AMD Dual Fiji videocard surfaces
Click here to post a comment for AMD Dual Fiji videocard surfaces on our message forum
MBTP
Possibly a perfect one card solution for DX12 games.
DX11 will still use as a 4GB memory card. More than 4GB for 4k is a need.
Only HBM archtecture optimized games will run well with 4GB.
schmidtbag
I never said they LITERALLY double VRAM bandwidth. I know that isn't the case. I said they effectively double the bandwidth. Mathematically, it makes no sense how the bandwidth wouldn't be doubled. You have 2 sets of VRAM with loaded with the same resources working in parallel. That's like saying a dual core CPU doesn't double the performance of a single core. If you have a dual-threaded task, both cores are still working toward the same goal but they're not calculating the same thing. But you're still effectively doubling the performance.
rl66
Half true as a lot pass from one GPU to another in the board and not thru the PCIe bus so the real bandwith will be more like x1,something than x2
yes 2 x 4g and not 8g and even more true now that mem are inboard the chip 🙂
About the ghost of botlenecking... there is still range in PCIe 3.0 for more.
80% of people talking about it have the problem in another place of the computer (2g of DDR3 at 800mhz (you know the very 1st one lol), an HDD from 10 years ago at 5400 as main media, a Celeron (OEM version of course to be slowest 🙂 ) with triSLI 780... this last one is the latest i have seen btw...:3eyes:)
---TK---
You should read your prior posts again. Read up on alternate frame rendering.
schmidtbag
First, my prior posts were specifically not referencing AFR. Second, AFR might have less of a performance impact on memory bandwidth but it still greatly improves it, because again - the GPUs are working in parallel and they're not merging their memory. For each individual GPU, the memory bandwidth isn't doubled, but since both sources of memory are being read from simultaneously for rendering for the same application, that's EFFECTIVELY doubling the performance. I don't know how to make it any clearer than that.
I really don't get how you think that 2 sources working in parallel for the same task don't increase performance.
Just to clarify, if the GPUs had access to 8GB of total memory, the performance would not be doubled.
Denial
I mean you're right technically but your initial post basically inferred that having 2x4 (8GB) of HBM would be the reason why the card performs well. Which I guess is what I disagree with. The card is going to perform well because it's 2 Fury's put together. The HBM bandwidth is definitely going to help. But neither having 8GB or double the bandwidth is going to really assist it's performance.
schmidtbag
Fair enough. That was just a speculation on my part. In my personal opinion, I think the improved bandwidth will help the performance, but I accept that I could be wrong about that.
Fox2232
Fury X is bottlenecked by vram speed. From my testing it is 10.3% of performance lost to memory bandwidth.
On other hand it is reason why Nano, while lower clocked, does not lose that much performance. Since going from 1000 to 1050MHz does increase performance only by 3.2%, but this can be easily remediated with vram speed increase.
From information provided by OC experts, they got HBM from 500 to 1000MHz by moving voltage from 1.2 to 1.35V. That's 150mV for such big jump.
even mere 50mV would get us to 650/700MHz since most of Fury(X) start artifacting while running vram @ 580~590MHz.
Denial
How did you measure that?
Fox2232
Downclock to 500MHz on core, and moving back to 1050 in 50MHz steps.
Performance gain is quite linear till GPU reaches 800MHz, then I did same for HBM @550MHz.
This time linearity was quite good till 900MHz on GPU.
With HBM at only 500MHz, performance gain from 950 to 1000MHz and from 1000MHz to 1050 is only ~59.7% that means 40% of given clock increase is lost.
And move from 1050 to 1100MHz lost 54% of performance.
OCing Fiji GPU is pointless without higher HBM clock. So pointless that 1050/500MHz performs same as 1000/570MHz.
And 1100/500MHz is same as 1050/580MHz.
Maybe it is not bandwidth itself what is needed, but some timing/latency stuff.
Mda400
No way thats like a "Duano". :infinity:
nexxusting
Pegged it.
With Relayer saying this we at least realize that considering the Nano exists we can put two of them on a single regular sized dual gpu pcb. I'm not sure of the maths however it seems do able.
*edit, said this before I looked at the picture of the card. Clearly a Dual Nano. 100%.
---TK---
I know how multi gpu works, where did I say you do not get increased performance? I been using multi gpu since 3dfx and nvidia sli since 7800gt series. In afr both gpus have exactly the same data mirrored in vram and each gpu takes turns in rendering alternate frames.
FerCam™
This last generation of GPU's from both brands are way too cheap for me, I'll be waiting for TPP to be approved and put in action for them to get even cheaper...
Koniakki
http://i.imgur.com/ppPJiAH.png
Andrew LB
Hmm... I should buy a few nice graphics cards and have my brother sell them for me in India. He "commutes" from Southern Califormia to Nashik, India every other month while working for a company that develops drugs to fight cancer. Talk about a brutal amount of air travel. lol. each way is two 10 hour legs. Whats funny is he used to complain about flying to Atlanta every 2 weeks. 🙂
I actually just finished reading the summary of the finalized TPP agreement and there was no mention of consumer electronics. It talked about removing tariffs on industrial and agricultural products, as well as textiles and apparel. It said tariffs on cars and trucks will remain. Of course I can't be certain since Obama won't release the full text of the agreement. In fact, the actual text of this agreement won't be released for FOUR years! The only reason they'd do this is because it's another HORRIBLE deal negotiated by idiots (or are simply evil), much like the Iran nuclear deal.
Ryu5uzaku
TPP will be horrible if it comes to be 😀
