If you could redesign the PC

[dornick]

what would you do differently?

Or to be more general, what changes would you make to current PC architecture for whatever reasons?

I don't really have any ideas of my own, I just want to hear what you highly technical people would think about it.

 

[Vegitto]

Good question. I'm aiming to designs processors in a couple of years (yeah, as a job), so this might get me started. I would get rid of the heat problem, first, since all CPU's you can get nowadays burn up in 2 seconds if you forget to attach a fan, or if the fan dies, and it sucks.

I'm assuming I'm an AMD engineer here, I'll get to the Intel part later. Here, you can see an image of the core of an Opteron. This is a Troy core, I believe. I'd get a smaller cache (not smaller, as in less than 1 MB, but smaller in size (I.E. not 90nm but 65nm)). Also, I'd improve the memory controller some more, since it sometimes craps out. Maybe some more data cache, and increase the size of the FP unit (floating point). But, this is just the Opteron core. With all AMD parts, I'd like to see the pins (on the processor) on the motherboard, rather than on the processor itself (as with Intel), since I hate ordering someting and missing 1 or 2 pins.

Now, the Intel part. This is a picture of a Prescott core. I want to have less heat, less cache (or lower latency cache) and an ondie memory controller. Also, I'd increase the width of the lines of, since that is what makes AMD king in the gaming area. You see, the depth of the lines makes for the clocking part (I.E. being able to clock high). This architecture is called NetBurst. So, with deep, wide lines, Intel'll be king, once again. Now, the current dualcore Intel solution is not acceptable. For example, the Smithfield core is just two Prescotts, glued together on a single package. They don't even work together. Now, it's rumoured that the Presler would improve this, but we'll have to wait and see.

Sorry for the long rant, but I just felt like it.

Have a good one.

 

[dornick]

thanks, that's interesting stuff.

Everyone else can chime in too because I wasn't referring specifically to the processor design. I'd love to hear any ideas, even stuff as simple as sticking the CD drives in the top of the machine vertically, if you thought that was a good idea.

 

[Ahkorishaan]

I for one would reorganize the PCB in PCI and AGP cards so that all the sexy chippage is facing up, not down.

But that's jsut the vanity in me talking...

 

[MrDudeMan]

Originally posted by: Vegitto

snip

before i say anything, i want to know what you have a degree in or what your current job is.

 

[BillyBobJoel71]

Originally posted by: Vegitto
Good question. I'm aiming to designs processors in a couple of years (yeah, as a job), so this might get me started. I would get rid of the heat problem, first, since all CPU's you can get nowadays burn up in 2 seconds if you forget to attach a fan, or if the fan dies, and it sucks.

I'm assuming I'm an AMD engineer here, I'll get to the Intel part later. Here, you can see an image of the core of an Opteron. This is a Troy core, I believe. I'd get a smaller cache (not smaller, as in less than 1 MB, but smaller in size (I.E. not 90nm but 65nm)). Also, I'd improve the memory controller some more, since it sometimes craps out. Maybe some more data cache, and increase the size of the FP unit (floating point). But, this is just the Opteron core. With all AMD parts, I'd like to see the pins (on the processor) on the motherboard, rather than on the processor itself (as with Intel), since I hate ordering someting and missing 1 or 2 pins.

Now, the Intel part. This is a picture of a Prescott core. I want to have less heat, less cache (or lower latency cache) and an ondie memory controller. Also, I'd increase the width of the lines of, since that is what makes AMD king in the gaming area. You see, the depth of the lines makes for the clocking part (I.E. being able to clock high). This architecture is called NetBurst. So, with deep, wide lines, Intel'll be king, once again. Now, the current dualcore Intel solution is not acceptable. For example, the Smithfield core is just two Prescotts, glued together on a single package. They don't even work together. Now, it's rumoured that the Presler would improve this, but we'll have to wait and see.

Sorry for the long rant, but I just felt like it.

Have a good one.

if you want a good processor, you must invent a super conductor. its the only way to eliminate heat and get super fast processors.

 

[Fox5]

I'd say PCs should have a smaller case.
I'd like expansion cards to lay down flat, rather than stick out to the side. Or even better, just put in the chips directly into the mobo. Airflow should go straight front to back without having many obstructions, perhaps even a path should be made(ducts) to lead the air over the critical parts.

Since I want good air flow, move all the drivers to the bottom of the computer, or lay the computer sideways, leave the drives in the front, and have the air exhaust straight across teh computer.

