AMD Polaris 10 GPU Reportedly Offers Near 980 Ti Performance For 300 USD

[Notional]

Just now, patrickjp93 said:

Shh... the plebes still don't understand it's all about the cache anyway. It may come to matter eventually, but games are not so complex currently (nor will be for a few years to come) that they actually require 512GB/s of bandwidth from VRAM, much less 1 TB/s. 

Yes just look how much better the XBone with it's DDR3 and large on die cache compared to PS4's GDDR5 in minimal cache. XBone is clearly better /s.

[patrickjp93]

19 minutes ago, Notional said:

Yes just look how much better the XBone with it's DDR3 and large on die cache compared to PS4's GDDR5 in minimal cache. XBone is clearly better /s.

Fewer shaders, TMUs, and ROPs, and far fewer ACEs, as well as a lower clock speed. You have to have balance in design, but performance in any computing environment is 90% determined by how well you use your cache. Don't believe me?

 

 

 

The same holds true for the GPU. The underlying paradigm of modern game engine design is based around optimizing for the best use of your data. Namely, it's all about caches. And mind you these guys only cover the basics. If you're going to main memory directly (not via a natural prefetch which the CPU caches handle in predictable access patterns to get the data and instructions before you need them), you've already lost. If you don't use your cache optimally, nothing else matters. And mind you, not all caches are created equally. Intel's cache structure is vastly superior to AMD's because it is fully associative. If you miss data/instructions in the L1 cache of a core in AMD's design and then can't find them in L2 either, you have to "snoop" the L1 of the other cores to check if it is still nearby before you can start checking L3 in the case of FX or going to main memory for the Athlons and APUs. In Intel's case, if it's not in a given core's L1 or L2, you know to prefetch from memory while searching the L3 cache. In the event of the L3 miss, you've already mitigated a lot of the damage by virtue of the design.

It is ALL about the caches, always, without exception.

[AyeJayKay]

Might switch to the red team if they can offer a competitive product. I'm still looking forward to the new GTX1070 or whatever its going to be called.

[Notional]

12 minutes ago, patrickjp93 said:

Fewer shaders, TMUs, and ROPs, and far fewer ACEs, as well as a lower clock speed. You have to have balance in design, but performance in any computing environment is 90% determined by how well you use your cache. Don't believe me?

 

 

 

It is ALL about the caches, always, without exception.

So it's not ALL about the cache. Wasting all that die space on on-die cache instead of more shaders, ROP's, ACE's TMU's was a mistake, that made the graphics not only slower, but less performing:

 

 

Playstation 4 on the left is smaller than the XBone chip on the right. All that space wasted on SRAM, which is super fast cache, did not yield better performance compared to the Playstation 4 chip.

All else equal than just SRAM, then sure it would be faster, but you always end up wasting a LOT of valuable die area that could be used much more efficiently for performance. Of course things have to be properly balanced.

[VagabondWraith]

I'm wondering... if nVidia does NOT support Async with Pascal, if it would be a wise decision for me to head to Team Red and pick up an AMD card. We've seen the benefits of Async, and more performance is always welcome. I could pick up a 490X, then get Vega when it's ready.

 

Hmmmm......

[That Norwegian Guy]

21 hours ago, ProKeero said:

Pretty sure AMD aint gonna win this time either xD, remember what they said about the 300 series? how it was gonna crush Nvidia? that was a great fantasy...

You mean the 300 series that is currently dominating every price point except the most expensive one? Yeah, about that....

[patrickjp93]

49 minutes ago, Notional said:

So it's not ALL about the cache. Wasting all that die space on on-die cache instead of more shaders, ROP's, ACE's TMU's was a mistake, that made the graphics not only slower, but less performing:

 

Playstation 4 on the left is smaller than the XBone chip on the right. All that space wasted on SRAM, which is super fast cache, did not yield better performance compared to the Playstation 4 chip.

All else equal than just SRAM, then sure it would be faster, but you always end up wasting a LOT of valuable die area that could be used much more efficiently for performance. Of course things have to be properly balanced.

From a software standpoint, it's all about the cache. You will saturate your hardware's capabilities very easily just by optimizing cache use.

[patrickjp93]

41 minutes ago, VagabondWraith said:

I'm wondering... if nVidia does NOT support Async with Pascal, if it would be a wise decision for me to head to Team Red and pick up an AMD card. We've seen the benefits of Async, and more performance is always welcome. I could pick up a 490X, then get Vega when it's ready.

 

Hmmmm......

The average 7% gain that Nvidia managed to surpass in AOTS without having Async? Okay...

[Agosto]

2 hours ago, patrickjp93 said:

And 40% higher clocks*

 

1400MHz boost clock is massive.

