SAM 460 poor performance, high price

[Louis Dias]

Actually with quake3 only textures are loaded to the graphics card memory.

Everything else is transferred per frame, and this accounts to typically several hundreds of KB of data up to 1MB.

Also since these transfers are synchronous on current OS4 3D system the CPU is busy performing these transfers. MorphOS 3D system uses asynchronous AGP transfers leaving the CPU free to perform other tasks.

Texture mapping is the bulk of what the gpu is doing.  They get loaded once into the gpu at the start of the level.  The level is fixed.  Only the players "move".  The level is much bigger than the players.  The rendering is done on the gpu based the camera position passed from the cpu and player positioning.  Please don't twist things.  The cpu is reading player input and keeping track of the two combatants and bullets.  Once the level starts, the bandwidth used to send data to the gpu is low.  It boils down to the gpu being able to render the scene without dropping frames.

On the DV player, data is constantly being streamed over the to the GPU for displaying hence the bigger improvement in framerates via the PCI bus driver upgrade.

Stop making it an OS issue when it's mostly a hardware issue.  SAM440 is underpowered.  We know this.  SAM460 is much better hardware than only the cpu comparison over the '440 shows.

[Louis Dias]

Oh, can you tell me what cache speed has to do with system bus speed?

Plenty.   That's what the cpu's working with until a flush is required.

[Louis Dias]

This is not how Quake3 works. Quake3 doesn't keep any static data on the GPU other than the textures. The current frame is calculated with the CPU and then this geometry is uploaded to the GPU for rendering. This amounts to considerable traffic. With OS4 this leads into major slowdowns due to CPU being busy uploading the data.


It very much is an OS issue. OS4 3D drivers are very very slow. It has little to do with HW.

This is easily validated by comparing Pegasos 2 with Radeon 9250 first by booting into MorphOS and then AmigaOS4. MorphOS is more than twice as fast as OS4. This is common knowledge.

I fail to see how OS boot speed has anything to do with the price of tea in China.

[Louis Dias]

Those benchmarks do not flush to memory at all, they work inside the cache. The L1 cache performance numbers have nothing to do with memory bus performance.


So if we're talking about the memory performance clearly we're not interested about the CPU internal cache. Or did we again change the subject?

No, but they have everything to do with cpu performance which you cite as poor on the SAM460 though it clearly beats the G4 in a # of aspects.

[Louis Dias]

to run quake3 was missing from my post. Fixed it.

You still argue that it's a HW issue?
Also I still don't understand how memory allocation speed affects the memory access speed.

I suppose if I had an OS4 and MOS box then I could see what you mean.

As for memory allocation speed, when initiating any transfer, is this step not required?  I don't know the size of the transfers in question but let's say they were done in 1k chunks.  The memory must be allocated for the transfer to begin.  If the tests included those requests then they would be skewed towards the OS with the most efficient allocator.

The smaller the chunk, the more exagerated the results of the the transfer.

On a lower level, was the OS4 transfer done in full DDR2 burst mode?  If not you wont' see speeds over DDR1.   There's alot of variables here.

Is the FPGA in the SAM460 the memory interface?  It may not be operating in burst mode yet.

All in all, I think you are prematurely bashing the hardware as an excuse to clandestinely bash the OS it's running on.  So, I repeat: port to SAM460 then tell us how bad it is...

[Louis Dias]

Those benchmarks do not flush to memory at all, they work inside the cache. The L1 cache performance numbers have nothing to do with memory bus performance.


So if we're talking about the memory performance clearly we're not interested about the CPU internal cache. Or did we again change the subject?

Aren't you the one changing the subject?

You called the SAM460 a poor performer.
Test show in some aspects it outpeforms a G4 Mac...
A G4 Mac does not even use DDR2 to my knowledge.  The memory test you posted is questional in the sense that I don't know if the memory controller in the '460 is actually running in burst mode because if it did, I would expect the memory tests to have favored it.

You then tried to blame to OS...which we know is your ultimate goal.

I pointed out that it may at some point outperform a G5 Mac on memory transfers since it used at best DDR2-533 and SAM460 *might* support faster memory (ala DDR2-800 or 1033) and also support faster video cards since is has native PCIex16 support.

So, no the SAM460 doesn't have Altivec, but it still has more than enough juice to hold it's own.

[Louis Dias]

Of course it is not. These memory speed benchmarks only allocate the buffer once, and this time spent allocating the memory is not accounted.


You're seriously confused.


If that is the case then this SAM460 has a lot more problems the issues listed so far...

