AMD Vega Preview: It's About Efficiency

[AliG]

Per PC World:

AMD crafted Vega?s high-bandwidth memory architecture to help propel memory design forward in a world where sheer graphics performance keeps improving by leaps and bounds, but memory capacities and capabilities have remained relatively static. The HB cache replaces the graphics card?s traditional frame buffer, while the HB cache controller provides fine-grained control over data and supports a whopping 512 terabytes?not gigabytes, terabytes?of virtual address space. Vega's HBM design can expand graphics memory beyond onboard RAM to a more heterogeneous memory system capable of managing several memory sources at once.

vega 512tb virtual address
AMD
That?s likely to make its biggest impact in professional applications, such as the new Radeon Instinct lineup or the cutting-edge Radeon Pro SSG card that graft high-capacity NAND memory directly to its graphics processor. ?This will allow us to connect terabytes of memory to the GPU,? David Watters, AMD?s head of Industry Alliances, told PCWorld when the Radeon Pro SSG was revealed, and this new cache and controller architecture designed for HBM?s blazing-fast speeds should supercharge those capabilities even more.

To drive the potential benefits home, AMD revealed a photorealistic recreation of Macri?s home living room. The 600GB scene normally takes hours to render, but the combination of Vega?s prowess and the new HBM2 architecture pumps it out in mere minutes. AMD even allowed journalists to move the camera around the room in real-time, albeit somewhat sluggishly. It was an eye-opening demo.

You guys already know it hits 60 FPS in Doom 4k Ultra, but the 512TB of addressable memory and twice as efficient pipeline is very intriguing. This could actually bite into Nvidia's deep learning market.

[SinOfLiberty]

~10%+/- to 1080.

[AliG]

~10%+/- to 1080.

That seems about right, but I don't think that was AMD's main point today. Enthusiast graphics is a nice market, but server computation is what recently blew up Nvidia's profits.

All that addressable memory and the new rendering heuristics will be huge in the corporate world.

[SinOfLiberty]

In General gaming.

It`s been known for a while lol.

Fail. But can OC to ~1080ti per.

[Tim]

Wow, how much longer do we have to wait, seriously. I'm not liking this very much. nVidia is just waiting and waiting and waiting, ready with the goods. GP102 Ti will most likely be good enough to counter the Vega. I'd be surprised if Volta wasn't in the end stages already, they've had ages without competition.

[xlink]

This is the right direction to head.
You can get a "good enough" memory size from HBM2 implementations and when memristors/similar-tech come out you're going to get DDR3 level memory performance for NAND like prices. Imagine a graphics card with 16GB HBM cache and 512GB

[zanzabar]

can amd hurry up and put out a high end card so NV will do the same and put out a non titan high end card. samsung and hynx are supposed to finally be shipping bulk hmb2 in february so i guess we will see it in march or april?

~10%+/- to 1080.

it should be better than that, the fury X is already close to that with the fixed drivers and capped tessellation, it is thermal limited so the die shrink should get it much better clocking so i would expect it to be closer to the titan x pascal than the 1080 assuming amd has 4096 shaders @ 1.5-1.7ghz.

Wow, how much longer do we have to wait, seriously. I'm not liking this very much. nVidia is just waiting and waiting and waiting, ready with the goods. GP102 Ti will most likely be good enough to counter the Vega. I'd be surprised if Volta wasn't in the end stages already, they've had ages without competition.

i would be willing to be volta is designed for HMB2 only or with an HMB2 cashe, since it is not shipping for a couple months and should be very costly i dont see them doing it until micron starts production of the cheap stuff.

[SinOfLiberty]

It is gonna come out with such per numbers, want it or not.

[SinOfLiberty]

Besides, some chinese sources who attended closed amd conference in China mainland, did say Vega can reach 1.7ghz at which it can easily beat 1080. Which is I have been reporting much earlier.
http://tieba.baidu.com/p/4929207543

Uno problema: will need way more than 400w

[iddqd]

my kilowatt power supply is going to be useful again???

[Tim]

I should think so that it beats GP104!! LOL, that was nvidia just putting insult to injury IMO!


400W? I assume they mean total system wattage with that?

[SinOfLiberty]

never heard of custom bios?...

[AliG]

Besides, some chinese sources who attended closed amd conference in China mainland, did say Vega can reach 1.7ghz at which it can easily beat 1080. Which is I have been reporting much earlier.
http://tieba.baidu.com/p/4929207543

Uno problema: will need way more than 400w

Even the original Fury X would handily beat a stock GTX 1080 at 1.7 GHz.

It's great if Vega can hit that, but I don't think it's necessary. Given Vega has the same shader count and is supposedly twice as a efficient per shader (lets call it realistically being closer to 1.5X in games), even a base clock of 1.3 GHz would be enough to compete with stock GP102.

[zalbard]

The level of performance of GTX 1080 would be disappointing.

Vega will be released 9+ months later than GTX 1080, which isn't even Nvidia's fastest card, with the bandwidth of only 320 GB/s.

[SinOfLiberty]

it is overclockable...

[zanzabar]

The level of performance of GTX 1080 would be disappointing.

Vega will be released 9+ months later than GTX 1080, which isn't even Nvidia's fastest card, with the bandwidth of only 320 GB/s.

at stock clock speeds the furyX and 980ti handily beat the 1080. dont forget that the actual clock speed is 1607mhz for the 1080 but they all run at 2050-2100mhz with the stock overclock running in the background just long enough to bench it for 30min.


vega also has a new ROP that no longer requires it to dump needed tasks to the cashe before getting a new one. that should really help out considering how amd has a ridiculous shader to rop power ratio. one of the big advantages NV has is that they have a more balanced shader to rop ratio.

[AliG]

vega also has a new ROP that no longer requires it to dump needed tasks to the cashe before getting a new one. that should really help out considering how amd has a ridiculous shader to rop power ratio. one of the big advantages NV has is that they have a more balanced shader to rop ratio.

I'm wondering how much of the Vega benefits come from simply fixing the ROP bottleneck. Their cards have lately performed horribly (except for DX12) with respect to how much raw processing power they have. Removing the backend bottleneck via cache tricks should help them perform similarly to a balanced design in synchronous compute engines.