Mr. Smartass, please explain to us all, how can MI25 GPU can have 32 GB of HBM on 2048 bit memory controller.Yes, I can see the lineup of excuses is long. Not the right drivers, its really a dual card for 300W and so on and on
Vega performs as expected when you look at Polaris.
Single GPU - yes, that is correct. But naming scheme, and technicalities are not in line. MI25 should stand for Machine Learning 25 TFLOPs of FP32.
So how you will the explain potential of 32 GB of HBM?
http://semiaccurate.com/2016/12/12/amd-launches-radeon-instinct-ai-card-lineup/
If that is the case, then it has to be dual GPU setup.
32 GB on 2048 bit memory bus. Please try to explain this. Vega has only 2048 bit memory controller.SK hynix also have 8 stack HBM2 in their Product list, so this could be a possibility.
"But first things first. What does this name even mean? If you look closely those MI numbers are equal to computing power (5.7 TFLOPs = 6, 8.2 TFLOPs = 8). But wait, there’s a catch! MI25 does not have 25 TFLOPs of FP32 computing power. MI25’s value comes from FP16 calculations. So it means that the card has 12.5 TFLOPs in 32-bit operations"It's FP16 performance, not FP32.
With Vega having the possibility to pack FP16 this explains the 25 TFLOPs.
12.5 tflops for big Vega isn't bad, but it's really late.
"As AMD’s sole training card, the MI25 will be going up against NVIDIA’s flagship accelerator, the Tesla P100."
http://www.anandtech.com/show/10905/amd-announces-radeon-instinct-deep-learning-2017/2
WOW. Anandtech sure popped this AMD preview up fast!
"But first things first. What does this name even mean? If you look closely those MI numbers are equal to computing power (5.7 TFLOPs = 6, 8.2 TFLOPs = 8). But wait, there’s a catch! MI25 does not have 25 TFLOPs of FP32 computing power. MI25’s value comes from FP16 calculations. So it means that the card has 12.5 TFLOPs in 32-bit operations"
http://videocardz.com/64677/amd-ann...erator-radeon-instinct-mi25-for-deep-learning
32 GB on 2048 bit memory bus. Please try to explain this. Vega has only 2048 bit memory controller.
Nope.This is just 7970 successor.It will compete with GP104.Probably 300-350mm2 SKU.I am not sure when AMD will release 290x successor/GP102 competitor.That is big vega
Mr. Smartass, please explain to us all, how can MI25 GPU can have 32 GB of HBM on 2048 bit memory controller.
THEN we can talk about excuses. Your excuses.
Google translate said:AMD, the specialist press has given a first look at the graphic architecture of Vega. We can not report on this much at this point, but we already know about our experiences in a demo room. There were two systems with Vega graphics cards, on one ran doom with a consumer Vega. The expanded statistics in Doom showed 8 GiByte HBM2 memories.
VideoCardz said:Device ID 687F:C1 confirmed as VEGA, 8GB RAM
One more thing, it appears that 687F:C1 device that was leaked just a few days ago through AOTS benchmark website is indeed VEGA, as confirmed by PCGamesHardware.
More importantly, we learn that Radeon VEGA has 8GB of RAM.
So tiny bit faster than a 6+ months old GTX 1080 in an AMD demo with Doom Vulkan, which means best case scenario. From the looks of it Vega is late to party, might be using more power than GP104 for similar performance, probably costs more to produce than the NVIDIA equivalent (interposer, HBM2, bigger die?) and has little to no chance against Titan X or a hypothetical full GP102.
So tiny bit faster than a 6+ months old GTX 1080 in an AMD demo with Doom Vulkan, which means best case scenario. From the looks of it Vega is late to party, might be using more power than GP104 for similar performance, probably costs more to produce than the NVIDIA equivalent (interposer, HBM2, bigger die?) and has little to no chance against Titan X or a hypothetical full GP102. I wonder if NVIDIA will bother to launch revised GP104 to match this performance at 150-180W using cheaper GDDR5X or just go straight to Volta.
Please read what you post.
They talk about 24 TFLOPs for a single V10 @ 225W ... by this logic MI25 can not be a dual gpu card.
http://www.pcgameshardware.de/AMD-Radeon-Grafikkarte-255597/Specials/Vega-10-HBM2-GTX-1080-1215734/
http://videocardz.com/64706/amd-vega-doom-4kultra-gaming-performance-demo-possible-specs
Maybe they were only showcasing small Vega in the Doom demo-room ? Which would be the perfect contender vs GTX 1080.. Afterall we already saw the 687F:C1 numbers matching 1080 in the Ashes of the Singularity leak.
http://videocardz.com/64475/are-those-radeon-rx-490-vega10-benchmarks-results
Pretty sure we haven't seen big Vega gaming numbers yet
Sure it can. Hasn't AMD lowered the clocks on dual GPU cards in the past to keep power consumption down?
So tiny bit faster than a 6+ months old GTX 1080 in an AMD demo with Doom Vulkan, which means best case scenario. From the looks of it Vega is late to party, might be using more power than GP104 for similar performance, probably costs more to produce than the NVIDIA equivalent (interposer, HBM2, bigger die?) and has little to no chance against Titan X or a hypothetical full GP102. I wonder if NVIDIA will bother to launch revised GP104 to match this performance at 150-180W using cheaper GDDR5X or just go straight to Volta.
Hmmm ....a card in a poorly ventilated case [throttling] and with Fiji drivers 'just ' beats the 1080 by 10% ...jeeze AMD should just give up now lol!
Volta is, best case, 12 months away.
Yes, I can see the lineup of excuses is long. Not the right drivers, its really a dual card for 300W and so on and on
Vega performs as expected when you look at Polaris.
Not really worst case either. Certainly unoptimised drivers (Fiji drivers if to believe), and closed case for possible throttling. But also 4K which is to be one of the best scenarios to show AMD product. If they were to show 1080p under DX11, we might see very different behaviour (yes talking about AMD's driver overhead once again).
Anyways, looks quite decent.
It seems, if the translation is good that its a Vega 10 with 8 gb of memory.
I suspect as they say , "its consumer based Vega 10".
A 8gb cut down version would be Vega 11 .and will compete with a gtx1070.
That's my guess.
The case looks to be quite good.
http://www.corsair.com/en/carbide-series-air-240-high-airflow-micro-atx-and-mini-itx-pc-case
Note, however, that first, no Vega-optimized driver was used, but simply a Fiji driver with a little additional debugging work. Second, any view on the graphics card should be denied, so that all the fan slots were taped and there was no airflow in the housing. The graphics card was noticeably noisy and dusted neat hot exhaust air backwards.
The case looks to be quite good.
http://www.corsair.com/en/carbide-series-air-240-high-airflow-micro-atx-and-mini-itx-pc-case
Second, any view on the graphics card should be denied, so that all the fan slots were taped and there was no airflow in the housing.