Sidebar Sidebar
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.

AMD Confirms RX 480 At $199 USD, Other APU & Polaris Announcements

ActionableMango said:
Hmmm, AMD building their GPU with a competitor's CPU.

I wonder how the AMD CPU division feels about that. The company meetings must be awkward.
Click to expand...
As this CPU can be comparable in some sense to a higher end FX and better for low multiprocessing, it seems like a temporary Climate Change bundle which makes more sense before Zen if you would also need to buy a motherboard and RAM anyway.
 
koyoot said:
AMD is one company that has two divisions: CPU and GPU. For both responsible, as well as the rest of the company is Lisa Su - CEO of AMD.

The meetings are exactly like their should be done. If both, Intel and AMD can win in this that means that one company loses - Nvidia.
Click to expand...

How were AMD's, Intel's and Nvidia's numbers last quarter?
 
Last edited:
http://semiaccurate.com/2016/12/12/amd-shatters-petaflop-rack-barrier-instinct/

More on the topic of Vega, and Deep Learning.
http://radeon.com/en-us/introducing-radeon-instinct/
http://radeon.com/en-us/radeon-instinct-ecosystem-open-source-platform/

And last link: http://semiaccurate.com/2016/12/12/amd-launches-radeon-instinct-ai-card-lineup/

MI6 and MI8 stands for: Machine Intelligence 6 TFLOPs FP32, and Machine Intelligence 8 TFLOPs FP32. MI25 stands for Machine Intelligence 25 TFLOPs FP16?. Isnt that strange? Either that is the case, or MI25 is Dual Vega GPU.

GP 100 competitor.

687F:C1 is really Vega GPU. My understanding was a bit wrong here. The GPU itself has 8 GB of HBM2. Its outperforming in 4K, in Doom GTX 1080 by 10%, and that is while using Fiji driver, not custom, Vega drivers. http://www.pcgameshardware.de/AMD-Radeon-Grafikkarte-255597/Specials/Vega-10-HBM2-GTX-1080-1215734/
https://www.computerbase.de/2016-12/amd-vega-details/

Quite a lot of information is going on, and to prove my point about MI25 being dal GPU :
http://videocardz.com/63715/amd-vega-and-navi-roadmap
http://videocardz.com/63700/exclusive-first-details-about-amd-vega10-and-vega20
Compare what they have posted and what turned out to be true(nearly...) Nano as MI8, 300W Vega 10 GPU, but dual: it might actually be dual GPU, that MI25.

Vega arch, has 2048 bit memory bus. So only way to get 32 GB's on MI25 is to use dual GPU setup on single board. So after all: MI25 - 25 TFLOPs FP32, 50(!) TFLOPs FP16, 300W TDP, 32 GB HBM2.
 
Last edited:
Vega is looking like the 16 nm shrink of Fiji. The architecture has been iterated, but the stream processors stay the same and now with HBM2 memory. Graphics will probably be a little better than the GTX 1080, but it should be much better at compute.

This should probably end up in the 350mm2 to 400mm2 ballpark for the size of the GPU. Roughly similar to Tahiti and the D700. Lets hope Apple puts these to use in a new mac pro in the not too distant future.
 
koyoot said:
Vega arch, has 2048 bit memory bus. So only way to get 32 GB's on MI25 is to use dual GPU setup on single board. So after all: MI25 - 25 TFLOPs FP32, 50(!) TFLOPs FP16, 300W TDP, 32 GB HBM2.
Click to expand...
Raja Koduri in a response on TT to one of questions about MI25 have confirmed that it is single GPU.
 
koyoot said:
Vega is GP100 competitor.
Click to expand...
And by the time that Vega ships, Nvidia will have Volta on deck.

Nvidia can always check on what ATI is doing by looking in their rear view mirrors.

And, ATI will be very lucky if the first Vega parts are competitive with GP104. Very unlikely that they'll be comparable to (what will then be Nvidia's year old) GP102 - let alone GP100.

And, at that point in time - they'll be competing with GV100/GV102/GV104 - at least in mind share even if the Volta chips are shortly behind Vega.
 
Last edited:
  • Like
Reactions: Mago and tuxon86
AidenShaw said:
And by the time that Vega ships, Nvidia will have Volta on deck.

Nvidia can always check on what ATI is doing by looking in their rear view mirrors.

And, ATI will be very lucky if the first Vega parts are competitive with GP104. Very unlikely that they'll be comparable to (what will then be Nvidia's year old) GP102 - let alone GP100.

And, at that point in time - they'll be competing with GV100/GV102/GV104 - at least in mind share even if the Volta chips are shortly behind Vega.
Click to expand...
First, make sure that Nvidia delivers the hardware they advertised. 32 GB version of the GPU was advertised as Q1 2017.

Secondly, make sure that Volta on 14/16 nm process is true Volta, and not reused GP100 architecture. Which to which, everything we know about Volta points at this point.

Back to topic: There are two GPUs with 687 DeviceID: 687F:C1, and 687F:C3.
 
I have something for people interested in on chip engineering:
http://www.eetimes.com/document.asp?doc_id=1330981&page_number=2
MD decided to create a superset of Hypertransport to replace separate on-chip interconnects that it used for its CPUs and GPUs. The resulting Infinity fabric debuts in Summit Ridge before April and in AMD’s Vega GPUs before July.

