AMD goes fermi

[Cerb]

I have to say that was not at all what I expected. I thought we would see nVidia backpedal on fermi rather then AMD embrace it.

First, NV would have no reason to backpedal on Fermi. GF100 sucked, but the rest have been alright. Fermi is also their 2nd generation of compute-first GPUs.

Second, GCN had rumors for awhile, and months ago, they were confirmed and expounded upon. The rumor then was the 7xxx series would be a GCN/VLIW4 mix. So, it should have been expected.

Third, they would need to have been working on this for years before Fermi came out. Fermi will have influenced only fairly minor decisions. IMO, it is likely they began work on what became GCN immediately after buying ATi.

Fourth, they've been talking about GPGPU since around the time they bought ATi.

Fifth, it's hardly a slow poke, and the the power use ain't half bad.

C. Games and consumer software is going to, very very soon, render non GPGPU focused video cards obsolete (has been nVidia's view from day one and it never materialized)

This is going on. Very soon is somewhat vague, but it needs to be after cheap hardware is in consumer hands. Run of the mill consumer apps are now using GPGPU, game engines are more complex, and so on. By the time it has happened, you will hardly have noticed, because it will have occurred gradually.

 

[SirPauly]

I didn't think it was at all inflammatory and flaming was not my intention.

Also it is more accurate because they didn't just focus on GPGPU, but did so in a way that looks to be the same as what fermi did to me. Unless I am misunderstanding.



Backpedal as in, backpedal for gaming while keeping fermi around for a separate compute line. There is more then enough money in each market to justify that.
And because the architecture makes significant sacrifices in the gaming arena which has cost them.

What are these significant sacrifices? The biggest leap for gaming may be because of GPU Processing from more dynamics to RayTracing.

 

[n0x1ous]

That "absolute minimal boost" in performance was enough for it to smash NVDA's flagship GPU,yet it runs fairly cool,uses a far smaller die and doesn't use a lot of power doing so....other than that..its "just like Fermi"....

I wouldn't call an average of 20-25% smashing. Comparing this to the jump the 5870 made over the 280/4870, this is much less impressive. Kepler will easily regain performance leadership.

same old same old. Those who care about efficiency (AMD) and those that care about absolute performance (movies)

 

[animekenji]

http://www.anandtech.com/show/5261/amd-radeon-hd-7970-review

So, I am looking at the anantech review of the AMDs new graphics architecture and all I can see is fermi.
It outright says that this is not as good for gaming, but better for compute. And so for necessity they are going that way, providing absolute minimal boost in gaming performance over current gen in graphics due to dumping VLIW4 for SIMD.

I have to say that was not at all what I expected. I thought we would see nVidia backpedal on fermi rather then AMD embrace it.

Bitcoin miners who prefer AMD cards over nVidia ones would dispute the idea that nVidia cards are better at GPU computing than AMD cards. In fact, they have proven that AMD cards have more processing power than equivalent nVidia cards. If AMD is changing their architecture to be closer to that of nVidia, they will be making a serious error.

 

[Vesku]

AMD VLIW architecture does well at hashing which is what bitcoin is about. It looks like AMD's GCN architecture has retained most of the hashing ability while gaining a lot of GPGPU ability in other areas.

 

[3DVagabond]

Bitcoin miners who prefer AMD cards over nVidia ones would dispute the idea that nVidia cards are better at GPU computing than AMD cards. In fact, they have proven that AMD cards have more processing power than equivalent nVidia cards. If AMD is changing their architecture to be closer to that of nVidia, they will be making a serious error.

No issues with bitcoin mining. GCN is nothing like Fermi.

 

[Will Robinson]

I wouldn't call an average of 20-25% smashing. Comparing this to the jump the 5870 made over the 280/4870, this is much less impressive. Kepler will easily regain performance leadership.

same old same old. Those who care about efficiency (AMD) and those that care about absolute performance (movies)

Where is this Kepler of which you speak?
You mean some unreleased project that may or may not be faster than HD7970..if and when it finally arrives?
Perhaps waiting to see how awesome it is before stating how easily it will beat AMD's latest GPU would be advisable...