As far as the processor goes, home PCs would have very centrinoish designs, and only workstations would get the cpus with more computational power and higher heat dissapation.

 

[Vegitto]

Originally posted by: Bigsm00th

Originally posted by: Vegitto

OMFG! You quoted my story! w00t

before i say anything, i want to know what you have a degree in or what your current job is.

No job, no degree. I'm in school, but it's holiday now. Why?

 

[Fox5]

Originally posted by: Vegitto

Originally posted by: Bigsm00th

Originally posted by: Vegitto

OMFG! You quoted my story! w00t

before i say anything, i want to know what you have a degree in or what your current job is.

No job, no degree. I'm in school, but it's holiday now. Why?

You were very vague with your ideas for one.
Sure, it's great to say you'd design it with less heat, but do you really think that AMD and Intel tried to design hot chips? And no way you'll make chips running without a fan considering even some fixed function chips now run hot enough to require a fan, like those in TV sets and set top boxes.
And a denser cache and a more advanced fabrication process? I'm sure if AMD could instantly do that they would. And I'm sure AMD is working on the memory controller as well.
More cache costs money, and the Opteron has more than enough FP power for now, considering no other x86 designs offer as much. And beyond that, dual core will be nearly doubling the FP power anyhow.

Pins on the motherboard? This was something intel did that was whined about because the pins break easier that way. There are upsides to having the pins on the mobo instead, but downsides too, so it's not a clearly superior choice, how about we go back to slot type designs?

And your analysis of Netburst...well it was interesting. BTW, the G5 cpu is deep and wide, yet is generally outperformed by both P4s and Athlon 64s, and very severely so, generally only when Altivec is used does it show worthwhile performance.

Basically you're noticably exceeding your current level of knowledge and it shows.

 

[RaynorWolfcastle]

I personally would ditch the BIOS and move to something more modern for bootstrapping, the BIOS is really an archaic piece of hardware that we can't really get away from. Apart from that, I'd probably want to ditch all the support for legacy x86 instructions, I'm sure there's some die space to be saved by doing that. I'd probably add more registers as well. Really once you change the instruction set, you're not really talking about x86 any more so I don't know if this is a valid discussion.

FWIW, IA64/Itanium is Intel's mid '90s answer to your question

 

[harrkev]

What I think would be completely cool would be to have all-optical interconnects. Your processor talks to your chipset via a little optical cable, Same for RAM, etc. You would not need a motherboard -- just a "power distribution tray."

Of course, other than a major redisgn of the chips, this assumes some major progress in packaging, as well as implementing lasers and optical sensors into the standard silicon process....

On a more practical level, come up with some sort of FPGA API, and include an FPGA or two on each motherboard.

 

[Fox5]

I personally would ditch the BIOS and move to something more modern for bootstrapping, the BIOS is really an archaic piece of hardware that we can't really get away from. Apart from that, I'd probably want to ditch all the support for legacy x86 instructions, I'm sure there's some die space to be saved by doing that. I'd probably add more registers as well. Really once you change the instruction set, you're not really talking about x86 any more so I don't know if this is a valid discussion.

What limitations does the bios bring up? And some companies do have updated bioses, so who knows.

 

[TuxDave]

Originally posted by: Vegitto
Good question. I'm aiming to designs processors in a couple of years (yeah, as a job), so this might get me started.

I just started probably 7 months ago and I still have no freaking clue how the rest of the processor works outside my block that I'm building. It's so monsterous I really wonder if anyone really gets the 'big picture'.

 

[johnnqq]

i wish computers could fit in a 1x1cm box

 

[CrispyFried]

I would go back to S100 style computers. IE the motherboard is just a backplane with slots, you would put CPU/memory on one card, all slow I/O on another card, all storage (HD controllers and such) on another card, video on another etc. All communication between boards would be via the backplane, perhaps super fast serial or maybe even do away with that and use fiber optics. To get more power, just change the CPU/memory card or add a second CPU card to the 1st. Basically with about 16 slots. You could even add drop in different CPU types.

Like this.

Slot Item

1... i86 CPU #1 with memory
2... i86 CPU #2 with no memory
3... Mac CPU with memory
4... Physics CPU
5... HD/Optical I/O
6... everything else I/O
7... Soundcard #1
8... Soundcard #2
9... Video card #1
10.. Video card #2
11.. Ramdisk card
12.. ROM/EEPROM card
13-16 etc

if the backplane were planned out well enough all you would need to do every couple years is add another CPU card and replace the video. should be cheaper than a new mobo/cpu/etc as much stuff that really doesnt change much doesnt get replaced. The CPU/mem card would be smaller, less complicated and thus cheaper.

