Pandora's Closest Rival Cpu Speed Wise?

[icurafu]

Just remember the blackfin is very slow per MHz. Much slower then the ARM9 in the GP2X. I'm sure one of the most overclocked GP2Xs would outperform it.

I'm ot saying blackfin is bad, because its cheap and energy efficient, which makes it perfect for chinese devices. That need perfformance in ways not available though a DSP.

Strangely, a totally unoptimised port of snes9x to the JXD301 runs very well, even with all the graphical effects turned on. The same can't be said for the GP2X.


Interesting. I'm sure you've already read this Orkie, but for anyone else, this is a good 2 year-old read on perforance and price.

http://blackfin.uclinux.org/gf/download/do...ntroduction.ppt

entitled (uClinux on a $5 Processor, that's $5 two years ago)

The DMIPS score for the 600MHz CPU is 235 which is an improvement based on a new toolchain. GP2X should be 220 DMIPSs at standard 200MHz. Anyway, in relation to SNES performance, I guess it means you should never trust DMIPS. Or that an FPU has a massive effect on most unoptimised software.

 

[Svartalf]

It has nothing to do with ARM or XScale, except perhaps Intel used its XScale knowledge to help Analog Devices develop it.
Its instruction set is completely unrelated to any ARM, even the register file is completely different

Hm... Guess that's what I get for listening to one of the guys they work with on a regular basis- he's one of the guys they hire for making their eval boards for the accelerometers, etc. on a regular basis. What I said was what I was led to believe was the case from him. Guess he's wrong.

 

[ldesnogu]

Most Embedded Proc's are benchmarked in terms of Int OPs(DMIPs,etc.), while DSPs are benchmarked in FLOPS. Which re-itterates my point that the Blackfin line is a DSP first and CPU second.

That's not true. Most DSP's are fixed-point only, you can count those that are using floating-point

QUOTE
That and the fact that the ARM Tool chain is more advanced and cheaper.

Ouch, that means that Blackfin tool chain costs more than a few thousand dollars? :blink:

Hm... Guess that's what I get for listening to one of the guys they work with on a regular basis- he's one of the guys they hire for making their eval boards for the accelerometers, etc. on a regular basis. What I said was what I was led to believe was the case from him. Guess he's wrong.

You should stop listening to him, he is bullshitting you or simply doesn't know what he's talking about

 

[quartercast]

Say, anyone try using SmartQ firmware on their JXD 301? There appears to be a d/l link here.

GBA emulation seems to have improved a lot over the X760 and 301.



Never mind.. the hardware and firmware is different. In the SmartQ fw, emulators are actually packaged individually!

 

[netbsd]

How about Ingenic's Jz4732/Jz4740? They are used in Chinese MP4 players and are said to be cheaper than Blackfins and/or faster. SoC, 360 (400?) MHz, energy efficient, MIPS32, SIMD instructions, Linux port.

Recently, they have started appearing in cheap laptops (3K Computers, Yinlips, Bestlink, etc).

Unfortunately I haven't been able to find any decent documentation, not even in Chinese. As an example, the Jz4732 isn't mentioned on their ingenic.cn site, while you can find pictures of them in the wild.

 

[craigix]

Did we actually ever find out for sure that the SNES9X port isn't optimised? As the other emus on the JXD don't run very well at all in comparison.

 

[DisgruntleElf]

Did we actually ever find out for sure that the SNES9X port isn't optimised? As the other emus on the JXD don't run very well at all in comparison.

In all fairness all the GP2X's emulators have had many revisions. The JXD's emulators haven't changed since its release. If the JXD had as much devoted development time as the GP2X only then would it be fair to compare the two.

 

[Cervante]

Honestly this looks more like a product I would pay $79.99 for, that $100 extra is highway robbery.

 

[Exophase]

There are some similarities between the XScale's added DSP (40bit accumulator and whatever) instructions and some of the ones in the Blackfin, but that's saying very little.

Last benchmark I saw put 600MHz Blackfin at around 365 DMIPS, not 235. Of course it was already established that DMIPS are useless for a number of reasons, but in this case the glaring problem is GCC. Blackfin is a VLIW CPU and GCC is not good at generating code for VLIW (see Itanium). In this case there's probably even less commercial interest to improve it, especially when the company pushing it wants to sell you their own compiler.

SNES emulation might be helped a lot on the Blackfin by having more cache. That can do a ton for the way the video rendering works. Who knows what they did there though. Still enough, the raw power even with GCC should be well above GP2X at 200MHz.

But Blackfin is still not very interesting as far as CPUs or DSPs go. It does have parallel execution, it can do more than one operation per cycle (and the latency for operations in general isn't terrible) but the restrictions on what you can pair severely limits you. On top of that, there are tons of instructions that are only useful for very limited applications, but it lacks a lot of better general purpose CPU design (like a lot of general purpose registers, instead opting for a number of more special purpose ones), which probably further explains the poor DMIPS score via GCC. Probably doesn't do very well with branches and aside from "zero overhead loops" doesn't offer much to help it (like conditional execution).

The thing is, there's no sense really even comparing the Blackfin to the Cortex-A8 in the OMAP3530 (although I have little doubt that the Cortex-A8 would stomp the Blackfin 5 times over in GCC output); instead you should be comparing it to the C64x+ DSP it has. And I can tell you off the bat that even though the C64x+ is clocked a little under 600MHz (435MHz iirc) it's a much more powerful and versatile DSP than Blackfin. Even then, TI had enough sense long ago to realize that a great DSP is still not going to cut it for general purpose programming and hence why they've included ARM CPUs on their SoCs as well. OMAP3530 has the best of both worlds, and by delegating DSP operation to a second CPU some focused, hand optimized ASM code can be grinded down for it to really deliver high performance when necessary.

Re this PMP: I'd love to know what GBA emulator this is. I doubt it's VBA. Which leads me to believe they either wrote one themselves, got Vampent to port vBagX, or they're using gpSP. The latter is the most free and easy solution so I bet that's what they did. It'd be running off of the interpreter but that's probably enough for this CPU, especially if they hack the timings to be different.

 

[CronoTriggerfan]

I hate to branch a bit off-topic, but does anyone know where the JXD 301s are even still available for purchase? I've been doing some hunting as they'd make a cool birthday gift for a friend, but it seems they've vanished from the interwebz...

 

[Firefox]

I hate to branch a bit off-topic, but does anyone know where the JXD 301s are even still available for purchase? I've been doing some hunting as they'd make a cool birthday gift for a friend, but it seems they've vanished from the interwebz...

They've stopped making them. I believe Craigix sells a very similar Blackfin-based PMP on gbax.com.

 

[Svartalf]

You should stop listening to him, he is bullshitting you or simply doesn't know what he's talking about

Heh... Either that or I misunderstood him when he mentioned the things in passing.

At the time we were more busy with trying to wrap our minds around the client's OpenGL drivers, which were...heh...entertaining, to say the least.

Doesn't matter much, looking at things. I'd say the things aren't even IN the same league as what we're doing here anyhow...

 

[craigix]

They are very hard to track down indeed, the JXD isn't made anymore but we do keep getting similar consoles in stock which are exactly the same inside but the case and controls change.