Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Evans]

* Significant improvements in the amount of low-level cache; this is Intel's responsibility, and it's not likely to get huge jumps.  Theoretically, you could design a variant CPU package that would normally have a large number of compute cores, but instead had only a few cores swimming in a huge L2 / L3 cache.  It would be a niche product, however. 

This discussion got me thinking.
Isn't that what graphics card actually are? Fast, specialized chips with a huge, fast RAM next to it for efficient graphics processing.
From what you guys are saying, GTX 1080 with 8gb onborad ram would run DF faster (say, as a dedicated machine) than an i7 laptop with 16gb ram half a motherboard away.

Again, we don't know for sure whether DF is thus constrained, or more likely what fraction of the time DF is thus constrained.  And this will depend on the rest of the computer as well.  It's entirely possible that DF is CPU limited on weaker processors, and RAM access limited on stronger ones, for a very simple example.  And anything that gets anywhere near a disk, even a SSD, is so ridiculously slow by comparison that even momentary glitches there can have significant effects on overall speed; this discussion assumes you already have enough system RAM to store all of DF's code, all of DF's working data, and all of your operating system, web browser, music player, and whatever else in RAM; otherwise they will be stealing precious slices of memory access bandwidth with the very slow process of paging or streaming.

I have i7-4710HQ with 16gb RAM running at 1600mhz.
The OS might be an issue however, as this is under win7 pro 64bit version.
I am not upgrading to any newer OS any time soon, but I can consider purchasing a dedicated win8/10 machine to run only DF at some point in the future.

[wierd]

now df would need to be compiled for opencl or some other gpgpu language.

even the GPU has its ram outside of the vector processing unit. for the same reasons.

additionally, now instead of a call to fetch from ram, we also have to fetch over the io bus when communicating with the card. (we have to move the mapped memory window around.)

the ideal chip to run DF on, is a single core chip, clocked at 3ghz or faster, that is swimming in an ocean of cache.

[Evans]

I can't recall where, but I read once that someone somewhere linked a bunch of fast GTX cards to create some sort of processing machine.

Don't remember what it was done for, but it was something along this concept that I was thinking about.

[Bumber]

I can't recall where, but I read once that someone somewhere linked a bunch of fast GTX cards to create some sort of processing machine.

Don't remember what it was done for, but it was something along this concept that I was thinking about.

It was probably doing multi-threaded calculations. No good for DF.

[Evaris]

We come full circle back to the idea of an APU with fast data access via something like HBM connected via interfacer die whole SoC thing.  In otherwords, AMD's Zen on an HBM interfacer would do the trick. 

Here's to hoping for them throwing in HBM on an embedded system for the sake of DF and RAM-speed-intensive niche roles.      (Like that would ever happen.  The only reason Zen would have the possibility is that it has the complete chipset on-die as it is, so an SoC solution should be doable, if they wanted to do something with that route.)   

[wierd]

if intel adopts it (wont be the first innovation from AMD they ended up adopting), then it will end up in pretty much all of intel's embedded platforms.  Currently, low power, high performance is the niche intel is slowly carving for itself, at least in the x86 chip arena. (ARM has it beat still, but intel has more oomph.)

Still, total spanning tree distance with HBM will still start to run into issues. there will be sensible limits to how many modules you can stack onto the interposer before there is significant performance degradation.

[Evaris]

if intel adopts it (wont be the first innovation from AMD they ended up adopting), then it will end up in pretty much all of intel's embedded platforms.  Currently, low power, high performance is the niche intel is slowly carving for itself, at least in the x86 chip arena. (ARM has it beat still, but intel has more oomph.)

Still, total spanning tree distance with HBM will still start to run into issues. there will be sensible limits to how many modules you can stack onto the interposer before there is significant performance degradation.

I honestly don't see this as an issue in the consumer space, given HBM 2 is launching next month with a rated capacity of 32GB of RAM on an interposer without excessive performance loss, and 16GB/s for full performance, I think it would be just fine until the 3.0 or w/e spec comes out.  At least for embedded / mobile / AIO solutions. 

But given HBM's capacity for 256GB/s data transfer rate @ 16GB, versus DDR4-3200's 25GB/s data transfer rate, well... it might be noticeable for DF. 

[Evans]

That would be amazing.
I never considered APU before, but seeing what you guys writing here I might actually start paying attention now.

Also Wierd, I read somewhere that AMD is developing some sort of cpu-in-ram to speed up calculations. Can't find the article now, maybe more info will show up over time.

Things are looking good.

Maybe in the next decade we will have machines with ultra-fast RAM next to CPU that will run DF with decent FPS

[wierd]

That sounds like the HBM previously mentioned.  Basically, the memory bus is literally millimeters away from the logic inside the chip, accessed through a simple but fast switching circuit. By being so close to the CPU, it can be driven much faster.

and an APU (Arithmetic Processing Unit) has been around forever... It's built in any x86 chip newer than a 486. (Those were the last chips where the APU could be sold separately.)  It was the difference between an SX (without) and DX (With) chip.  For something like DF, not having the APU (aka, math-coprocessor) would be crippling.