 

[dornick]

That would be pretty awesome if it worked well.

 

[Vegitto]

Originally posted by: Fox5

Originally posted by: Vegitto

Originally posted by: Bigsm00th

Originally posted by: Vegitto

OMFG! You quoted my story! w00t

before i say anything, i want to know what you have a degree in or what your current job is.

No job, no degree. I'm in school, but it's holiday now. Why?

You were very vague with your ideas for one.
Sure, it's great to say you'd design it with less heat, but do you really think that AMD and Intel tried to design hot chips? And no way you'll make chips running without a fan considering even some fixed function chips now run hot enough to require a fan, like those in TV sets and set top boxes.
And a denser cache and a more advanced fabrication process? I'm sure if AMD could instantly do that they would. And I'm sure AMD is working on the memory controller as well.
More cache costs money, and the Opteron has more than enough FP power for now, considering no other x86 designs offer as much. And beyond that, dual core will be nearly doubling the FP power anyhow.

Pins on the motherboard? This was something intel did that was whined about because the pins break easier that way. There are upsides to having the pins on the mobo instead, but downsides too, so it's not a clearly superior choice, how about we go back to slot type designs?

And your analysis of Netburst...well it was interesting. BTW, the G5 cpu is deep and wide, yet is generally outperformed by both P4s and Athlon 64s, and very severely so, generally only when Altivec is used does it show worthwhile performance.

Basically you're noticably exceeding your current level of knowledge and it shows.

I know I don't know much, but hey, that's what schools are for.
Now, you said that the Opteron had enough FP power already, but it is my oppinion that improvement is always welcome, even when the current one is good enough. The G5 is indeed less strong than current processors (Intel and AMD ones), but it's not based on the x86 architecture, but on the PowerPC architecture (I believe, not sure). By the way, don't trust those benchmarks, because they only work when a program is also designed/coded for Macs.

Also, your idea of slot-type designs tickle my fancy. I hate buying a processor and missing two pins, I'm sure I'd hate buying a motherboard and missing some pins, but slots don't carry any pins. Good thinking on that area.

I also know that AMD is probably working on a smaller fabrication process of the cache, and I don't know how they should make it smaller, but if I had the knowledge, and worked for the company, I most definately would. I also want to get rich, like my dad, but I don't know how. Yeah, my dad wanted to get rich, too .

 

[Calin]

Originally posted by: Ahkorishaan
I for one would reorganize the PCB in PCI and AGP cards so that all the sexy chippage is facing up, not down.

But that's jsut the vanity in me talking...

So they would receive all the dust? I wouldn't (or so I think)

 

[Calin]

Originally posted by: harrkev
What I think would be completely cool would be to have all-optical interconnects. Your processor talks to your chipset via a little optical cable, Same for RAM, etc. You would not need a motherboard -- just a "power distribution tray."

Of course, other than a major redisgn of the chips, this assumes some major progress in packaging, as well as implementing lasers and optical sensors into the standard silicon process....

On a more practical level, come up with some sort of FPGA API, and include an FPGA or two on each motherboard.

Laser LEDs don't have a very long life (I don't think they would go over 5 years of continous use). Compare this to electrical connections on hardware made 10 years ago and still working

 

[randym431]

no wires, at all!

 

[Valkerie]

Originally posted by: Fox5

I personally would ditch the BIOS and move to something more modern for bootstrapping, the BIOS is really an archaic piece of hardware that we can't really get away from. Apart from that, I'd probably want to ditch all the support for legacy x86 instructions, I'm sure there's some die space to be saved by doing that. I'd probably add more registers as well. Really once you change the instruction set, you're not really talking about x86 any more so I don't know if this is a valid discussion.

What limitations does the bios bring up? And some companies do have updated bioses, so who knows.

This is like saying that printers shouldn't have any buttons...

Personally, I'd say the BIOS should be equally accessible from an open OS without even the use of safe modes. it is reasonable to say that BIOS's should be upgraded with more features, way more features.

 

[andrewln]

no mechanical parts....ie) HD, fan, Cd-rom!

 

[Vegitto]

Originally posted by: Valkerie

Originally posted by: Fox5

I personally would ditch the BIOS and move to something more modern for bootstrapping, the BIOS is really an archaic piece of hardware that we can't really get away from. Apart from that, I'd probably want to ditch all the support for legacy x86 instructions, I'm sure there's some die space to be saved by doing that. I'd probably add more registers as well. Really once you change the instruction set, you're not really talking about x86 any more so I don't know if this is a valid discussion.

