Slashdot Mirror
AMD Radeon Fury and Fury X Specs Leaked, HBM-Powered Graphics On the Way
MojoKid writes: A fresh alleged leak of next AMD Fiji graphics info has just hit the web and there's an abundance of supposedly confirmed specifications for what will be AMD's most powerful graphics card to date. Fiji will initially be available in both Pro and XT variants with the Fiji Pro dubbed "Fury" and Fiji XT being dubbed "Fury X." The garden variety Fury touts single-precision floating point (SPFP) performance of 7.2 TFLOPS compared to 5.6 TFLOPS for a bone stock Radeon R9 290X. That's a roughly 29-percent performance improvement. The Fury X with its 4096 stream processors, 64 compute units, and 256 texture mapping units manages to deliver 8.6 TFLOPS, or a 54-percent increase over a Radeon R9 290X. The star of the show, however, will be AMD's High Bandwidth Memory (HBM) interface. Unlike traditional GDDR5 memory, HBM is stacked vertically, decreasing the PCB footprint required. It's also integrated directly into the same package as the GPU/SoC, leading to further efficiencies, reduced latency and a blistering 100GB/sec of bandwidth per stack (4 stacks per card). On average HBM is said to deliver three times the performance-per-watt of GDDR5 memory. With that being said, the specs listed are by no means confirmed by AMD, yet. We shall find out soon enough during AMD's E3 press conference scheduled for June 16.
This card has 400GB/sec throughput on memory. Not that far away, and that's just the first model limited to 4 stacks.
No idea where that imaginary goal of yours came from though. They always marked this as around 100TB/stack. And of course like all such memory, it's going to run in parallel, so the more stacks on die, the more bandwidth.
YAY for the new chip and memory. I just hope the TDP isn't so power sucking as the R9 series.
Don't be apathetic. Procrastinate!
In November 2001, one of the Fury X cards would beat the worlds top supercomputer on raw FLOPS.
"performance of 7.2 FLOPS compared to 5.6 TFLOPS"
I think I would stick with the faster card...
HBM1 gives 1GB and 128GB/s per stack, so 4GB and 512GB/s in this model with 4 stacks.
HBM2 will double both performance and capacity, and is expected some time next year.
They aren't stacked directly on top they are stacked to the side of the GPU .... sort of like how intel puts a GPU on the same processor package beside it's CPU ... just in this case there is an interposer that allows a much more massive number of interconnects than previously practical.
It probably makes things easier to cool than before since the memory now gets cooling it would not have previously gotten... and Memory is really the most important part of the GPU.
Because they aren't Shimano.
The HBM memories run at a much lower clockspeed than the GDDR5, but compensate it by using a very, VERY wide bus, so they're probably a lot colder.
Dude, if you needed a CUDA card, the 980ti is DOGSHIT compared to a K2 GRID.
Only in double precision.
Quit fronting and admit you don't know jack about hardware.
His hardware choice might make perfect sense, depending on his use case. Perhaps you should lighten up a little.