Slashdot Mirror
New VIA x86 CPU Takes Aim At Intel Silverthorne
Kaz writes "While not operating on the same scale as the two major CPU designers, VIA has been gaining traction in the world of UMPCs and thin clients with its Eden and C7 lines of processors. While past architectures have been considerably out-of-date in terms of modern features, the new Isaiah architecture looks to be very competitive with what AMD and Intel have lined up for future ultra-mobile products. It features an out-of-order, superscalar execution core, 64-bit support, virtualization, and even SSE3 — all on a 94M-transistor, 65nm process die. The initial offering will be single-core only, though VIA says that multi-core ability is already designed in. Is Isaiah going to replace your Core 2 system for gaming? No, but it might give Intel's Silverthorne a run for the money."
The next big step in integration is integrated memory. Cache memories are consuming most of the die in your typical high-performance CPU, these days. If you can find a CMOS-compatible, high-density (e.g. - SRAM's six transistors per cell is toooo big) memory technology, then we're going to be at the point where we can simply replace the cache memory with on-board memory. If said on-chip memory technology is nonvolatile, then we're talking panacea cakes, batman.
Naturally, this will first occur in low-performance devices where huge amounts of memory are not necessary. Then, it will work its way into the PC and up from there.
This is why Intel is divesting itself of discrete memory technologies - they don't want to be holding the bag when they're obsoleted by on-chip memory.
SPU manufacturers had better be ready for this because discrete CPUs will be going the way of the horse and buggy if anyone can ever do such a thing.
More
Then if I read right they go on to say Isaiah will be similar. Sorry, but that's not even in the same league as Silverthorn. Silverthorn will be more like a sub 5 watt product. If this is right, they'll be competing against Core 2 processors and performance won't even be close.
I want a fully recharged UMPC after two minutes of Seiko Kinetic shaking, baby!
"He Who Dares Wins"
While past architectures have been considerably out-of-date in terms of modern features
They may not be bleeding edge, but their Eden processors used to compare very favorably to Intel's low-power chips, and have unique features like Padlock accelerated encryption (which is supported at least partially by the Linux kernel to accelerate cryptographic stuff.) Padlock made it possible to have a very low power VPN server..
The only real problem I've had with the VIA processors has been availability, pricing, and cheesy 3rd party motherboards. Mini itx dot com for example wants to bend you over backwards for some pretty old systems; the latest stuff you practically need to take out a mortgage from. You can't really buy the boards from but a handful of places. VIA also seems to be ignoring the networking market (if they sold a low-power board with 3 gigabit ports, they'd put Soekris out of its misery once and for all- overnight.)
Same thing with AMD's low-power Geode (which is plug-compatible with certain athlons.) You can't buy them anywhere except bundled with really shitty motherboards.
Please help metamoderate.
Ooh I love this game:
http://www.pcper.com/images/reviews/511/isaiah_arch.jpg
If you mod this up, your slashdot background will turn into a beautiful sunset!
While the performance is pretty slow (Maybe 800 mhz PIII range), it's low power and low heat, which was what I desired. Email doesn't require much processing power, so why waste the electricity on a high performance machine?
If they make a higher performance chip that get within the range of a Core 2, I'd consider buying one to replace my higher performance server in a few years. I hate paying for more electricity, and then paying to get rid of the waste heat. I'd even consider it for a workstation PC if the performance is good enough. Quiet fans are desirable to me, super-duper performance matters fairly little.
AccountKiller
I used VIA (and Cyrix) back in the days of Socket 7 and they worked reliably and well for me. I have not used VIA in any new configurations, primarily because I've been rooting for AMD and a long-time supporter of their CPUs. All that aside, I want to see VIA succeed and succeed admirably. Why? Because competition for Intel (and yes, AMD too) will only benefit the consumer in the long run. If the VIA processors force AMD and Intel to rethink their designs and then _innovate_ to keep up with (or keep ahead of) VIA then the consumers win, win, and win.
What could we get out of this? Loads, of course. One thing I'm not worried about is speed of the chips. Yes, faster CPUs are generally a good thing but I'd like to see more efficient chips coming out in all areas from the chip makers. I'd like to see less heat, less power usage under load, less standby power usage, reduced need for fans/cooling, and more along the lines of efficiency. More efficient chips, especially power usage, equates to less money I spend on utility bills or batteries or whatever. More money in my pockets, more efficient chips, more competition among the chip makers - big and small - all equals "the goodness".
My $.02 for the day...
Dream as if you'll live forever.
Live as if you'll die tomorrow.
~Anonymous~
You, and whoever last moderated the grand parent's post, aren't getting what he's saying.
What he means is: forget on-chip cache -- on-chip main memory. IOW, instead of having main memory on the motherboard, it would be embedded into your processor, running, presumable, at the same speed as the CPU.
