Quote:

Hammer wrote: @BigGun Quote: The Amiga chipram always was so fast that you could put a program in chip mem. "Chip Ram" is like UMA i.e. unified memory architecture as used in PC's IGPs and XBOX. 68K's "Fast Ram" is noted to be faster than "Chip Ram" hence "Fast Ram" label i.e. shared memory architecture is ussually $uck compared discrete memory architecture. UMA is used for cost reduction reasons e.g. low end Amigas, low end PCs and XBox.

Chip RAM was faster than the CPU could access memory on A1000/A500/A2000 -
as long as you were not using too many bitplanes.
The Natami memory will too be faster than the CPU can use it.

Quote:

Quote: This is something that you would not want to try in the PC world. Erm, note why the modern PC has cache coherent hardware features i.e. you have DMA operations from non-CPU sources writing to main memory.

You are missing the point.
How many hundred NS does innitiating a DMA take on your system?
The point is that the Natami has a unbeatable low latency on memory access.



Quote:

Quote: The Natami chipram is faster than the 2nd Level cache of your x86, PPC or CELL CPU Define faster? Any memory test benchmark results?

PC's "RightMark Memory Analyzer" benchmark test memory in random mode.
One of the methods to minimise latency in the modern X86s and PPCs is to implement pre-fetch techniques.
[/quote]

You miss the point.
The point was that if you have code that sets single pixels on the Natami.
Like a Star Routine, Texturemapper, or Voxel routine
then the low latency of the Natami is a major advantage over any other GFX card.


Quote:

Lower latency is not a silver bullet, e.g. AMD K8’s lower memory latency didn’t save it from being crushed by Core 2 i.e. Core 2 has speculative pre-fetch for both instructions and data.

The comparison with K8 is very weak -
The Natami has not much lower latency than the K8 memory !
The Natami has even lower latency than the AMD "on chip cache" !

The point was to outline some advantages of the Natami Chipmem over using a normal a PCIe GFX card.
Your example of "speculative prefetching" is a good example what does NOT work over PCI at all.
GFX memory has by design to be "uncacheable" or you can not use the blitter!
And NOT cacheable memory can off course never be cache prefetched.

Quote:

Quote: Then 800 means that this memory can transmit 1 word in 800 MHz in a burst. That 800Mhz is not a true 800Mhz i.e. 400Mhz double rate. Mainstream architecture leans towards streaming computation. To lower latency, some modern PC NB and MCHs has pre-fetch techniques e.g. nVidia's nForce 600i (with DASP 4.0) Series For Intel and Intel 840. nVidia's nForce 600i's DASP has both NB cache, NB pipelining and NB pre-fetch techniques. nForce 680iSLI with DDR2 800(4,4,4,T1) has 27.2 ns @400Mhz. **Intel 965**, 975**, P35**, P38** has Intel FMA (Fast Memory Access) i.e. out-of-order processing. http://www.intel.com/products/chipsets/q965_q963/demo/demo.html In real world application benchmarks, there's very little gain in nForce 600i's DASP vs Intel 965.

I'm not sure what you are trying to prove.
So you are saying that the Natami with 10ns is about three times faster than the nVidia's nForce ?
You are probably right.
I think there is no question that the Natami its fast.

One thing is fact:
While CPU clock rates did increase a lot in the recent years, the latency of DRAM did not increase at all.
If a memory program is latency bound then it runs at a fraction of the CPU clockrate.


For example a typical pointer chase algorithm on your 3Ghz x86 runs at memory latency speed.
In other words such an algorithm will run with max 25 Mhz !!!

If you have a low latency design as the Natami then you will have a much better performance for such cases. The Natami will easely run 3-4 times faster than your K8 in such case.

The normal way of trying to hide the latency is as you correctly pointed out prefetching.
This is the reason why many CPU have huge cache lines sizes of 128 byte or more.
But prefetching does not always help. E.g If you are pixeling a Voxel screen then you draw the screen in colums, prefetching a whole cache will not help but hurt performance majorly.

And if you are manipulate GFX memory both with COU and blitter then the GFX mem is uncacheable by design.