 

[taltamir]

Bitcoin miners who prefer AMD cards over nVidia ones would dispute the idea that nVidia cards are better at GPU computing than AMD cards. In fact, they have proven that AMD cards have more processing power than equivalent nVidia cards. If AMD is changing their architecture to be closer to that of nVidia, they will be making a serious error.

Bitcoin mining is a very specific type of compute that has similar parallelism as video games.
Better for computer, in general, does not mean better for every compute task out there.

And there is no "if", they already did it. It was soft launched already, reviewers have the cards, and you should see store availability in 3 weeks.

 

[Cerb]

Bitcoin miners who prefer AMD cards over nVidia ones would dispute the idea that nVidia cards are better at GPU computing than AMD cards. In fact, they have proven that AMD cards have more processing power than equivalent nVidia cards.

Bitcoin mining, and other crypto, generally favored Radeons, as they were the epitome of simple, wide, and fast. As software gets more complex, that breaks down. If you go look at any set of GPGPU benchmarks, you'll see Radeons tended to dominate or trip over themselves, with not much in between. That is a hallmark of VLIW, and compilers that can get rid of those poor cases simply don't exist.

If AMD is changing their architecture to be closer to that of nVidia, they will be making a serious error.

They aren't. GCN still favors high throughput over fast complex memory access, compared to Fermi. It's not that AMD is trying to mimic Fermi, it's that NVidia got there first, because AMD has had poor management. 64-bit addressing, virtual memory, and virtual functions were going to be needed, whether they added them to VLIW4 or made a clean break w/ GCN. While some game engine developers want those features, HPC users have been chomping at the bit for them.

OTOH, they do have that error that is the launch-version BD to deal with, so it's a good thing GCN performs.

 

[Vesku]

GCN was scheduled with 28nm, 28nm delay=GCN delay. Basically what we are seeing now is what 6000 series would have been without the delayed node.

Bitcoin mining, and other crypto, generally favored Radeons, as they were the epitome of simple, wide, and fast. As software gets more complex, that breaks down. If you go look at any set of GPGPU benchmarks, you'll see Radeons tended to dominate or trip over themselves, with not much in between. That is a hallmark of VLIW, and compilers that can get rid of those poor cases simply don't exist.

They aren't. GCN still favors high throughput over fast complex memory access, compared to Fermi. It's not that AMD is trying to mimic Fermi, it's that NVidia got there first, because AMD has had poor management. 64-bit addressing, virtual memory, and virtual functions were going to be needed, whether they added them to VLIW4 or made a clean break w/ GCN. While some game engine developers want those features, HPC users have been chomping at the bit for them.

OTOH, they do have that error that is the launch-version BD to deal with, so it's a good thing GCN performs.

 

[Cerb]

GCN was scheduled with 28nm, 28nm delay=GCN delay. Basically what we are seeing now is what 6000 series would have been without the delayed node.

Aside from the missing 32nm in that statement, that was kind of my point. Why did they wait until 28nm to add useful memory features, instead of just talking about what the future may bring, while their competitor was making it happen?

 

[StrangerGuy]

The scale of fails:

Bulldozer (slower, hotter) <<<<< 1st gen Fermi (faster, hotter) << 7970 (faster, cooler)

 

[Madcatatlas]

I didn't think it was at all inflammatory and flaming was not my intention.

Also it is more accurate because they didn't just focus on GPGPU, but did so in a way that looks to be the same as what fermi did to me. Unless I am misunderstanding.



Backpedal as in, backpedal for gaming while keeping fermi around for a separate compute line. There is more then enough money in each market to justify that.
And because the architecture makes significant sacrifices in the gaming arena which has cost them.