The company declined to give data rates or latency figures for Infinity, which comes only in a coherent version. However, it said that it is modular and will scale from 30- to 50-GBytes/second versions for notebooks to 512 Gbytes/s and beyond for Vega.

AMD does not plan to license the link, which uses the Hypertransport messaging protocol. Instead, it will use Infinity both as a network-on-chip and as a clustering link between its GPUs and x86 server SoCs. It supports the open CCIX standard as a link to third-party accelerators such as FPGAs.
Click to expand...
Its HSA hardware implementation. Will be very good addition for Mac Pro, and software designed with HSA in mind.

It appears that in hardware design, at least for this moment, AMD may be 1-2 years in front of Intel and Nvidia in terms of coherency for their hardware. When you read stuff like that you start wondering, what would be possible if big software companies, would care about initiatives like HSA.

Just on paper, it looks absolutely amazing. Ability to deliver the data where its needed, when its needed, and after that immediate idling out the hardware resources, is really tempting for future of efficiency.
 
ManuelGomes said:
nVidia confirmed 1080Ti coming soon, I guess Vega might have forced them to put it out the door.
980Ti owners get first silicon.
Some OCers are taking the 1060 past 3GHz - impressive!!
Click to expand...
I really doubt that ATI does any significant driving of Nvidia's product cycles - especially "forcing" Nvidia to bring something to market.

Nvidia so firmly owns the top end of the GPU space that they have the luxury of timing releases for tactical value. The best thing about the GTX 980ti was that it was close to the performance of the Titan X (M), but much cheaper.

So right now, ATI doesn't have anything for the GTX 1080 space (or really even for the 1070 space). Nvidia sits on the 1080ti silicon, and sells Titan X (P) at a premium. Meanwhile, the fabs are cranking out chips that have plenty of working CUDA cores to be an 1080ti, but not enough for a full fledged Titan X (P).

When (if) Vega ships with performance in the 1070/1080 range, Nvidia can start shipping 1080ti that are faster than the 1080, and much cheaper than the Titan X (P). All based on chips with too many defects to be Titan X (P), but are fine as 1080ti. (Check your MBA textbooks for "sunk costs".)

Nvidia won't be "forced" to respond to Vega, they're laying the groundwork for a response that maximally disrupts ATI's Vega rollout. (And they're likely building inventory so that 1080ti cards will be everywhere just at the time that Vegas are in early stage supply restriction, and Nvidia will be able to price them for maximal disruption.)

And this assumes that Vega is actually viable. HBM is a risk factor, and if it limits Vega to a paltry 8 GiB of VRAM it could sink Vega. 4 GiB would be fatal.
 
  • Like
Reactions: tuxon86 and Flint Ironstag
koyoot said:
To counter Vega, Nvidia will have to reuse GP100 architecture/chip in consumer market.

Also, because AMD will refresh the Polaris lineup, Nvidia will do the same thing with Pascal. So hold your horses, Aiden.

Manuel, you brought Nvidia to AMD thread. GTX 1080 thread is here: https://forums.macrumors.com/threads/nvidia-gtx-1080.1971528/page-21

On the first page of this subforum ;)
Click to expand...
Is Vega due in 2017 or 2018 - or will it slip into 2019?

And I'm not sure what you mean by "refresh" in the context of Nvidia. Nvidia didn't "refresh Maxwell", they brought out Pascal. Are you setting the stage to dismiss Volta as a "Pascal refresh" - perhaps even before the often-delayed Vega is actually delivered?
 
Last edited:
AidenShaw said:
Is Vega due in 2017 or 2018?

And I'm not sure what you mean by "refresh" in the context of Nvidia. Nvidia didn't "refresh Maxwell", they brought out Pascal. Are you setting the stage to dismiss Volta as a "Pascal refresh"?
Click to expand...
2017. Pascal refresh means that in 2017 we will see 20XX GPUs using the same Pascal architecture. Volta most likely will come out in 2018. Also, it will be done by reusing GP100 chip, and uarchitecture in the mainstream. Can we move this discussion to Pascal thread?
 
koyoot said:
2017. Pascal refresh means that in 2017 we will see 20XX GPUs using the same Pascal architecture. Volta most likely will come out in 2018. Also, it will be done by reusing GP100 chip, and uarchitecture in the mainstream. Can we move this discussion to Pascal thread?
Click to expand...
Why, does a discussion of Pascal threaten your precious thread about the imaginary Vega?

It's really absurd to insist that comparative discussions of ATI and Nvidia architectures and implementations have to be in separate threads.
 
  • Like
Reactions: tuxon86
AidenShaw said:
Why, does a discussion of Pascal threaten your precious thread about the imaginary Vega?

It's really absurd to insist that comparative discussions of ATI and Nvidia architectures and implementations have to be in separate threads.
Click to expand...
Information about Pascal refresh has nothing to do about comparing AMD and Nvidia. Secondly, this is not my thread. Thirdly you are threadcrapping, by posting comparison of technology, between two brands, which is rendered moot, because is based on non-actual information. It is a possibility, but not in the face of what we know about what is Nvidia and AMD doing.

We all know that you hate AMD, but at least you could do better, with spreading FUD. For example by moving Pascal discussion to Pascal thread on this forum.
 
Polaris 12 seems to be real.
I wonder what the suffix V means. P (performance?) for Fiji/10, M (mobile?) for 11, V for 12, value?

10.12.3 beta 2 out for developers. Any Polaris 12 reference?
 
Last edited:
You must log in or register to reply here.
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.
Top Bottom
Back