In sort: Todays CPUs try real hard to increase speed by using prefetching and huge amounts of cache. Prefetching often helps but sometimes it can NOT help and sometimes it will even hurt performance a lot. On 68k and Coldfire both the positive and negative effects of cache line bursting aresmall as the cache lines as small too.
I think the Natami goes the best possible direction for improving performance for an Amiga like working system.


Cheers
Gunnar

_________________
APOLLO the new 68K : www.apollo-core.com

@BigGun
.Quote:

Either way you do it : The theoretical very high performance of the high clocked memory on your PC GFX card can not be used properly.

http://www.digit-life.com/articles2/video/g80_units2.html

Note;
shader_busy (avg),
shader_busy (max),

Most shader idle percentage is less than 25 percent.

Quote:

If you have a PCIe GFX and want to set pixels with the CPU then you need to create the whole frame in mainmemory and then copy the whole frame to the GFXcard.

The whole point of having DX10/OpenGL2.x (ICDs from ATI/NV) is to move gfx workloads to the GPU.

Quote:

If you try to do this on PCIe then your performance will be much slower than A500 .

My Gfx’s VIVO and memport are functioning just fine. In OfficeFX Pro, realtime effects DX3D are sent back to the main memory (for storage), it's faster than A500 and it’s working at HD 720p and HD 1080p resolution.

Last edited by Hammer on 22-Jan-2008 at 02:27 PM.

_________________
Ryzen 9 7900X, DDR5-6000 64 GB RAM, GeForce RTX 4080 16 GB
Amiga 1200 (Rev 1D1, KS 3.2, PiStorm32lite/RPi 4B 4GB/Emu68)
Amiga 500 (Rev 6A, KS 3.2, PiStorm/RPi 3a/Emu68)

@BigGun
Quote:

Like a Star Routine, Texturemapper, or Voxel routine

Where's FPS bench?

Quote:

While CPU clock rates did increase a lot in the recent years, the latency of DRAM did not increase at all.

Benchmark AMD64 DDR266 vs AMD64 DDR400 vs AMD64 DDR-II 800 vs AMD64 DDR-II 1066 in random access mode. You will find that the overall bandwidth increased.

Quote:

Your example of "speculative prefetching" is a good example what does NOT work over PCI at all.

Nope, refer to software prefetching i.e ATI binary driver automatically handle this.
Note the reason why you have shared gfx memory and dedicated gfx memory.

Refer to texture prefetching from shared gfx memory to gfx on-board memory.

Quote:

This is the reason why many CPU have huge cache lines sizes of 128 byte or more.

Intel Core 2 has L1/L2 cache lines of 64 bytes.

Quote:

And if you are manipulate GFX memory both with COU and blitter then the GFX mem is uncacheable by design.

In PC IGP (and lesser extent GPU), refer to cache coherency hardware feature.

Last edited by Hammer on 22-Jan-2008 at 02:33 PM.
Last edited by Hammer on 22-Jan-2008 at 02:24 PM.

_________________
Ryzen 9 7900X, DDR5-6000 64 GB RAM, GeForce RTX 4080 16 GB
Amiga 1200 (Rev 1D1, KS 3.2, PiStorm32lite/RPi 4B 4GB/Emu68)
Amiga 500 (Rev 6A, KS 3.2, PiStorm/RPi 3a/Emu68)

@BigGun

Quote:

Chip RAM was faster than the CPU could access memory on A1000/A500/A2000 - as long as you were not using too many bitplanes.

Actually ......

The 68k can access RAM with atleast one cycle in between, with chip-mem it only gets every 2nd cycle (unless the chipset demands more), meaning it allready pushes the RAM to it's limit.

From time to time an operation would take an odd number of cycles to finish, with chip-mem the CPU will have to wait one cycle before it is it's term again, while it would get instant access to any (real) fast-mem.

So yes chip is slower than fast, even on a plain 68000, it's just that the difference is barely noticeable.

_________________
- We don't need good ideas, we haven't run out on bad ones yet
- blame Canada

@BigGun

Quote:

Chip RAM was faster than the CPU could access memory on A1000/A500/A2000 - as long as you were not using too many bitplanes.