If you follow the trends happening in CPUs, including this one, faster CPUs aren't the big issue. The real issue is the bus. The bus is slow. The more you put on the other side of it, the better. A CPU like this new VIA CPU might be slow, but if you had sufficient memory integrated right on the CPU die, it would blow the pants off your latest 4+GHz Core 2 Duo.
My blog
Wow! A 94 meter transistor, that's one big transister! How big is the laptop going to be? Shades of Oldenberg!
-mcgrew
No spam for YOU!
mcgrew's razor: Never attribute to stupidity that which can be explained by greedy self-interest
I've had nothing but good luck with them. Combined with a mini-itx fanless case, these things make great appliances. Here's a great place to get them:
http://www.logicsupply.com/
At work, we used the mini-itx with fanless case for branch office VPN solutions using linux + openswan (which in turn connected back to checkpoint clusters as well as other branch office openswan gateways). At home, I have a VIA chipset m/b with an Athlon 3000+ processer which has been running great for me for a few years.
The curious thing is that the Isaiah is heading towards OoO, whereas Intel's going to build the first in-order chip since the Pentium in Silverthorn.
C7 already has a good track-record for small form factor, low power, and providing acceptable performance at that category. IMO with the OoO they're heading more towards the laptop market, and I think they could've done something at least less conventional with the design.
Imagine that they modified the C7-M in-order execution core to a 4-way, fine grain interleaved multithreading, and have 2 cores. The existing C7-M has a short pipe, so pipeflushes aren't as penalizing. At the clockspeed that they're starting at (2GHz), each thread would have acceptable performance for your typical workload. And as OSes are becoming more thread happy (OSX is definitely one of them), such design would be at least something different than ordinary. It would be like having a cut down Sun Niagara in your laptop.
The current design would make it work decently well for low end laptop and desktops, but I can't help but think that the core now has a bunch of stuff that they can't exactly turn off - I haven't heard of a CPU that could switch off its OoO and retire queue, and the die size has increased significantly compared to the C7.
These VIA CPUs and their motherboards would do a lot more good if their nVidia drivers were completely open. Quite a lot of the overall processing power is in the nVidia chip on the mobo. But when the drivers for Linux (and probably Windows, too) don't fully expose all the video features, the CPU has to do a lot more work preprocessing, at much lower efficiency than the nVidia chip can.
--
make install -not war
Actually, the new Via architecture shares a lot more in common with the Core 2 -- Its out-of-order, spends a lot of die space on speculation, has a fairly wide execution pipe, has something similar to the Intel's uOp-fusion and much more cache than the old C7. Its also prepped to go dual-core, but the company says that'll probably only happen once they go to 45nm. This has basically nothing in common with Silverthorn, which goes back to in-order, narrow execution pipe and smaller caches in the interest of saving die-space and power envelope. Of course, Via's chip is still focused on low-power, so it doesn't scale past 2ghz (at least at 65nm, they say) but its probably comparable to Core 2 on a clock-for-clock, core-for-core basis, give or take 20% or so.
The Summary kind of has it backwards, Via's new chip competes more closely with Core 2, while Intel's Silverthorn competes more closely with Via's C7 chips.
Eden, Isaiah...
Jesus is coming soon!
Agreed. Hopefully we'll see more of technologies like Z-RAM, which sounds like it has great promise. You get the speed of SRAM, but with only one transistor per bit, and greater density than DRAM. That should lower the power consumption significantly for current cache sizes, or allow much larger caches.
but as the article said, this time it's more powerful. The C7 is not particularly strong because of its in-order execution core, and the new CPU appears to fix this.
For the record, my 2 GHz C7 machine can play a 720p h.264 video smoothly, but only without sound :) This is using MPlayer, no hardware acceleration except Xvideo.
Escher was the first MC and Giger invented the HR department.
"Troll" is not a synonym for "stupid," although deliberate stupidity can be an effective trolling tactic. The original post is not stupid or a troll, but I though I would point that out in case you decide to sling that word around in the future.
"What he means is: forget on-chip cache -- on-chip main memory. IOW, instead of having main memory on the motherboard, it would be embedded into your processor, running, presumable, at the same speed as the CPU."
Memory on the die has been done in micro controlers for years. It isn't going to happen on PCs for a long time.
"A CPU like this new VIA CPU might be slow, but if you had sufficient memory integrated right on the CPU die, it would blow the pants off your latest 4+GHz Core 2 Duo."
What is sufficient memory? 4 GB or Maybe 512 MB? There is a reason that they use Static ram for cache. It needs to be fast. So lets say that you get 512 MB on the die are you not going to allow the user to add more memory? Or how about this. You put 512 MB on the die and then let them add memory on the buss if they need more. And then you could have it swap memory from the slower buss memory in to the fast on die memory to speed everything up... Yea and we could call it a cache!
Until you can put the full address space on the die it will not work for anything but microcontrollers.
