News and Rumours: PS4

[mrcorbo]

I'm not saying I don't like AMDs efforts. It's just not cohesive enough for me personally. You are putting two fundamentally different paradigms together effectively producing a SoC but not an orthogonal architecture. It's a lot more work to get going and if you build a network of such devices you are looking at a lot of redundancy.

Development of their APUs has been an iterative process for AMD, necessitated by the fact that they can't afford to continue to invest in development without being able to turn them into shipping products. So what you have seen so far are the first steps in the integration and not the finished product. The cohesion will come with further iterations and will arrive courtesy of both further hardware development and API support. The next couple of iterations will each be large leaps, though. Especially,the move to bring a unified address space and memory coherency across CPU and GPU operations which is scheduled for 2013.

CELL had the advantage of accessible parallelism wrapped up in a neat package. I like the ARM approach because they are easy to put together in heterogenous groupings (I may be swayed by a little patriotism as well )

ARM are awesome at what they do. Any company that can stonewall Intel from entering a market that they are coming after so aggressively deserve a great deal of respect. And the Exynos in my SGSII-derivative cell phone is damn impressive technology by any standard, even though it's old news now . What is it about ARM that you find makes it easier to put together in to heterogenous groupings?

And I admit I am more interested in the fringe exotic architectures. But only because that is where all the exciting things happen. Intel and AMD have been too busy with each other to notice that a lot of powerful silicon has been creeping into server tech and beyond. None of it bound by decades old legacy support.

I remember when Motorola brought out the 68060 and Intel was on the 386 with the 486 round the corner. The '060 was a lovely harvard design, risc core, the works. And the 386 was a faster 286 with a dodgy FPU (and a proper MMU). That's kind of like the situation between Intel and AMD today. Meanwhile Acorn and others, including AMD, were working on interesting designs that did things differently. I liked that!

I think that you could draw a parallel between the evolution of the FPU (which started as a separate chip in it's own socket, was eventually integrated physically and then was gradually made more integral to the processor design as a whole while continuing to expand in capability) and what AMD are trying to do with the stream processors in their APU. I even suspect that AMD will attempt to deprecate their existing vector units by using the GPU elements to provide those functions where possible thereby eliminating redundancy.

 

[mrcorbo]

And CELL didn't put together two fundamentally different paradigms when they included a 1xPPE with 7xSPEs? We are talking degrees but SPEs with LS (instead of cache) and relying strongly on SIMD is a big departure.



Yet ARM performance is the pits and just throwing, oh, say 32 of them on a chip without any concerted effort for inter-chip communication and smart memory access/sharing would be a disaster.



Sadly exciting and good aren't synonyms of good design.



Like?

Intel's and AMD's offerings trend toward volume. Intel's Itanium series aside they aren't doing stuff like IBM where they are selling entire platforms with chips with insane TDP for tens of thousands of dollars. That isn't their business model--but neither are those crazy designs even relevant because the *cost* of such isn't relevant. Don't get me wrong there is really cool stuff out there; when you limit the discussion to chips and technology that is financially feasible for consoles and, more importantly, conducive to a 2 year development cycle for games, those neat technology pieces become far and few in between.



Ok, so Intel has an exotic serve chip in Itanium, their core business of x86 chips that stretch from servers (Xeon), high end enthusiast rigs all they way down to low end desktop (i3, i5, i7) with performance mobile counterparts, they have their Atom series which has dual core variants in all sort of nettops and netbooks and now even pushed down into the mobile sector with Medfield, etc. Intel has been actively working on their GPU technology (see Haswel) and has been expanding their SIMD performance (e.g. AVX). They have been progressing their interconnects with Quickpath and dabbling in stacked memory.

I read exotic as "we have no prayer of competing with Intel on their turf so we are going to try something completely different, throw it at the wall, and see if it sticks in a niche market."

And to be blunt, that is exactly where ARM has excelled. It isn't unlike the problem MS has had getting Windows into the mobile space: x86 as a modern architecture until recently couldn't compete in the mobile space. Its features and legacy support made it too big and it wasn't designed with the power constraints; strip the features and fit it into a power envelop that is acceptable and performance would be so poor it wouldn't matter if it was x86 compatible. This, along with cheap licensing, allowed ARM to have a niche market uncontested by Intel/AMD. Now that there is volume to be had in that sector and node progression has knocked down some barriers (plus the fact successive generations of mobile software have needed more power pushing performance needs more into territory Intel could comfortably address) and the foresight of Intel to use Atom as a baseline architecture with an eye toward the ultra mobile sector ... and things change. A lot.

The funny thing about ARM is seeing how much people SLAG Atom processors but Medfield, a single core with HT processor, is right up at the top of the pack against ARM. And people (not necessarily you, but some posters here) seriously talk about wanting a sea of these chips, as if Amdahl's law magically vaporized into thin air and ARM has some magic sauce of inter-core communication and memory management. I don't see how ARM in any way is addressing the needs of console makers making a progressive platform (until what time someone like Nvidia releases a high end performance variant of ARM).

Anyways, popcorn time. With Sony losing $5B it will be interesting to see what they do.

Don't agree with your assessment of ARM. Too late and too much to address right now, though. How do you type so much so fast?

 

[-tkf-]

I believe ERP, among others, replied to this earlier. Why not quote him and specifically state what you don't understand or what you disagree with an experience console programmer about.

Actually Shifty had the arguments that worked for me. I always thought about 2 things. Performance and investments in the current development tools. Both which i considered important.

Having to rely on a outdated and expensive memory type and not having a economic partner in the development process is major problems.

 

[mrcorbo]

Yet ARM performance is the pits and just throwing, oh, say 32 of them on a chip without any concerted effort for inter-chip communication and smart memory access/sharing would be a disaster.

That's overly harsh, IMO. Place the same die size, voltage and TDP restrictions on a design based on any other architecture and what kind of performance would you expect?

I read exotic as "we have no prayer of competing with Intel on their turf so we are going to try something completely different, throw it at the wall, and see if it sticks in a niche market."

And to be blunt, that is exactly where ARM has excelled.

ARM is not niche. It's ubiquitous. They are the dominant architecture in the fastest growing computing segment and that market growth is accelerating.

It isn't unlike the problem MS has had getting Windows into the mobile space: x86 as a modern architecture until recently couldn't compete in the mobile space. Its features and legacy support made it too big and it wasn't designed with the power constraints; strip the features and fit it into a power envelop that is acceptable and performance would be so poor it wouldn't matter if it was x86 compatible. This, along with cheap licensing, allowed ARM to have a niche market uncontested by Intel/AMD.

I can't see how you define the smartphone/tablet/integrated market as niche. It's HUGE.

Now that there is volume to be had in that sector and node progression has knocked down some barriers (plus the fact successive generations of mobile software have needed more power pushing performance needs more into territory Intel could comfortably address) and the foresight of Intel to use Atom as a baseline architecture with an eye toward the ultra mobile sector ... and things change. A lot.

The funny thing about ARM is seeing how much people SLAG Atom processors but Medfield, a single core with HT processor, is right up at the top of the pack against ARM.

People slag Atom as a mobile processor because despite very public smack-talk from Intel they have yet to release a compelling product in this market. Medfield-based products are going to appear on the market just in time to be completely eclipsed by products that use the next generation of ARM-based designs.