Dual Fiji Card May Finally Be Here Soon

[littleg]

384bit bus and GDDR5X is plenty to beat 4096bit HBM1. 672GB/sec vs 512GB/sec for example.

HBM needs to drastically get its static cost down. Else it can end up even losing to HMC in the long run.

How much VRAM will you need to get a 384 bit bus?

 

[ShintaiDK]

How much VRAM will you need to get a 384 bit bus?

Depends on the densities you pick. Anywhere from 3 to 24GB. Perhaps soon up to 48GB.

 

[ShintaiDK]

Intel's propriety ram-tech last time didn't end too well..

What memory technology was that? And no, RDRAM is the wrong answer.

This factor is why I strongly believe next-gen high-end is HBM2 only (they will certainly need it for Teslas and Firepros to out compete Intel's Phi) as such, it's launch will coincide with how well HBM2 shapes up.

HBM2 needs to deliver first, and with the rumours going on it looks bad, really bad. HBM1 is a mere shadow of what was promised. And it backfired so hard for AMD it made Fiji useless for any HPC/Pro segment where cost is close to irrelevant unlike the gaming segment. Instead AMD sells Hawaii GPUs with GDDR5 for that segment.

 

[littleg]

Depends on the densities you pick. Anywhere from 3 to 24GB. Perhaps soon up to 48GB.

Possibly 'modules' would have been a better term. You'd need 12 modules for a 384 bit bus. Standard 512MB modules gets you 6GB like the 980ti.

More modules, more power, more cost.

 

[littleg]

HBM2 needs to deliver first, and with the rumours going on it looks bad, really bad. HBM1 is a mere shadow of what was promised. And it backfired so hard for AMD it made Fiji useless for any HPC/Pro segment where cost is close to irrelevant unlike the gaming segment. Instead AMD sells Hawaii GPUs with GDDR5 for that segment.*

*Citation needed

 

[ShintaiDK]

*Citation needed

Really? You cant be bothered to visit AMDs site and check the FirePro cards?
http://www.amd.com/en-us/products/graphics/server/s9170#
http://www.amd.com/en-us/products/graphics/workstation/firepro-3d#
http://www.amd.com/en-us/press-releases/Pages/amd-delivers-worlds-2015jul08.aspx

4GB limit is a joke there. Hence no Fiji.

Possibly 'modules' would have been a better term. You'd need 12 modules for a 384 bit bus. Standard 512MB modules gets you 6GB like the 980ti.

More modules, more power, more cost.

Its not a problem for 980TI or TitanX. Nor is it an issue for AMDs 32GB FirePro card or Telsa/Quadro cards.

 

[ShintaiDK]

I always applaud either side for pushing the envelope so Kudos to AMD for their work in releasing this soon.
However, I sure hope the HDMI 2.0 issue is addressed since of prospective purchaser of this is surely thinking a 4k Monitor. I thought I read that the total memory is 8G?

AMD already started their road to HDMI 2.0 is via a DP to HDMI active adapter.

Memory is 2x4GB. So its 4GB per GPU. 4GB in late 2015 or 2016 for a high end dual GPU card is just silly. Everything points to an inflation in VRAM usage. We already see lower end cards having 4GB now. like 380/960. The 2GB 960 is even discontinued.

And we already know AMD needs 2 dedicated engineers to manage VRAM usage on Fiji. What happens when these 2 people gets relocated to a newer GPU or something else. I think we all know.

 

[littleg]

Really? You cant be bothered to visit AMDs site and check the FirePro cards?
http://www.amd.com/en-us/products/graphics/server/s9170#
http://www.amd.com/en-us/products/graphics/workstation/firepro-3d#
http://www.amd.com/en-us/press-releases/Pages/amd-delivers-worlds-2015jul08.aspx

4GB limit is a joke there. Hence no Fiji.

They're all pre-fiji. AMD released the fury line to generate some cash (imo) and to at least have something out there that could compete. The high end workstation cards will come in the next 9 months or so, fiji based first, die shrunk at the back end of 2016.

Its not a problem for 980TI or TitanX. Nor is it an issue for AMDs 32GB FirePro card or Telsa/Quadro cards.

Yeah, 980ti or TitanX. Great examples of low cost, low power GPUs

 

[ShintaiDK]

They're all pre-fiji. AMD released the fury line to generate some cash (imo) and to at least have something out there that could compete. The high end workstation cards will come in the next 9 months or so, fiji based first, die shrunk at the back end of 2016.

Citation please.

Yeah, 980ti or TitanX. Great examples of low cost, low power GPUs

Like Fiji is any different.

 

[Erenhardt]

Really? You cant be bothered to visit AMDs site and check the FirePro cards?
http://www.amd.com/en-us/products/graphics/server/s9170#
http://www.amd.com/en-us/products/graphics/workstation/firepro-3d#
http://www.amd.com/en-us/press-releases/Pages/amd-delivers-worlds-2015jul08.aspx

4GB limit is a joke there. Hence no Fiji.



Its not a problem for 980TI or TitanX. Nor is it an issue for AMDs 32GB FirePro card or Telsa/Quadro cards.

Or FP64 performance worse than 7970 is the key... VRAM can be increased easily. But why release more expensive card that is slower than the one you are already selling?