What happens if AOS2 with 640x512i workbench screen and 16 colors on ECS?

Quote:

You are missing the point. How many hundred NS does innitiating a DMA take on your system?

My point was for
"This is something that you would not want to try in the PC world."

My answer was; the PC world is using (not trying) UMA for low end PC systems.

Can NatAmi AGA++ beat Intel GMA X3100 in OpenGL (shader)?

_________________
Ryzen 9 7900X, DDR5-6000 64 GB RAM, GeForce RTX 4080 16 GB
Amiga 1200 (Rev 1D1, KS 3.2, PiStorm32lite/RPi 4B 4GB/Emu68)
Amiga 500 (Rev 6A, KS 3.2, PiStorm/RPi 3a/Emu68)

@Kronos

Quote:

And what would be the point ?

Dual boot systems: AmigaOS4 / MorphOS / AROS / AmigaOS3 / Linux PowerPC.

Quote:

Putting 2 (different CPUs) onto one board is pure nonsense, as best seen by the way the 603/604s are crippled on the P5-cards. Now imagine how bad it would be for a modern PPC to be dragged down by an 68k.

Yes but you only need one CPU, you can put E-UAE cpu core in side a flash, allowing to run the classic OS3.x software whit just one CPU, you probably get close to 060 / 50mhz speed on modern PowerPC, even if the assembler code is interpreted, because there is no chipset emulation in software, fast memory will extremely fast when using the onboard memory one accelerator.

Quote:

Combining a PPC with AGA also doesn't make sense, as you still won't be able to run games on it directly and if I'm useing UAE I have no need for AGA. PPC-"Amigas" should be legacy free boards like the Pegs, the SAM and the Efika.

I fail to see your point of view.

Quote:

On top of that we have the legal death-lock surounding OS4, pretty much invalidading you statement about development being more PPC than 68k at least for the moment.

It’s just matter of designing hardware that is almost compatible whit BizzardPPC or CyberStorm PPC.

_________________
http://lifeofliveforit.blogspot.no/
Facebook::LiveForIt Software for AmigaOS

@Kronos

Quote:

Kronos wrote: @BigGun Quote: Chip RAM was faster than the CPU could access memory on A1000/A500/A2000 - as long as you were not using too many bitplanes. Actually ...... The 68k can access RAM with atleast one cycle in between, with chip-mem it only gets every 2nd cycle (unless the chipset demands more), meaning it allready pushes the RAM to it's limit. From time to time an operation would take an odd number of cycles to finish, with chip-mem the CPU will have to wait one cycle before it is it's term again, while it would get instant access to any (real) fast-mem. So yes chip is slower than fast, even on a plain 68000, it's just that the difference is barely noticeable.

The 68000 could never use all cycles of the chipmem.
As an 68000 needed 4 clock per instruction word, or loaded operation-word anyway.
The chip mem was twice as high clocked as the usable bus transfers that an 68000 could do.

_________________
APOLLO the new 68K : www.apollo-core.com

@Hammer

Quote:

Hammer wrote: @BigGun Quote: Chip RAM was faster than the CPU could access memory on A1000/A500/A2000 - as long as you were not using too many bitplanes. What happens if AOS2 with 640x512i workbench screen and 16 colors on ECS?

What in "as long as you DO NOT USE TOO MANY bitplanes" do you not understand?
Everybody knows that you could use for free 4 planes in lowres or 2 plane if hires.
Are you just arguing for the sake of it?


Quote:

Quote: You are missing the point. How many hundred NS does innitiating a DMA take on your system? My point was for "This is something that you would not want to try in the PC world." My answer was; the PC world is using (not trying) UMA for low end PC systems. Can NatAmi AGA++ beat Intel GMA X3100 in OpenGL (shader)?

Can your Intel run Amiga OS-3 ?

The point of the Natami is recreating an original Amiga which feels and behaviors like an original,
runs OS 3, boots in 2 second, is passive cooled, but is much faster than the original and has more possibilities.

If you see a better way of creating an improved AGA chipset then please speak up.
If you are more interested in a System for running Windows (as your posts are implieng) then maybe this is the wrong website for you?

