Two technical questions about clock speed

[Drizzt]

1) Why Intel and AMD

• with a .15 process
• with more transistors than the parhelia

can reach 2 ghz, while Matrox and Nvidia are stuck on really low frequencies?

2) Why every one try to guess the clock speed of Matrox cards, and no one put the appropriate apparatus on the right pin and measure it?



(for question 1, I know that memory bandwith would be a bottleneck to a 2 ghz Parhelia, but it's just curiosity :-D )

[Umfriend]

Ok, I know less than nothing about these things and probably less, but AFAIK:
1. P4 and Athlon XP have less than 80 mln transistors (64 mln???)
2. A lot of these are for L1 and L2 cache, wheras the 80 mln P transistors are mainly processing units

Having said that, I got your point and I have been wondering about this for quite some time now....

[D0ug]

Re: Two technical questions about clock speed

1)The process and number of transistors have absolutely nothing to do with clock speed. Clockspeed is meaningless unless your comparing two identical chips. Why do you think that an nvidia card is able to do TnL faster in its GPU, then what can be done on the CPU of the computer, even if the CPU is many times faster MHz wise? I dont even want to imagine the cooling a 2ghz parhelia would need either. Have you already seen the size of the heatsink on the pictures of the alpha board that have been floating around?

2) There is no need to guess the clock speed of matrox cards. There are software utilities out there that will tell u what clock speed your card is running. The reason everyone is guessing what speed the parhelia is running at is because no one besides matrox physically has a card to mess around with yet.

Originally posted by Drizzt
1) Why Intel and AMD

• with a .15 process
• with more transistors than the parhelia

can reach 2 ghz, while Matrox and Nvidia are stuck on really low frequencies?

2) Why every one try to guess the clock speed of Matrox cards, and no one put the appropriate apparatus on the right pin and measure it?



(for question 1, I know that memory bandwith would be a bottleneck to a 2 ghz Parhelia, but it's just curiosity :-D )

[D0ug]

I should also add, processors like Intel and AMD's CPUs process data in chunks of 32bits. Whereas Nvidia's GPU processes data in 256bit chunks, and the Parhelia's GPU will process data in 512bit chunks. Per clock cycle the Nvidia and Parhelia GPUs can process more data then an Intel or AMD cpu. This is why i said clockspeed is meaningless between diffrent chips in my last post

From what ive been able to calculate a 150mhz Parhelia 512bit GPU, and a 300Mhz Nvidia 256bit GPU can process the same ammout of raw data as a 2400Mhz 32bit CPU

Heres how i calculated

2400mhz * 32bit = 76800 for the 32bit CPU

76800 / 256bit = 300Mhz for the 256bit nvidia GPU

76800 / 512bit = 150mhz for the 512bit parhelia GPU

I know there are many other factors to the speed of a CPU/GPU, this is just a rough estimate

[Drizzt]

Hey, DOug, what is this?
Paranoids from Nvidiot questions?

Why have I asked for a technical-electronical info, and you've replied with "Mine it's bigger than yours?"

Is there anybody out there who is an electronic expert who can answer to MY question instead of try to save the Matrox economy?

(Umfriend answer, however, seems reasonable :-) )

[Drizzt]

(YES, I'M A BIT ACID TODAY, but it's 4 days I'm without my motorcycle and I will be without it for A WEEK OR TWO)

(And I want a stable relation with a girl only, I'm tired of changing girl every month....grrrrrr)

(I WANT MY BIIIIIIKEEEEE!!!!!!!)



(I want a shotgun, now...)
(really want a shotgun, now...)
(some human target, too...)

[frankymail]

Well, a CPU's pipeline is conceived to make small operations on data at each stage of it's pipeline. The pipeline is very long and a lot of data enters the pipeline, but it spends a lot of cycle just waiting. In a graphic chip however, things aren't as "pipelined": the chip has multiple pipelines that makes "big" operations on each chunk of data that enters it, and the data gets out pretty quickly (there are much less stages then in a CPU); it's like comparing a Honda S2000 V4 2L with a Chevy Caprice Classic Station Wagon V8 5.7L: both do their respective job well, but one take a lot of small chunks, while the other makes big bites but less of them.

Doug: you can't really make a side-by-side comparaison of CPUs and GPUs; A CPU has to be much more versatile than a GPU, because the spread of its tasks is much wider, while a GPU's task are very repetitive; How do you think nVIDIA/ATI can come out each year with a "new" (i'm talking about Radeon->Radeon 8500 evolution style) architecture each year? Because a GPU is much easier to design than a new CPU; to Increase a GPU's power, you can always increase the number of pixel pipelines, add vertexes-processing units and other blocks, but you can't really do that with a CPU. Designing a robot is easy, but designing the control system and AI is much tougher...

C YA

[D0ug]

I am no nvidiot. I am a happy owner of a g400 max. You wanted to know why gpus dont run at 2ghz, and i told you why. If you want even more info, i highly doubt the AGP slot can provide the 50+watts of electrical power that would probably be required for a chip that fast. Im not sure how much power intel's recent chips use. however i beleve the amd athlon is already up into the 70 watt range. The heatsink on a GPU like that would be so big that you would probably loose 2-3 PCI slots on your motherboard.

Originally posted by Drizzt
Hey, DOug, what is this?
Paranoids from Nvidiot questions?

Why have I asked for a technical-electronical info, and you've replied with "Mine it's bigger than yours?"

Is there anybody out there who is an electronic expert who can answer to MY question instead of try to save the Matrox economy?

