Could MIPS make a comeback in consumer gaming devices?

[Mobius1aic]

Supposedly there is a new push new make MIPS a new rival to ARM in the tablet market, where the architecture is cheap, powerful enough and efficient enough to take on ARM's pretty much exclusive hold on tablets and phones. With the release of the PSVita, it will also be the only CPU architecture in current handheld gaming consoles.

With MIPS now running Ice Cream Sandwidth it seems that MIPS is due for a comeback, especially with the possible push for cheap $100 tablets. In my MIPS related travels, I came to understand that the Chinese developed Loongson series CPUs are in fact a MIPS compatible family of processors, which I had not previously known, and I also learned of the little beasts known as the Loongson 3B.

Loongson 3B
The updated 8-core 65nm Loongson 3B processor runs at 1.05 GHz, with a peak performance of 128 GFLOPS double-precision, or 256 GFLOPS single-precision.[10] This is accomplished by having two 256-bit vector processing units in each core. They produce a peak performance of 8 double-precision floating-point fused Multiply-Add results per cycle, or 16 GFLOPS per core operating at 1 GHz. The Godson-3B has exceptional energy efficiency in terms of performance per watt -- executing 128 GFLOPS using 40 watts.

There are also versions like the Godson T which are meant to have 64 cores. They are aimed at scientific computing though but it is still very interesting.

Considering MIPS in new and very cheap tablets, and the sheer GFLOPS performance of these currently developmental high end Loongson 3 processors, is it possible we could see a return of MIPS with a vengeance to gaming, at least in the mobile sector? The potential of the Chinese market for a gaming console is immense and something like the Loongson 3B (especially on a very small node size) or a derivative perhaps could form the basis for a very decent CPU if the non vector unit performance of the cores is up to snuff. Such a system, with the right publishers and partners could probably make the jump across the pond to the US or Europe. Even without a home console, MIPS on tablet running Android 4 at such a low price is a very big development if the performance is there to make games a possibility.

It makes me wonder what the video acceleration hardware is though....

It's a Vivante GC860: http://www.vivantecorp.com/p_mvr.html (GC800 series is midway down the page). Not bad specs for a mobile GPU in a $100 tablet. Not bad at all.

 

[Grall]

Sure it COULD happen, but why would it?

If there's no particular or compelling reason to choose MIPS over ARM, there's little chance any manufacurer would. It'd be cheaper, simpler and easier to go with the established standard and get very predictable power consumption, performance and software support.

Smart phones and app stores are becoming more and more important, customers would be very annoyed if a phone they bought (which contains a MIPS platform) doesn't deliver the kind of grunt or battery life they're used to.

 

[Shifty Geezer]

As for the Loongson 3B, those theoretical specs won't amount to squat if the memory architecture isn't there to support that throughput. That's 16 vector units that need to be supplied. Cell is feeding 8 cores with single 128 bit vectors each. Loongson is quadrupling the throughput!

 

[Mobius1aic]

As for the Loongson 3B, those theoretical specs won't amount to squat if the memory architecture isn't there to support that throughput. That's 16 vector units that need to be supplied. Cell is feeding 8 cores with single 128 bit vectors each. Loongson is quadrupling the throughput!

Either way it's pretty awesome isn't it? :smile:

Anyways, I'd reduced it to a quad core for any kind of home platform, and as I pondered earlier, the non-vector part of the cores for all we know could be worthless for anything beyond managing those vector units.

Why I wonder about MIPS making a return to the forefront is due to the introduction of the architecture in tablets, which could gain momentum in markets like Asia, which could trickle into let's say, a Chinese market only gaming console. In order to keep the tech in house, I'm just theorizing that perhaps Loongson could be part of a comprehensive console set up. We know how they are about letting outsiders introduce consoles into their market. And perhaps with the right push behind it, Loongson based PCs + Linux with the newer Loongson 3 architectures could make a place in the PC gaming market there.

 

[Shifty Geezer]

Either way it's pretty awesome isn't it? :smile:

At this point I'd conjecture it's a flawed design that's wasting silicon on an excess of execution units. You could burst-process some data whixxing through the caches, but sustained function should see lots of idle units, unless they are doing something jolly clever.

