HD 5000 series: New architecture or more a major refresh?

[CarstenS]

What's your take on this?

Me, I'm not quite decided. On the one hand, they have of course integrated support for DirectX 11 and for example moved attribute interpolation into the shader core. On the other hand though, the way AMD presented the new chips made me feel quite familiar with their general design (i.e. SIMDs of 16x5 with a quad tmu) more in the way, GT200 was a refresh to G80 (the former of course lacking in support for a new DX version).

So, how would you classify this - if at all?

 

[Ailuros]

Boy I smell some serious catfights for that topic again

My own KISS approach on the issue is that execution and especially for an entire DX11 family is more important at the moment than anything else. If NVIDIA would have launched it's own DX11 family at the same time as AMD and as quickly from top to bottom, things would be different. With NV's delay though any possible architectural disadvantage one could come up with gets overrun to a degree with AMD being earlier there and I'm quite sure that ATI engineers won't sit idle in the longrun either.

That of course under the presupposition that Fermi truly poses as something far more efficient which is still unknown.

Personally I don't have any significant objections for the Evergreen family as they stand today.

 

[mapel110]

From a gamer view: A little bettter AF and DX11 - All the rest was driver work ATI did, afaik. So I am disappointed by this new generation and would call it a careful step forward.

 

[fellix]

The similarity in the general design often contradicts with the numerous subtle changes, upgrades and additions, somehow obscured for the common perception. I think there's enough "critical mass" of those small tweaks to consider HD5000 a new architecture, after all. One have just to sum and quantify those in a proper manner to project the complete picture.

 

[CarstenS]

With NV's delay though any possible architectural disadvantage one could come up with gets overrun to a degree with AMD being earlier there and I'm quite sure that ATI engineers won't sit idle in the longrun either.

Personally I don't have any significant objections for the Evergreen family as they stand today.

I am not talking advantages and disadvantages here - you could only summon up those for HD 5k compared to 4k. And there's going to be almost none of the latter (disadvantages) as far as i can tell.

I also don't have objections about Evergreen - otherwise I surely wouldn't have spent 350 Euros on an HD 5870 (Dragon Age even looks ok with Supersampling, looking forward to Drakensang: Am Fluss der Zeit though). I just want to know, what you think it is: new or refresh.

 

[dizietsma]

Refresh to me I think, though a good one.

 

[Ailuros]

I just want to know, what you think it is: new or refresh.

You can't jump to a new technology generation with a simple refresh. I consider it a new architecture with the least possible changes on the base elements of their preceeding architecture. And yes that might pose an oxymoron, but not necessarily a bigger one then "going to a new tech generation with a refresh".

 

[Blazkowicz]

From a gamer view: A little bettter AF and DX11 - All the rest was driver work ATI did, afaik. So I am disappointed by this new generation and would call it a careful step forward.

great AF, and the supersampling modes. this differenciates HD 5000 from the HD 4000 and G8x/G2xx geforce.
so it appeals to a fringe user like me.

they're milking the R600 architecture, and there's nothing wrong from a practical gaming standpoint. with of course a much bigger difference between R600 and RV870, versus G80 and G215.

 

[Lonbjerg]

The final evolution of the R600 in my book.
Same cores, the DX11 part of the GPU die is acutally very small...and since the have had "tessalation" (aka trueform in one form or another for years) really not a new "architechture".

For me it's a dieskrunken R770 with very little addded...not that it is a bad thing, it's the current top GPU

 

[MfA]

The LDS is fundamentally different, it has "coherent" L2,all the FP is IEEE compliant ... I could see them putting in everything they lack over the coming competition feature wise (writeable L1 cache, multiple rasterizers, branching to shader computed addresses and virtual addressing ... coherent L2 they already put in) while still retaining the basic familiar SIMD structure.

Fermi as different as it is from GT200 still is still has 32 wide SIMD engine just like it's predecessors, by all appearances the fundamental design aims for GPUs have not changed that much. You only have to change what you didn't get right the previous time around.

 

[trinibwoy]

I think it's an easy question to answer once you look at what we've considered new architectures in the past. It's certainly not a G80 or R600 level change.

 

[Tahir2]

If we look at the number of enhancements the architecture:

For DX11 compliance several blocks have been improved, increased bandwidth to the L1 caches and increased L2 caches, improvements in AF, audio video decoding has been improved also.