It sure looks inflammatory, credit to those who ignored the obvoius referance to fermi as a negative. And when tahiti shares NONE of the known issues with the core fermi negatives like hot, powerhog, noisy. It is right to call your topictitle inflammatory

 

[taltamir]

the obvoius referance to fermi as a negative.

Odd that my intention was so obvious to you, when your assumption about what my intentions were is not even true.
It doesn't refer to it as such, and in going into it further I clearly stated it sacrifices some gaming performance for better GPGPU performance.

And when tahiti shares NONE of the known issues with the core fermi negatives like hot, powerhog, noisy. It is right to call your topictitle inflammatory

Ah, I see your problem. You think fermi = bad = hot, noisy, powerhog.
Tahiti = good =! hot, noisy, powerhog.
Therefore tahiti =! fermi.

Thing is, rather then making hyperbole, I was actually talking about architectural design. Things like VLIW4, SIMD, FP64, etc.
Looking back at your "fermi is negative comment" I would say that this seems to be you projecting your own opinions unto me.

Hot & noisy are just symptoms of powerhog. Ultimately all 3 are determined by how agressively the part is clocked as well as manufacturing process and chip design.
Which is ultimately determined by its efficiency & market forces (do they feel threatened enough to be forced to redline it).

A more efficient architecture can be clocked higher without needing to increase power consumption compared to a less efficient one.

My point wasn't that GCN is "bad like fermi" (hyperbole; and assumes fermi was inherently "bad"), but that GCN is architecturally very similar to fermi, and like it is a GPGPU power houses, but that comes at a cost of reduced efficiency for video games. The overall chip can be faster and cooler if coupled with a suitable process shrink (which fermi didn't have) but the results are underwhelming for gamers. Very impressive for a GPGPU though.

They aren't. GCN still favors high throughput over fast complex memory access, compared to Fermi

Could you elaborate a bit about this difference between the two designs?
I don't recall seeing that difference in the anandtech article.

 

[hectorsm]

It sure looks inflammatory, credit to those who ignored the obvoius referance to fermi as a negative. And when tahiti shares NONE of the known issues with the core fermi negatives like hot, powerhog, noisy. It is right to call your topictitle inflammatory

To me it was obvious that the OP was refering to the GPGPU capabilities of Fermi which has been attacked by some people when it was released as a waste of die space.

Fermi GPGPU strategy has been further validated by AMD as the future of GPUs so I think the comparison of valid...at least from my point of view.

On the other hand, AMD approach to GPGPU seems more efficient and they deserve credict for that. At this point I see the 7970 as superior product. I am planing to to replace my gtx260 next year and the AMD cards are looking like a good choice.

 

[Lonbjerg]

To me it was obvious that the OP was refering to the GPGPU capabilities of Fermi which has been attacked by some people when it was released as a waste of die space.

Fermi GPGPU strategy has been further validated by AMD as the future of GPUs so I think the comparison of valid...at least from my point of view.

On the other hand, AMD approach to GPGPU seems more efficient and they deserve credict for that. At this point I see the 7970 as superior product. I am planing to to replace my gtx260 next year and the AMD cards are looking like a good choice.

You cannot make that conclusion until we see "kepler" on 28nm.

I also took the OP to be talking about GPGPU and the arch.

 

[Will Robinson]

Odd that my intention was so obvious to you, when your assumption about what my intentions were is not even true.
It doesn't refer to it as such, and in going into it further I clearly stated it sacrifices some gaming performance for better GPGPU performance.


Ah, I see your problem. You think fermi = bad = hot, noisy, powerhog.
Tahiti = good =! hot, noisy, powerhog.
Therefore tahiti =! fermi.

Thing is, rather then making hyperbole, I was actually talking about architectural design. Things like VLIW4, SIMD, FP64, etc.
Looking back at your "fermi is negative comment" I would say that this seems to be you projecting your own opinions unto me.