Cheers

[correcting quotes]

Last edited by BigGun on 22-Jan-2008 at 06:12 PM.

_________________
APOLLO the new 68K : www.apollo-core.com

@BigGun

I am with you man!

Gives the new Amiga! Yeaaah!

I am getting carried away sorry!

Waiting impatiently for more news, pictures and videos of the new Amiga!

_________________
www.loriano.pwp.blueyonder.co.uk

@NutsAboutAmiga

Why put the "E-UAE cpu core" into the flash, when one can just run EUAE ???

This whole idea to seperate the cpu-emu and have it running against a "real" Amiga-chipset sounds like a nighmare to implement, and would probraly be far less compatile as a straight EUAE-port.

Copying the P5-cards is another PIC ..... the design was flawed back then and would be even worse now. Think about it, you would limit yourself to 603-based cores. And there is the question on how to boot OS4 on this ... does it need the 68k for startup ? Maybe even parts of the (copyrighted) Phase5-flashROM ?

You would also need to mimmik one of the supported HD-controllers, either dog-slow PIO-only IDE or extremly expensive UW-SCSI.

And than your gonna stick a GFX-card into it as those modern OSes aren't much fun even with an updated AGA.

Sticking an PPC into a MiniMig/Natami/CloneA makes just about as much sense as sticking a 68060 into C-One ...... you know you could run a 6502-EMU on it

_________________
- We don't need good ideas, we haven't run out on bad ones yet
- blame Canada

@Kronos

Quote:

Copying the P5-cards is another PIC ..... the design was flawed back then and would be even worse now. Think about it, you would limit yourself to 603-based cores.

Classic OS4 runs on 603 and 604, but PowerPC assembler code is mostly genetic, how ever there is a HAL to take in count, I don’t remember the specifics, this it has some thing to do whit interrupts etc.

Quote:

And there is the question on how to boot OS4 on this ... does it need the 68k for startup ?

680x0 is turned off after the kickstart modules are loaded.

The kickstart modules will need to be loaded using the E-UAE CPU emulation.

Quote:

Maybe even parts of the (copyrighted) Phase5-flashROM ?

Don’t know maybe, (or some thing that looks compatible.)

Quote:

You would also need to mimmik one of the supported HD-controllers, either dog-slow PIO-only IDE or extremly expensive UW-SCSI.

Way not simply write your own drivers, you can install the driver after OS4 has booted on the dog-slow PIO-only IDE, possibly implemented in FPGA.

Quote:

And than your gonna stick a GFX-card into it as those modern OSes aren't much fun even with an updated AGA.

Well If its going to have AGA+ chipset whit chunky support, then there is no need to stick any other graphics card in to it, all you need is Truecolor P96 driver for it.

Last edited by NutsAboutAmiga on 22-Jan-2008 at 07:12 PM.
Last edited by NutsAboutAmiga on 22-Jan-2008 at 07:11 PM.

_________________
http://lifeofliveforit.blogspot.no/
Facebook::LiveForIt Software for AmigaOS

hello
in context of new hardware i just wanted to link ur attention to some german forum
the subject of the thread is a replacement board for a1k with better specs built privately by someone there. unfortunatelly its all in german but please regard the benchmarks.
http://www.a1k.org/forum/showthread.php?t=8947

as 4 natami: a faster and bigger aga style genlock compatible graphics interface would be of much interest to me still some years ago, but now how about an a4k multiple-xeon turboboard with hardware jit 86k-code interpreter built in?

Quote:

It’s just matter of designing hardware that is almost compatible whit BizzardPPC or CyberStorm PPC.

Sorry, but if AInc wins they will bury the sourcecode and stop the sales of AOS 4.0. You then can only get it used or via .torrent - and no updates except through spatch...

Quote:

And than your gonna stick a GFX-card into it as those modern OSes aren't much fun even with an updated AGA.

If i read BigGun's postings right, people will not feel the need to do that. It seems we have till summer to talk about first benches of 68060 based development systems...

Quote:

Sticking an PPC into a MiniMig/Natami/CloneA makes just about as much sense as ...