12bit sub pixel precision, EDC, improved ALU usage, better power management ... everything has been tweaked.

So it's the same architecture, just more efficient in many cases and with more features.

What was the overall difference between G92 and G200?

 

[rpg.314]

The LDS is fundamentally different, it has "coherent" L2,all the FP is IEEE compliant ... I could see them putting in everything they lack over the coming competition feature wise (writeable L1 cache, multiple rasterizers, branching to shader computed addresses and virtual addressing ... coherent L2 they already put in) while still retaining the basic familiar SIMD structure.

Fermi as different as it is from GT200 still is still has 32 wide SIMD engine just like it's predecessors, by all appearances the fundamental design aims for GPUs have not changed that much. You only have to change what you didn't get right the previous time around.

Is L2 in r8xx writable or read only? I was under the impression that it was the latter.

 

[MfA]

I don't have one to measure throughput/latency of read after write vs. read from memory (would be interested in the results though) but in the Siggraph Asia PDF they said the global memory cache was R/W with atomics.

 

[no-X]

The question is: What is "new architecture"? If an architectural concept suits well for current and forthcoming games, then there's no reason to change it. Major changes always come after a fruitless generation (NV3x -> NV40, G71 -> G80 etc.)

RV870 is based on similar concept as RV770 (which was based on heavily modified R600). Adoption of DX10.1 was very easy; in fact e.g. DX10.1 RV535 had less transistors than DX10 RV530. Adoption of DX11 didn't consume much die area, too - RV840 (800SPs/40TMUs/16ROPs) = 10% more transistors than on RV770. Low-end RV830 GPU is even more efective per transistor, than its predecessor.

I think the architecture was developed with DX-next adaptability on mind and that's the reason why RV870 is close to RV770. I'd say RV870 is new architecture based on existing concept.

 

[PSU-failure]

Refresh.

It's a bigger refresh than GT200 over G92 (1 more SIMD for each partition and almost nothing else), but nothing really changed except some DX11 prerequisites (more local/shared memory) and some tweaks probably free given these prerequisites (more local and shared memory = wider SIMDs, so the addition of a few instructions didn't affect size).

R600 = new arch // G80 new arch
RV670 = shrink // G92 refresh (G84 based upscale)
RV770 = refresh // GT200 refresh
Cypress = refresh

 

[trinibwoy]

GT200 added double precision and doubled register file capacity per SM. Nvidia claims TMUs were tweaked as well. Memory coalescing was also greatly improved over G80. I don't see Cypress being much more of a change.

 

[TKK]

I think it's an easy question to answer once you look at what we've considered new architectures in the past. It's certainly not a G80 or R600 level change.

Agreed.

If we consider "new architecture" as something "completely new in every aspect", then I'd say HD5k series is only a refresh (though with some significant changes, obviously). Even performance-per-SIMD-per-clock didn't change much compared to RV770, so from a performance point of view, RV770 was a more significant step than Cypress.

Btw by that definition, I don't consider Fermi/GF100 a brand-new architecture, either. Many changes over G80/GT200, sure, but it's still scalar cores, optimised to clock 2.x times as high as the rest of the chip.

So basically, I agree with PSU-failure, everythingthat came after G80/R600 so far can be considered refreshes.

I would, however, subcategorize them into those two categories:
Minor Refreshes (RV670, G92, 40nm-GT21x), and
Major Refreshes (RV770, Cypress, GT200, Fermi)


Should AMD go from current superscalar cores to scalar cores with separate clock domain in the future, that would be a new architecture IMO, even if that arch was still DX11.

 

[Richard]

Are new architectures expected before the next-gen console cycle?* Anyway, refresh (obviously).

It comes from a lineage descendant from the troubled R600 of old, but which has since atoned for all past transgressions, and especially with RV770 which was a truly excellent chip. Cypress isn't a major departure, but rather a natural evolution of past designs, and evolution made possible by the fact that DX11 itself is also more evolutionary than revolutionary.

* This does not mean retail launch date.

 

[MfA]

Creating a suboptimal schedule for your main market for a contract you might never get is rather risky ... so no, I don't think the consoles play a huge part in it.

PS. if Microsoft/Sony had promised some major contracts already would that show up in financial filings?