(Umfriend answer, however, seems reasonable :-) )

[frankymail]

Power might seem like an issue, but it really isn't; there alwasy AGP Pro, an on-board molex conenctor (like some Canopus cards or Voodoo 5 5500), or even an external power supply (Voodoo5 6000 anyone??? ) For the heat issue, there'S always other solutions: larger low-profile heatsink-fan assemblies, heatpipe-heatsink-fan asemblies. But so what if one of the requirement for my next graphics card is to have a free slot next to the AGP connector? On my main workstation, I already have (in addition to the video card) a sound card, two LAN adapters and a TV-Tuner; that leaves me two free slots next to the video card. and you can also that this example: Abit's latest MAX motherboard; the AT7 has only 3 PCI slots and the IT7 has 4 of them; that means space equal to 2-3 PCI slots next to the AGP connector free for any use...

[DGhost]

i am guessing there are several things at play here, but i am not an EE so i dunno...

1) most processors have power/heat saving features. if a part of the chip isn't running, it turns off. last i checked, most video cards don't have this feature, only the mobile parts do.

2) as was said earlier, a chip like the Parhelia or an Nvidia chip does more work per clock than a CPU. maybe in the neighborhood of 4-8 times as much if you count bits. have you ever done calculations as to how much data its pushing around?

1280 x 1024 x 32 = 41943040 bits (5MBytes) - on my monitor thats being refreshed at 60hz, moving about 300MBytes/sec sustained. Its also rendering 5MBytes of data at any given time. thats just from the video core to its memory. thats not assuming it has to do any alpha blending, any special processing of it, etc. thats just flashing a picture. granted, the cards may have bandwidth and processing saving features on them now rather than just brute forcing it.

try using a VESA frame buffer at 1280x1024 sometime and see how it performs when you move windows around, and tell me your host processor is faster than your video card.

and then tell me why a video card would even need to run over 300-500mhz.

3) some FPU instructions on a P4 take 2-3 cycles to finish. whereas to contrast, the GPU on your graphics card is usually able to execute them in a single cycle, or execute multiple instructions in a single clock. in contrast try executing matrix math on a P4 - again, this is something that a GPU should be able to do in a single cycle, but the p4 will have a lot harder of a time doing. SSE/SSE2 might fix this, but i have not looked into it. my basic understanding of physics is that the more work being done by the processor in a cycle means more transisters being used to do it means more components that produce heat due to resistance.

4) smaller die process doesn't affect clock speed directly. small pathways = less resistance = less energy lost as heat = a cooler running chip that consumers less power. if you clock that higher than the older chip, it starts producing more heat again and can then be balanced against the rated temps of the die

as far as transistor counts... the Palomino core as it stands has roughly half the number of transisters that a parhelia does and is on a .18 micron process. the P4 contains about 3/4th of the amount of transistors the parhelia contains, and is on a .13 micron process.

consider one last thing... most video cards have heatsinks/fans on them that would cause an Athlon to burn up within 5 seconds...the thermal guidelines that a video card has to work within are a lot stricter than the ones that a processor does...

again, i am not an EE and really rather have had nothing much to do with hardware design, so forgive me if my comments are off base, inaccurate, or just plain silly...

[Ashley]

Re: Two technical questions about clock speed

Originally posted by Drizzt
2) Why every one try to guess the clock speed of Matrox cards, and no one put the appropriate apparatus on the right pin and measure it?

Because there is no such pin. You can measure the *memory* clock, but not the engine clock.

[Wombat]

Power might seem like an issue, but it really isn't;

Sure it is. How many AGP Pro motherboards are out there? Also, anything you put in has to come back out - go ahead and try to dissipate 50W off of a video card. It would be <I>very</I> difficult.

There's lots of other things that are off in this thread, but I'm not about to lecture on them all. One thing I will mention is that you can't make any conclusions about power just from transistor count. It all depends on what kind of transistors you have, usually how often they switch, and sometimes even what state they tend to stay on. The really fast logic tends to burn the most power. Also, shrinking to a smaller process really tends to hurt you in the power area, as transistors dissipate more leakage, and your interconnects bite you.

[Marshmallowman]

CPU's are sequential processing units(mostly), they have caches so the can they are not limited to much by limited memory speeds. They have to process one instruction after another, and those instructions are complex

whereas a GPU in a video card is a very parellel processing setup . It processs big chunks of data in one unit and is tightly linked with the memory interface speed since it is processing bulk memory dataand the processing done is simpler and more uniform than on a CPU

The processing that occurs in a GPU is all about throughput, and that is easily increased just by adding more parallel processing units. Where as CPU can only process one instruction at a time because of the very nature of computer programs(they are sequential)

Note: modern CPU do actualy process several instructions at the same time (and out of order as well) and do have a lot of "parallel" procesing going on but it is not even approaching the level of parellelism that is possible on a GPU

I would think the number of "electronic switches" per second is similar for a GPU and a CPU, its just that in A GPU they are all hapening at the same time.

[R.Carter]

Regarding issue #2 of measuring the clock generator directly. Unless they have a test point, it isn't always easy to measure, especially for BGA parts as there are no pins to measure.

Also, any external clock frequency will usually have a multiplier applied to them internally. So even if you could measure the clock input into a part, that will not always tell you how fast it is running internally.

[R.Carter]

Stolen from Matrox reveals Paharhelia clock speed.

"Now that production preparations are further along, a Matrox representative could tell us that "we are looking at [a clock speed] range in the low 200s, and 220MHz is a reasonable estimate at present." The greater the complexity of a chip, the harder it is to achieve high clock speeds. It's understandable that larger, next-generation chips like the .15-micron, 80-million transistor Parhelia won't match speeds with smaller chips or even future chips produced with a denser .13-micron process."