Depends on how fast these things are

_________________
<- Amiga 1260 / CD ->
Looking for:
A1200/CF CFV4/@200,256MB,eAGA,SATA,120GB,AROS

@BigGun

Quote:

The 68000 could never use all cycles of the chipmem. As an 68000 needed 4 clock per instruction word, or loaded operation-word anyway. The chip mem was twice as high clocked as the usable bus transfers that an 68000 could do.

Only even cycles were assigned to 68000 and blitter DMA could also steal memory cycles from 68000 (or 68000 from blitter). Even when 68000 had precedence blitter would not release the bus until 68000 was unsatisfied for three consecutive memory cycles.

Nevertheless it can be said that 68000 run (nearly) at full speed on chip ram only machine if there were not too many bitplanes in use.

_________________
Amiga Developer
Amiga 500, Efika, Mac Mini and PowerBook

@Donar

Quote:

Sorry, but if AInc wins they will bury the sourcecode and stop the sales of AOS 4.0. You then can only get it used or via .torrent - and no updates except through spatch...

He he, you think Amiga Inc is going to come knocking down my door to get my bountiful AmigaOne-XE 1Ghz 512Mb ram? I don’t think so

_________________
http://lifeofliveforit.blogspot.no/
Facebook::LiveForIt Software for AmigaOS

@itix

He he, all the talk about biplanes reminds me about some old Z80 trainer, member it was able to display 2 colours, green and black

The thing was basically designed like this the clock was divided, so that video card was able to clock trough the pixels in sequence that did not interfere whit the CPU, I don’t remember any thing more about it, but Amiga500 / OCS is most likely a lot more advanced

_________________
http://lifeofliveforit.blogspot.no/
Facebook::LiveForIt Software for AmigaOS

Quote:

He he, you think Amiga Inc is going to come knocking down my door to get my bountiful AmigaOne-XE 1Ghz 512Mb ram? I don’t think so

Where did you actually see that i wrote something like that???

_________________
<- Amiga 1260 / CD ->
Looking for:
A1200/CF CFV4/@200,256MB,eAGA,SATA,120GB,AROS

@Donar

Well in your comment you stated “You then can only get it used or via .torrent” and you quoted me, as I already have OS4, I’m not going to go-around pirating software unless some one like Amiga Inc comes takes it from me and if I can’t buy it from computer store or from Hyperion / ACube-Systems.

Last edited by NutsAboutAmiga on 22-Jan-2008 at 08:45 PM.

_________________
http://lifeofliveforit.blogspot.no/
Facebook::LiveForIt Software for AmigaOS

@BigGun

Quote:

So you are saying that the Natami with 10ns is about three times faster than the nVidia's nForce ?

A 100Mhz with 64bits wide bus has a theoretical 800MB/s of bandwidth. Now use RM memory benchmark on modern PCs. You will find that 800MB/s of bandwidth is an easily target to hit.

Quote:

For example a typical pointer chase algorithm on your 3Ghz x86 runs at memory latency speed. In other words such an algorithm will run with max 25 Mhz !!!

Depends on the hardware involved. Google “predicate pointer” patents.

_________________
Ryzen 9 7900X, DDR5-6000 64 GB RAM, GeForce RTX 4080 16 GB
Amiga 1200 (Rev 1D1, KS 3.2, PiStorm32lite/RPi 4B 4GB/Emu68)
Amiga 500 (Rev 6A, KS 3.2, PiStorm/RPi 3a/Emu68)

@BigGun

Quote:

What in "as long as you DO NOT USE TOO MANY bitplanes" do you not understand? Everybody knows that you could use for free 4 planes in lowres or 2 plane if hires. Are you just arguing for the sake of it?

I don’t recall running my Amiga500 like a monochrome Mac.

Last edited by Hammer on 23-Jan-2008 at 06:58 AM.

_________________
Ryzen 9 7900X, DDR5-6000 64 GB RAM, GeForce RTX 4080 16 GB
Amiga 1200 (Rev 1D1, KS 3.2, PiStorm32lite/RPi 4B 4GB/Emu68)
Amiga 500 (Rev 6A, KS 3.2, PiStorm/RPi 3a/Emu68)