It would have been nice to see a 20nm chip. probably could have lowered the power draw anyway but maybe hbm is more efficient and will help out. have not seen anything about the efficiency...
Yeah, but understandable given cost/yields/etc. 20nm would have likely saved 20% on the gpu side. HBM should save about 15w on the memory side (vs the old product, granted this is faster).
640GB/sec, why? Unless AMD is planning to address the "4K problem" by swapping out textures all the time, I don't see any benefit to this, and lots of drawbacks (price being one of them). Considering nVIDIA's GPUs have always been able to match AMD's for performance, while using much narrower bus widths (Hawaii: 512bit, Maxwell: 256bit), I'm not seeing any good reason, unless of course AMD's architecture is far more bandwidth-dependant than nVIDIA's.
Okay, that's a loaded question. The short-ish answer is yes, AMD's arch is more memory dependant because of (typically) greater available floating point/shader resources (amd's units do double-duty for sf as well....where-as nvidia uses less fpu and smaller special function units at a ratio that is often close to fully utilized in many scenarios) plus the fact that around 1/3 of what would be nvidia's similarish required bandwidth is derived from a mixture of greater on-die cache and whatever benefits their newer compression allows. If I had to guess, the split on that is something like 75% cache improvements, 25% compression improvements. IOW, the compression improvements help around 8% to slightly more, just like Tonga for AMD.
The really odd thing about this lineup, is what AMD expects to field in the discrete mobile arena. Presently, the top part is Pitcairn based (M290X) in its third generation of cards. The M295X's (Tonga) heat production in the iMac probably preclude its use in laptops, and Hawaii is clearly unsuitable.
That's where HBM starts. Better too have too much bandwidth than too little.
Hey, someone has to lead the charge. Just imagine the marketing mileage from 640GB/sec. It's like those nutty theoretical fillrates pumped up to eleven!
Fiji will do double duty as a compute chip, where on-card bandwidth will play a much greater role in GPGPU. FWIW, even Nvidia are unlikely to go below 384-bit for their compute chip. The one thing that will hold Fiji back is the 4GB as a FirePro (not the bandwidth). AMD already has the W9100 with 16GB of onboard GDDR5 for a reason.
First, never say never...IIRC nvidia sold big-ass Fermi (granted cut wayyy down and clocked in the basement) in laptops.
Right about HBM...plus, if they shrink/reconfigure the core design on 14/16nm for 50% more resources, the mem controller could probably be reasonably recycled....it's also possible they could cut cache or what-not because of that off-die bw. Not saying they did/will...but who knows? It's possible it's there for more than being ahead-of-it's-time (but a necessary evil) on bw and behind-it's-time on density configuration. Even if they didn't change anything, it should be good for a good extra chunk of performance (double over what is required gives typically around a 16% boost...this in essence could give something like 8% over what one might expect given the other specs and typical core usage).
Either way you look at it, this thing *has* to compete with a GM200 21SMM part. Say that can do 1400mhz best-case, that essentially means this has to do 1200 to compete. The bandwidth required for quite literally 10TF is..well...a lot. You'd be talking needing an 8ghz/512-bit controller which wouldn't exactly be small nor power efficient (if even possible with within die size limits). As odd as it sounds, twice the controllers at (what apparently amounts to 5ghz) is likely both less transistors and more efficient within the gpu logic..
Don't try comparing the arguments over a 970's memory shenanigans to AMD's next uber chip (Fiji). I'm not clued up on it but many say HBM only caters for 4GB memory allowance (for now?...). The 970 is the cheaper Maxwell performance part whereas 390X will be the single gpu top tier.
And yes, those that bought cards with 4GB (or not as the 970 is) would have figured that into their choice. If 390X is to be AMD's next gen top tier card, you would hope it would have more as they have already seen fit (AIB's) to release a 8GB 290X with albeit small rewards at 4k.
IMO, I don't know if we need >4GB for gaming purpose except on poorly coded things (look at the recent CoD for bad memory hogging, or Titanfall IIRC). But if we do need >4GB in the next year or so, I'm pretty sure there will be developments to allow higher memory usage on the AMD chips.
So, to be plain - 4GB won't be an immediate worry and I'm sure it will be addressed when needed.
Correct. HBM is currently 1GB. Implementation, unlike the setup of gddr5, is limited to four chips. That means 4GB. 2nd gen is due end of year. Does that mean a refresh before 14/16nm? Conceivably...but who knows how fast amd is transitioning to the smaller process. I continue to disagree about 4GB being enough...If one was to argue things should be properly coded for 4k/8GB (or possibly 6GB in many instances), we could have a conversation. That said, it's not going to stop HBM memory density from increasing and badly optimized console ports targeted toward that shared pool of memory at a low resolution from being a scaling factor in some regards. I still stand by GM200/R390x for the most part being 1440p-targeted chips (think in terms of a 900p 30fps ps4 game scaled to 1440p/60)...just like GM204 is mostly a 1080p-targeted chip. In those respects, it can be argued 4GB(/6GB in some cases) is sufficient.