Hot & noisy are just symptoms of powerhog. Ultimately all 3 are determined by how agressively the part is clocked as well as manufacturing process and chip design.
Which is ultimately determined by its efficiency & market forces (do they feel threatened enough to be forced to redline it).

A more efficient architecture can be clocked higher without needing to increase power consumption compared to a less efficient one.

My point wasn't that GCN is "bad like fermi" (hyperbole; and assumes fermi was inherently "bad"), but that GCN is architecturally very similar to fermi, and like it is a GPGPU power houses, but that comes at a cost of reduced efficiency for video games. The overall chip can be faster and cooler if coupled with a suitable process shrink (which fermi didn't have) but the results are underwhelming for gamers. Very impressive for a GPGPU though.



Could you elaborate a bit about this difference between the two designs?
I don't recall seeing that difference in the anandtech article.

Looks pretty good to me...

Even at full blast Fermi 480 is still lounder.

source:The Guru of 3D.http://www.guru3d.com/article/radeon-hd-7970-overclock-guide/5

 

[RussianSensation]

I use this on a near daily basis:
http://www.vreveal.com/

Took 10 minuttes to clean that video op...I wouldn't like to think on how long I would have spent if I were using my CPU...

Thanks for sharing. I think GPGPU has a bright future but we are probably at the early stage. Of course, with AMD now onboard with GCN, we might see much faster innovation in this space. Hopefully. :thumbsup:

To me it was obvious that the OP was refering to the GPGPU capabilities of Fermi which has been attacked by some people when it was released as a waste of die space.

People loved to talk about die space during Cypress/Cayman vs. Fermi. But subsequent NV and AMD earnings & financials revealed that NV had superior margins and profitability. So outside of Fudzilla constantly claiming that NV was in trouble, the reality was far different. By itself, the die size tells us nothing about profitability, yields, performance, etc.

When a gamer upgrades from HD4870/GTX260 216 to HD7970, I don't think they'd care if HD7970 was 100mm^2 or 600mm^2 as long as it ran cool and quiet and had the performance they were looking for at their specified budget.

Fermi GPGPU strategy has been further validated by AMD as the future of GPUs so I think the comparison of valid...at least from my point of view.

For sure. Nvidia was way ahead in predicting the future of computing and how a graphics card can be used to accelerate GPGPU compute tasks. I even created a thread why I thought going GPGPU compute for NV made sense, despite sacrificing performance/watt for gaming. AMD might have thought or talked about it, but it took them 3 GPUs since HD4890 to get there.

My summary from that thread was:

"My own takeway is that it probably would be better for NV to design 2 separate chips - 1 for professional graphics and 1 for games. However, due to R&D costs and manufacturing constraints, this is probably not possible given the size of the company. By designing Fermi as an "all-purpose" GPU, NV is thus able to amortize/share its R&D costs across at least 3 product lines including Consumer Graphics, Tesla and Quadro lines. In addition, this business strategy minimizes risks. If one of these product lines suffers a decline in revenue as a result of market forces, chances are that the other 2 product lines will remain healthy until a recovery can occur. By NV not putting all of its eggs in one basket (i.e., consumer products only), it is able to withstand changing market conditions to a greater degree. "

On the other hand, AMD approach to GPGPU seems more efficient and they deserve credict for that.

At the moment, yes. But you have to realize that 7970 is AMD's 1st GPGPU architecture, while Kepler will be NV's 2nd. HD7970 came out on 28nm, not on 40nm. The only way to fairly compare the efficiency of both architectures is to place them on equal nodes. When Kepler arrives, we'll be able to do that. For now, we are comparing a 2-year-old Fermi architecture on 40nm to a brand new GCN architecture on 28nm. It was obviously expected that GCN would be far faster than GF100/110 in compute and in games. The question is, did AMD do enough with 7970 to position it well against NV's 28nm GPU? We'll find out in 2012. Exciting times ahead.