If the fpga is the memory controller then it can be fixed.  This may have skewed the test to begin with and yes is limiting the OS now.  *now* but perhaps not *later*.

[Louis Dias]

I think I have been quite consistent.

Opinions vary.

I think we're getting to the core of the issue here. You expected something to to be true, made a bogus claim and now are unwilling to back off and admit your mistake.

What mistake?  The Mac uses DDR1, SAM460 uses DDR2.  @ 400 rating DDR1 will outperform DDR2 at the same rating because of latency and overhead.  If SAM460 can use memory faster than DDR2-400 then it will outperform the G4's DDR-400 memory.  These are facts.

From the source I didn't see what memories were used.  From the results, it looks like we are comparing DDR-400 to DDR2-400.

Let's see what happens when faster memory is used in the SAM, shall we?

I didn't bring quake3 to this thread. I merely commented your contradictory and clearly bogus claims.

I didn't bring it into the thread either and wasn't replying to you on it.  I still stand by the fact that the bulk of the processing is due to the work done on the gpu.  You can look at MacOS running Quake 3 and see that on the same hardware it's twice as fast as MOS.  So while yes the driver is not good on OS4, it's still not perfect on MOS and that to see an improvement in the actual video card used will still result in a direct improvement of the framerate more than a mild improvement in the PCI bus.  You simply just spun that into another reason to bash OS4 where as the poster was curious as to what to expect with OS4's Quake 3 port on SAM460 over the SAM440...

[Louis Dias]

According to Applied Micro the AMCC460 does max DDR2-400 (peak speed 3200 MB/s). Mac mini G4 uses 166MHz bus though so DDR-333. That evens up the situation a bit, but for some reason Mac mini G4 still is faster.


Until RadeonHD has 3D drivers the whole question of Quake 3 performance is academic anyway.

And just as a final remark the work in progress MorphOS 3D drivers already run Quake 3 faster than Mac OS X.

From what I read:

400 MHz clock DDR2 memory controller

If this refers to the i/o bus then it's DDR2-800.  The internal memory clock is 200MHz for DDR2-800...

There appears to be some confusion about what it actually supports via the wording.  System designers usually only reference the internal or bus clock and not the "labelling".

Basically, whomever ran the tests should drop in some DDR2-800 (as it won't hurt anyways) and rerun the tests.

Kudos on improved 3D drivers.

[Louis Dias]

in actuality, its a 4x pcie-lane, and once you take protocol overhead out of the 4x lane, it pushes the same amount as the macmini agp 4x slot. so any of the g4/g5 macs that have an agp 8x slot can push more video bandwidth than the sam460.

iirc, pcie 4x lane minus the overhead is about 1000mb/s, agp 4x is about 1033 or 1066 mbs but I dont recall if the agp 4x numbers are with protocol overhead removed.

PCI-E is a serial connection. With data transfer going bi-directional. 16X up and down...while AGP is 8X only one way at a time.

[Louis Dias]

Applied Micro says "up to 3.2 GBps peak data rate". This translates to DDR2-400 (PC2-3200, 100MHz memory bus).

Interesting, that site lists 1.0GHz max.
I wonder if ACUBE downclocked the bus but increased the cpu multiplier to achieve 1.16 GHz...  That would explain the poorer memory performance...

[Louis Dias]

Not sure if this has been mentioned yet, but the hardware interface is far from the be all end all of end results. Athlon64 in its day had much higher memory performance vs ddr2 when using only ddr1. Also, quality of drivers and support within an os can make a huge difference. Using Gallium3d (the eventually to be released gfx subsystem for OS4.x) itself is one example of proof of that. There's cases where a faster gpu is slower than a better supported weaker gpu, both with nvidia and ati/amd. Any benchamrks for mos vs. os4.x while mos is using a previous generation card on a pci interface vs. os4.x using agp on a card one generation newer? Going by the huge differences in performance for 3d between the 2, Id be surprised if there wasnt situations where mos (running same software) will outdo OS4.x, even when using a theoretically much weaker gfx card.

Here you include the software behind the hardware currently.  I was hoping to ignore the OS wars and was just looking at hardware (all else being equal) because nothing is stopping an MOS port except TeamMOS.

Now I actually say this as a fan of all "amiga" options, but I couldnt help refrain from commenting after your "I know better, Im smarter" tangen, which is followed up by inaccurate comments.

Now Ive also seen you comment about G5's being restricted to ddr2-533. One detail you seem to have omitted there is that it's capable of utilizing a dual channel memory controller.

None of this matters as the '460 won't support anything faster than DDR2-400 and it currently only running at DDR2-333 speeds.