $1000 already seems pretty bonkers to me, do you think some people would go higher than that if a faster chip came out?
I suspect so. I know someone who just stumped up the cash for a complete rig with a pair of water cooled 1080 cards in SLI. But then at his age I would probably have spent the money on cars and petrol which is no more sensible really.
*shrug* $999 cards/CPUs have always sold, in small volume at least. Suspicious of the idea of nvidia doing a silicon run just for a small-volume Titan release though - seems ... unlikely. And more to the point unprofitable - I'm pretty sure all those masks and tape-outs are expensive. Any chance that GP100 has a GDDR5 memory controller baked into the silicon somwhere? Harvested GP100 dies using GDDR5X memory would make more sense to me than a separate die for a single product line.
Expensive? I see mask costs of $10M bandied around for a modern big chip with lots of multi patterning going on. So if Nvidia can make $200 profit on a sale it would take 50K cards to break even. That sounds a stretch, but possible. That's assuming the factory is adding 20%, graphics card company makes 30%, retailer makes 30% that makes a $1000 graphics card about $500 in bill of materials. Lots of fast ram isn't going to be peanuts, but I expect the Nvida chip is going to be a big chunk of that cost. Let's guess $300, of which $100 goes to TSMC to make it.
Harvesting GP100 doesn't make much sense. It is rigged to do high precision maths, not the 32 bit that we want for graphics. That uses a lot of silicon, making it probably no faster than the 1080 silicon for gaming use.
scaryjim (14-06-2016)
Fair enough I guess - maybe not as expensive as I thought. Still not sure it really justifies a chip for a single product line though...
I could be thinking about this wrong though. What if GP100 is actually the single-product-line chip? A DP-unbalanced, HPC-specific chip that nvidia can charge as much as they like for because it'll basically pay for itself in just a few dozen sales. GP102 could be a similar size, but less DP-focussed - more like the top chips of the previous generations. It could fill out Titan and 1080 Ti SKUs for the consumer and the bulk of the workstation range for cases that don't require the additional DP throughput (no doubt with the DP capability chopped significantly to "persuade" people who need DP towards the (presumably) more expensive P100). The GP102 wouldn't be specific silicon just for Titan. GP100 would be specific silicon just for the P100. Hmmm....
Ah, yes, I'd forgotten about those. I believe the main wins are for Volta rather than Pascal, but presumably GP100 is essentially a development platform to test the planned architectural enhancements that will go into the Volta GPUs down the line. And then if you can bundle those test modules up into a "deep learning" platform and sell them for $300k each...
Yup, and don't forget there's a bunch of tesla/quadro products that will make use of GP102 as well that don't need the full fat double precision performance of the GP100.
There are currently 1 users browsing this thread. (0 members and 1 guests)