 

[Lonbjerg]

Thanks for sharing. I think GPGPU has a bright future but we are probably at the early stage. Of course, with AMD now onboard with GCN, we might see much faster innovation in this space. :thumbsup:

I hope so, my only consern is that AMD has no ECO system, unlike what CUDA has.
It's CUDA gradstudents tinker with...not DirectCompute or OpenCL.

AMD has a long haul in front of them and they don't need to onvince people about the benefits of GPGPU...they need to convince them to use something else than CUDA....that is a different battle.

The Vreveal software I linked to don't even figure on NVIDIA's HP:
http://www.nvidia.com/object/cuda_app_tesla.html

If you look at AMD's HP...it's sadly lacking...lot of fancy words, but no substance.

 

[Arkadrel]

I hope so, my only consern is that AMD has no ECO system, unlike what CUDA has.
It's CUDA gradstudents tinker with...not DirectCompute or OpenCL.

AMD has a long haul in front of them and they don't need to onvince people about the benefits of GPGPU...they need to convince them to use something else than CUDA....that is a different battle.

The Vreveal software I linked to don't even figure on NVIDIA's HP:
http://www.nvidia.com/object/cuda_app_tesla.html

If you look at AMD's HP...it's sadly lacking...lot of fancy words, but no substance.

Ironic?

http://www.vreveal.com/vision

Looks like it ("The Vreveal software") runs on OpenCL...


I personally think, OpenCL (which arm/amd/intel/nvidia) can use/support is going to kill off CUDA.

 

[Lonbjerg]

Ironic?

http://www.vreveal.com/vision

Looks like it ("The Vreveal software") runs on OpenCL...


I personally think, OpenCL (which arm/amd/intel/nvidia) can use/support is going to kill off CUDA.

LOL
We are talking about GPGPU, not the APU crap:
http://www.vreveal.com/premium

Sorry of that wasn't in your AMD press kit...

 

[Arkadrel]

thats just it... its not just the apus....ur gpu's do opencl as well.
That program uses opncl as well.

http://www.vreveal.com/

GPU Acceleration:
Faster than fast. vReveal now runs on x86, OpenCL (AMD GPUs) and CUDA (NVIDIA GPUs).
Enhance 1080p video at 30 FPS.

its right there on the front page.

 

[Lonbjerg]

thats just it... its not just the apus....ur gpu's do opencl as well.
That program uses opncl as well.

http://www.vreveal.com/



its right there on the front page.

And it's right here in my link:

Plus, take advantage of the fastest encoding to H.264 using NVIDIA CUDA (up to 6 times faster than the CPU).

Plus, take advantage of the fastest encoding to H.264 using NVIDIA CUDA (up to 6 times faster than the CPU).

Plus, take advantage of the fastest encoding to H.264 using NVIDIA CUDA (up to 6 times faster than the CPU).

Not via OpenCL...:biggrin:

 

[MrK6]

And it's right here in my link:

Plus, take advantage of the fastest encoding to H.264 using NVIDIA CUDA (up to 6 times faster than the CPU).

Plus, take advantage of the fastest encoding to H.264 using NVIDIA CUDA (up to 6 times faster than the CPU).

Plus, take advantage of the fastest encoding to H.264 using NVIDIA CUDA (up to 6 times faster than the CPU).

Not via OpenCL...:biggrin:

And it doesn't say "only via NVIDIA CUDA" either. So instead of being childish, why don't you wait to see what applications start to support the 7970 once it's actually released.

Also, the title should be "AMD didn't go Fermi" since they released a new architecture on a new process that wasn't a laughable disappointment.

 

[swilli89]

Aside from the missing 32nm in that statement, that was kind of my point. Why did they wait until 28nm to add useful memory features, instead of just talking about what the future may bring, while their competitor was making it happen?

Nope, 32nm IS the 28nm process. 32nm was so heavily delayed they changed the name to 28nm process, as the method used to measure this is subjective and not absolute.