What limitations does the bios bring up? And some companies do have updated bioses, so who knows.

This is like saying that printers shouldn't have any buttons...

Personally, I'd say the BIOS should be equally accessible from an open OS without even the use of safe modes. it is reasonable to say that BIOS's should be upgraded with more features, way more features.

Yeah, I'd like to be able to access my BIOS from within my OS, but I want it securely. What I don't want is hackers messing with my hardware. For example, if they would mess with my voltages, my CPU/PCI-card/AGP-card/PCIe-card would be a goner.

 

[phantom309]

On the front panel I'd like two switches with little flip-up guards on them like in an old-school airplane cockpit. One ON/OFF and one STANDBY. Flipping the ON/OFF switch would turn the computer ON or OFF. No Sleep Mode, no "Super Energy Saver" mode, no sit-there-and-hold-the-stupid-button-in-forever-then-yank-the-plug-in-frustrati
n mode. ON or OFF.

The STANDBY would activate whatever sleep mode the manufacturer fancies - as long as the system is immediately brought back online when STANDBY is switched OFF.

 

[BitByBit]

On the subject of processor architecture, I suppose only very few here could make any really credible suggestions born of experience, but I suppose there's no harm in having a go.

Pipeline depth

Perhaps the most talked about, and often misunderstood aspect of processor architecture is pipeline depth.
The most common misconception of which being that pipeline depth is directly related to IPC.
Provided you can keep the pipeline full, whether through superior branch prediction, caching algorithms, or faster memory interfaces, IPC will not be significantly affected - something made apparent in Prescott vs. Northwood comparisons.
If pipeline depth and IPC were directly related, then we'd see both the P4 and Athlon using a basic 4-stage pipeline.
This is something to bear in mind when AMD introduces the K10 core, which will almost certainly use a deeper pipeline.
If I was designing a processor, I'd try to evaluate my manufacturing abilities, and determine whether it was possible to achieve high clock speeds without encountering problems such as transistor leakage, which would not make a good case for a deep pipeline.

Execution units

Given the difficulty that exists in trying to extract sufficient ILP to make good use of many execution units, it doesn't make much sense to go much wider than current designs.
This was probably the philosophy of Netburst's designers - it was easier to increase performance through higher clock speeds than through IPC from increased numbers of execution units.
That said, efficient use of SMT can also allow efficient use of multiple execution units, by giving the OOO engine more instructions to pick from.
I'd probably go for a 4-issue design, with heavy emphasis on SMT.
There is also something I was wondering about: General execution units.
Someone on this forum will likely be able to tell me whether this would be possible or not, but I was wondering about the possibility of scraping the idea of dedicated ALUs/FPUs and using execution units that can handle both integer and floating-point code (GUs, if you will).
The biggest drawback of the current design is that the maximum IPC (in non-SMT able cores) is determined by the number of execution units of each type, not the combined number.
So, the K8's maximum IPC is 3, since it has 3 integer units, and 3 floating-point units, and they're not used simultaneously.
If this design was changed to 4 'GUs', then its maximum IPC would be 33% higher, with 33% fewer execution units.

Cache

One of the things that let down the K7 architecure was its slow L2 cache.
Not only did it have a high latency, it had a very narrow data width (64-bit), meaning updating the cache took a relatively long time.
Fortunately, the K7's large L1 reduced the need for a fast L2 most of the time, but when dealing with larger datasets, this became a disadvantage.
AMD seemed to address this problem with the introduction of the K8, however, by reducing its latency, and doubling the data width to 128-bits.
Dothan seems to have the most ideal cache subsystem - a 64KB inclusive L1, with a very fast, very large L2.
Processor architects should probably try to match this design atleast.

Efficiency

One of the reasons Dothan can match the K8's IPC in some areas is the emphasis its designers placed on efficiency through variously technologies, and Micro-ops Fusion foremost among them.
This allows Dothan to achieve an average IPC of 'wider' designs, despite having a narrow execution engine.
I don't know about anyone else, but elegant, clever technologies such as these impress me.
Brains over braun any day.

In spite of all the above however, the future seems to be about TLP rather than ILP.
Both Intel's and AMD's designs will probably become more Cell-like, possibly involving two larger, more powerful cores surrounded by lots of little cores.
In that case, the importance of single-core performance will diminish, but by no means become indecisive.