CPUs contain ALUs (Arithmetic Logic Unit), which they use to perform logic operations (OR, AND, NAND, XOR, IF THEN ELSE, etc.) against data stored in their registers. An APU is a very special set of ALUs that specializes in doing floatingpoint arithemetic (Division, multiplication, exponentiation, etc.)

In the old days (of the old DOOM, and pals), clever hacks were devised to perform floating point arithmetic without one of these special co-processors, because not everyone had them. (There were literal co-processor sockets on motherboards for them, if you did not have a DX chip! This is what they looked like)   These days everyone has one built into the CPU.  Intel no longer sells them as an "upgrade."

[Evaris]

That sounds like the HBM previously mentioned.  Basically, the memory bus is literally millimeters away from the logic inside the chip, accessed through a simple but fast switching circuit. By being so close to the CPU, it can be driven much faster.

and an APU (Arithmetic Processing Unit) has been around forever... It's built in any x86 chip newer than a 486. (Those were the last chips where the APU could be sold separately.)  It was the difference between an SX (without) and DX (With) chip.  For something like DF, not having the APU (aka, math-coprocessor) would be crippling.

CPUs contain ALUs (Arithmetic Logic Unit), which they use to perform logic operations (OR, AND, NAND, XOR, IF THEN ELSE, etc.) against data stored in their registers. An APU is a very special set of ALUs that specializes in doing floatingpoint arithemetic (Division, multiplication, exponentiation, etc.)

In the old days (of the old DOOM, and pals), clever hacks were devised to perform floating point arithmetic without one of these special co-processors, because not everyone had them. (There were literal co-processor sockets on motherboards for them, if you did not have a DX chip! This is what they looked like)   These days everyone has one built into the CPU.  Intel no longer sells them as an "upgrade."

The modern term for APU is "Accelerated Processing Unit" (basically SoC-lite, any x86 CPU with GPGPU/GPU compute compatible graphics cores on-die.  Modern Intel CPUs have been called APUs as well here and there, so.. yeah) not arithmetic logic unit.  Blame AMD's marketing :p

[wierd]

freaking markters.


cant they just call it something sensible, like an integrated device core (IDC), or something?  do they really need to muddy the waters by reinventing what an established industry term means?

oh, right-- marketers are masters of the spin, and fancy words like "advanced!" are crack cocaine to them. (rolls eyes).

[Dirst]

do they really need to muddy the waters by reinventing what an established industry term means?

Sometimes it helps comprehension... like when DOS went from its original technobabble Disk Operating System to what had been its real meaning all along, Denial Of Service

[wierd]

No no no..

That is the difference between DOS (Disk Operating System)
and DoS (Denial of Service)

  (And DOS never really denied you service. It just expected you to know what you were doing. It had no real security to speak of other than a read only flag in the filesystem that literally anyone could turn off at any time.)

[Evaris]

So for a particular note; AMD's upcoming Raven Ridge has apparently been confirmed to have SKUs with 16GB of HBM 2.0 on-interposer with the APU.  Albeit at half-data transfer rates (128GB/s vs 256GB/s, still 5x faster than DDR4, but.. yeah)


So, may be investing in a low power PC of some sort purely for DF in the next year, lol. 

[wierd]

Fun fact:

I learned that a Western Digital "MyBook" can be repurposed into a very swanky ABS shell for tiny form factor PCBs.

Let me get pictures of mine real quick.

Here we go-- Little HTPC for waching movies on in the living room.

(wrapped in spoilerite!)

Spoiler (click to show/hide)

Here is the innards with the outer shell removed.  It's a very tight fit, but the Minnowboard core, the power extension cable, the Mini-HDMI extender, a targus USB Hub and USB extender cable, spinny 2.5" 1TB SATA disk, and several USB expansions (Wifi, Wireless keyboard/mouse fob, and bluetooth dongle) all fit inside the case.

Front


Back


As you see here, I nibbled out a little window for a cooling fan. The intel Atom chip inside needs more than the passive cooling, especially when packed this densely.


and here is the back. Note how I repurposed the key lock antitheft hole for the mini HDMI output, and routed an external USB port to the USB connector.


These bad boys (the chasis) are designed for a 3.5" HDD. They have rubberized shock resistant mounts that just sit inside little plastic cups to hold the drive.  What I did was yank the drive out (I needed a spinny disk in a hurry. Dont ask.) and replace it with a bay converter for a 2.5" drive. I used the bay converter as a mounting plate for an Intel Minnowboard. Minnowboard supports a REAL SATA interface, which is why it has that spacious 1tb spinny disk inside.  I could have put a sata SSD just as well.)