Well, the average Titan X (full GM200) boost clock is around 1150-1200 MHz, so that's not 40% higher clockspeed...

If NVIDIA listed the GP100 as a 300W part, which is 50W higher than their current highest end cards, this could mean they had some minor issue with the 16nm process or the integration of FP64 units. At least I hope so since I'd like to see AMD to ~2010 market share levels... and latest NVIDIA drivers are quite disappointing.

[Scitesh]

For the fun of concept of competition let there be a ideal world where4/5 companies are fighting it out for GPU market that would lead to cut throat competition as a result the refresh cycle will shrink down and we will see strategies from them as we see from android based smartphone companies 

(Samsung launches S7 priced 54000INR challenged by xiomi mi5 at 24000 and it gets challenged by micromax canvas 6 at 13000) so each price point is covered and we don't get these big gaps b/w each iteration

[Prysin]

On 25/04/2016 at 2:20 PM, IMPERIUS said:

I hope thats true

 

edit: the 300- series was also apparently going to "destroy the 900- series from nvidia" tho that didn't really happen...

the 300 series was going to compete with. Not a single word of "destruction" or "crushing" or "drastically outperforming" was ever said by AMD. So take your shit elsewhere.

[patrickjp93]

3 hours ago, Agost said:

Well, the average Titan X (full GM200) boost clock is around 1150-1200 MHz, so that's not 40% higher clockspeed...

If NVIDIA listed the GP100 as a 300W part, which is 50W higher than their current highest end cards, this could mean they had some minor issue with the 16nm process or the integration of FP64 units. At least I hope so since I'd like to see AMD to ~2010 market share levels... and latest NVIDIA drivers are quite disappointing.

No, the rated boost clock of the Titan X is much lower. Now, what GPU Boost 2.0 may achieve on average is a separate matter.

[MrRavens]

7 hours ago, patrickjp93 said:

Shh... the plebes still don't understand it's all about the cache anyway. It may come to matter eventually, but games are not so complex currently (nor will be for a few years to come) that they actually require 512GB/s of bandwidth from VRAM, much less 1 TB/s. 

Lmfao, you did not just call them plebs!

[patrickjp93]

49 minutes ago, LtRavens said:

Lmfao, you did not just call them plebs!

I did.

[Energycore]

4 hours ago, patrickjp93 said:

I did.

Wow. So nobility. Many demeaning.

[patrickjp93]

4 minutes ago, Energycore said:

Wow. So nobility. Many demeaning.

Much right. No apologies. Such suckiness.

[Belgarathian]

On 4/25/2016 at 11:43 AM, Clanscorpia said:

AMD is not fine. Their stock is 4 dollars. THATS SHIT. They need money. And they dont care about their old card. If their best card is better, they can always underclock to make it cheaper. It's a very, very good strategy

Nothing wrong with $4 shares, it's about the market cap (price of shares * number of shares). 

 

For instance, a company with 1 x $1000 share has the same value as a company with 1000 x $1 shares.

[vanished]

I just can't imagine that will be true, but if it is... I might just try radeon again...

[MrDynamicMan]

13 hours ago, That Norwegian Guy said:

You mean the 300 series that is currently dominating every price point except the most expensive one? Yeah, about that....

Shhh, they cant see past the green.

[Fetzie]

1 hour ago, Ryan_Vickers said:

I just can't imagine that will be true, but if it is... I might just try radeon again...

If you are planning to use the graphics card for more than 2 years you'd probably be better off with amd anyway My HD 7950 is still getting performance increases from drivers 5 years after it was released.

[M4st4M1nd]

Looks very promising. I look forward to switch from Nvidia/Intel to AMD

[IMPERIUS]

13 hours ago, Prysin said:

the 300 series was going to compete with. Not a single word of "destruction" or "crushing" or "drastically outperforming" was ever said by AMD. So take your shit elsewhere.

Didnt say that AMD said those things

[79wjd]

5 hours ago, Fetzie said:

If you are planning to use the graphics card for more than 2 years you'd probably be better off with amd anyway My HD 7950 is still getting performance increases from drivers 5 years after it was released.

It helps that the 7950 was in AMDs current lineup until 11 months ago.

[Deedot78]

On another thought, if the graph graph is true, its gonna devalue ALL of amd cards present series cards on market, time to get my cards sold before the flood of AMD and Nvida cards come onto ebay with this release. 

[Fetzie]

2 hours ago, Deedot78 said:

On another thought, if the graph graph is true, its gonna devalue ALL of amd cards present series cards on market, time to get my cards sold before the flood of AMD and Nvida cards come onto ebay with this release. 

If it is true it'll devalue all of the cards on the market.