Why I wonder about MIPS making a return to the forefront is due to the introduction of the architecture in tablets, which could gain momentum in markets like Asia, which could trickle into let's say, a Chinese market only gaming console. In order to keep the tech in house, I'm just theorizing that perhaps Loongson could be part of a comprehensive console set up. We know how they are about letting outsiders introduce consoles into their market. And perhaps with the right push behind it, Loongson based PCs + Linux with the newer Loongson 3 architectures could make a place in the PC gaming market there.

That's a fair argument. That part of the world may embrace MIPS for licensing reasons leading to exported devices and eventually a flow of MIPS into the mobile space. I can't see any existing console or tablet going for it. It may feature prominently in those cheap Chinese knockoff consoles. A portable PlayStation, perhaps.

 

[Flux]

How much low latency eRam and main ram does a chip like this require for bandwidth?


16-24MB eRAM/4GB GDDRAM?

What about a 16(4x4) core ARM Cortex a15 @2.5GHz with 4GB ram and on fast low latency on die memory. The smart phones made arm's pretty cheap now.

PowerVR and their multicore GPUs are cheap too.

 

[Shifty Geezer]

512 bits per cycle * 8 cores is 4096 bits per clock, or 512 bytes per clock. At 1 GHz that's 512 GB/s data throughput. Latencies can be worked around, but setting up enough data throughput to keep those ALUs busy is gonna be nigh impossible unless they are using eDRAM and very expensive memory interfaces in their super computer, or something clever like working in SRAM for limited workloads.

 

[JohnH]

That would be the instruction to instruction BW not the external BW i.e. it's to internal register files. External BW would be substantially lower e.g. for 10 instructions per data fetch 51GB/s, for 100 instructions per data fetch 5GB/s etc.

 

[Flux]

That would be the instruction to instruction BW not the external BW i.e. it's to internal register files. External BW would be substantially lower e.g. for 10 instructions per data fetch 51GB/s, for 100 instructions per data fetch 5GB/s etc.

So is that why newer chips have higher caches?

Because the bandwidth to the register files?

 

[Shifty Geezer]

That would be the instruction to instruction BW not the external BW i.e. it's to internal register files. External BW would be substantially lower e.g. for 10 instructions per data fetch 51GB/s, for 100 instructions per data fetch 5GB/s etc.

True enough, and my reponse was far too from-the-hip. I'm just concerned over that many execution units actually being useful outside of specialist scientific computing. There's a reason Intel and friends don't just fit 4+ vector units per core for MASSIVE POWER.

 

[Flux]

True enough, and my reponse was far too from-the-hip. I'm just concerned over that many execution units actually being useful outside of specialist scientific computing. There's a reason Intel and friends don't just fit 4+ vector units per core for MASSIVE POWER.

What are the diminished returns for just lumping vector units on a core?

 

[Shifty Geezer]

How much vector workload you have to do relative to everything else, I'd guess. GPUs have a shedload of maths units per 'core'. A CPU isn't typically working with those problems for that amount of a time. As a general purpose CPU, this Loongson 3B wouldn't be expected to do such maths-heavy work. Cell only has to do as much maths as it's given because RSX isn't up the job alone! A future console would have a lot of that maths silicon sitting idle much of the time.

For what it's worth, follow the ISSCC link from Wikipedia and you get this:

The Godson-3B processor is an 8-core high-performance processor implemented in a 65nm CMOS LP/GP mixed process with 7 layers of Cu metallization. It contains 582.6M transistors in a 299.8mm2 area. The highest frequency of Godson-3B is 1.05GHz. Its peak performance is 128/256GFLOPS for double/single-precision with 40W power consumption.

Similar Cell like double/single precision tradeoffs. Cell was 174 mm^2 at 65nm.

 

[tunafish]

The only reason the Loongson series exists is that the Chinese govt wants to be less dependent on foreign tech. If it was a real commercial venture, it would have gone bust half a decade ago.

... Then again, that also means that the Chinese govt is perfectly happy to finance it's development at a loss for essentially forever, and possibly might be persuaded to straight-out subsidize it's use in commercial devices to promote domestic industry and additional development. So it could well pop out in places where there's no good reason to expect it.

As for the merits of the ISA's, 64-bit MIPS is essentially *really* close to the published 64-bit ARM ISA, so whatever you can expect from one of them, the other can do too.