Differences/Similarities in architecture and Design Philosophies of Fermi & Tahiti

[taltamir]

A cerebral discussion on the "Differences and Similarities in Microarchitecture and Design Philosophies of Fermi and Tahiti".
This is not about which is a better GPU, the better buy, etc. This is not comparing products (like the 5870 or the GTX580).
Rather, this is a discussion on the architectural design decisions only in an abstract academic manner.

Some reading about Fermi and Tahiti architectures:
http://www.anandtech.com/show/2849
http://www.anandtech.com/show/4008/nvidias-geforce-gtx-580/2
http://www.anandtech.com/show/4455/amds-graphics-core-next-preview-amd-architects-for-compute
http://www.anandtech.com/show/5261/amd-radeon-hd-7970-review

Architecture:
The most obvious change in tahiti is that it replaces VLIW with simple SIMD; according to anantech that gives much more stable and predictable performance for compute. Usually faster, although sometimes slower then VLIW but overall more compute friendly.

AMD groups their SIMDs into arrays of 16, each such group gets its own 64K cache.
A CU (Compute Unit) is 4 such arrays
nVidia basic grouping is arrays of 32 SIMD units called SM, each group gets its own 64K cache.
4 of those are grouped into a GPC unit

So from that it appears that each SIMD on tahiti gets 2x the cache of fermi. I suspect this is at lest partially due to being designed for a 28nm process, allowing it much transistor budget to make cache with.

Aside from AMD having double the cache per SIMD unit, the two look very similar overall. There is only so much info to be gleaned from staring at architecture overview pictures, so I was wondering if someone who is more familiar with the technical aspects of it could point out where they diverge and why.

Philosophy:
According to anandtech's titles, Tahiti is "Architected For Compute" and Fermi is "Architected for Tesla". I agree that both place heavy emphasis on compute.
It seems to me that both nvidia an AMD currently believe the following two design philosophies
1. GPGPU is the future
2. It is better to develop a single GPU that balances its performance in GPGPU and in video gaming rather then a unique part for each.

I am surprised by this convergence of philosophy between the two since it seems an odd position to take. I would instinctively feel that it would be best to design two separate architectures for those two market segments. One of the biggest advantages of such segmentation is that you can charge an arm and a leg for compute parts from your corporate/science customers as your compute cards (eg, nVidia Tesla, AMD FirePro) do not compete with gaming devices.
The second is that you could eke out slightly more performance in each field by tossing out unneeded parts... A tesla card can do without most of the fixed function gaming stuff and VLIW4 has done very well against SIMD arrays in gaming.

However, I don't have the actual cost numbers to back this up. How much does it actually cost to develop an architecture?

It occurs to me that another plausible explanation for the "one chip for both" choice is not saving money on design costs, but if they believe that games are going to make heavy use compute as well. If that occurs then the GPU architectures designed with compute in mind will do very well on such games.
IIRC the makers of the very popular unreal engine suggested as such in an interview. Are there other indications that games are going to be heavily compute dependent in the future?

 

[Concillian]

I think one of the biggest reasons that they didn't stick with two architectures is the direction gaming developers are headed.

Namely, doing less in-situ AA and more post-processing image enhancement.

We're just starting to see a glimpse of what looks like real consumer level products with actual resolution improvements with the rumored iPad3 HIGH display resolution and rumored apple 4 megapixel 27" display.

If displays (finally) start moving in this direction, then video game graphics will get more and more like compute loads, with just needing raw Gtexels/sec throughput + some FP16 processing.

Graphics as they are done now, will simply not be practical for 4+ megapixel displays. MSAA takes a lot of extra processing and memory, and as pixel pitch decreases, the disparity in quality between MSAA / TrAA, etc... and a post processed AA like FXAA diminishes. I think we'll see quality distinguishment coming in the form of ambient occlusion shading and such, rather than the AA modes that we see now. Things that are reasonably well suited to a compute architecture.

At the end of the day, the developers decide what we can do with our hardware, and like it or not, they are going to be developing for consoles and porting to the PC or vice-versa... which means they are going to want to do things a similar way on multiple platforms. High resolution displays will give PCs their quality niche while allowing developers to do things the easy way (optimize for FXAA on all platforms rather than optimizing for FXAA so the game can actually run on the lean hardware a console uses and then tack on some PC functions, even if they're not well optimized.)

I think the software side is already converging to some extent. It's possible AMD is just listening to developers. I mean our hardware is only as good as the software available that runs on it. I'm sure they were asking for more similarity between the two brands of video cards, it must be a nightmare for them to have such dramatically different architectures. Take Civ 5, where they add some compute functions and it means there is a large divide in performance in some areas. Not a huge deal in a game like Civ 5, but imagine how many games specifically avoided compute friendly development paths just because AMD wasn't that great at it and they want as large a potential customer base as they could get?

Long story short, I don't think it's necessarily the hardware that's driving the bus. It very well could be the gaming developers who steered AMD in this direction.