See my blog http://ilovecookes.blogspot.com/ for light hearted technical information.
Snubbing something that's perceived as a "tech tabloid" isn't really a good idea these days. If you remember, The Inquirer was first to report on the exploding lithium-ion batteries that ended up costing Sony a pretty penny.
1) This would be a new memory technology.
2) What's sufficient? Well, that depends on the application. Just as you have chips geared at different applications today, you'd have chips geared at even more different applications. In case you haven't noticed, we're moving away from the general purpose computing device anyhow and increasingly into more specialized devices to meet specialized needs.
My blog
NOT true! Yes, on-die cache is way faster than main memory, but for most applications, you're hitting the cache most of the time. I can't find any articles right now with numbers, but I know from looking at tests on the effect of cache size on performance that at some point you hit rapidly diminishing returns for cache size increases.
It's pretty analogous to main memory vs disk latency. If you don't have enough main memory, your system bogs down swapping to disk, but if you have enough memory that that doesn't happen, increasing the amount of memory beyond that won't make much of a difference.
I would say that parallelization is the next big thing(tm) in improving CPU performance. Clock speed seems to have hit a wall.
Prov 9:8 Do not rebuke mockers or they will hate you; rebuke the wise and they will love you.
You're right. But the poster has as point. The Unichrome support is really bad on Linux. There are about 3 different drivers to try, all with differing results:
:(
- The OpenChrome drivers, open source, some hw-accel support
- Unichrome drivers, open source but taking a purist approach that lacks features
- Via's own drivers, limited binaries for only certain distros, nightmare compile process, but most features supported
Unfortunately for me, I bought a VIA-epia ex1000 mini-ITX. It has some nice TV out connectors (component out!), so needs a driver that knows how to get this going. Having wasted a lot of time trying to build the drivers for FC7, I gave up and ended up using the Via binaries with FC5. The problem then is that other bits of hardware aren't detected under FC5, leaving me to patch PCI tables and rebuild the kernel to get the right southbridge driver (made a big difference to system performance - much smoother) and the SMBUS working.
Looking at forums I'm definitely not alone. This guy ended up with XP: http://cg-note.blogspot.com/2007/09/via-epia-ex1000-installation-adventure.html
Personally I think the problem is with Via. They claim to support open source, but throwing out the odd binary driver and giving mangled sources with not too easy to follow build instructions isn't much more than lip service. If they were serious, they could setup a yum repository for Fedora and make rpm's and debs for each major release of the distros they choose to support. Putting all the download packages on one page of their site would also help, as would openly releasing all their datasheets.
I hope they learn to do better, because I feel their products are held back by the poor Linux support
Mike
-- Mike
It's probably a bit of a moot point how much is enough. I'm sure OS vendors (and probably Linux and the BSDs first) could set aside as much of the on-board RAM as necessary/desired for critical parts of the OS. Perhaps we could even have a tool to say "run this executable in on-board memory, if possible." I don't claim to know a whole lot about processor tech, but I know that it wouldn't be hard to add this sort of optimization.
What he means is: forget on-chip cache -- on-chip main memory. IOW, instead of having main memory on the motherboard, it would be embedded into your processor, running, presumable, at the same speed as the CPU.
You presume wrong... a least for decent processor speeds and memory sizes. A DRAM chip accesses and cycles slower not just because it is on another chip, but because the smaller transistors in the much larger memory arrays can only drive their bit-line capacitances so fast; plus the chip has to restore the data back to the DRAM (the D is for Dynamic) transistor after reading it, and before it can be read again. In a properly designed memory hierarchy, the cache SRAM is small enough that its memory cells can drive their bit-lines fast enough that it takes significantly fewer clock cycles for the CPU to access any cache memory cell than any main memory cell (or higher level cache).
There will be some performance benefit from putting main memory on the same die, or just in the same package, if it were possible with enough yield, but that would not replace the need for (perhaps several levels of) cache for high performance CPUs.
RAM is tricky, some users need very little, others need a lot (even with the same processor requirements). Plus, there are soooo many transistors that it's cost inefficient to use the top-end manufacturing for bulk jobs like RAM.
GPU's on the other hand are fairly constant in requirements. Once it can handle HDTV, it'll be good for a lot of low and medium end use - more than 90% of users, IMHO.
It's inevitable... when's the last time you bought a Floating Point Processor? Every PC needed to do FP, so it got integrated when technically possible.
Of course gamers and media designers will still be able to get faster add-in boards.
"I don't claim to know a whole lot about processor tech, but I know that it wouldn't be hard to add this sort of optimization."
It is hard and it has been done. That is exactly what the cache does.
The tasks that need the memory the most stay in the cache longest.
See my blog http://ilovecookes.blogspot.com/ for light hearted technical information.
"1) This would be a new memory technology."
Okay well get back to me when you finish with that little detail.
"2) What's sufficient? Well, that depends on the application. Just as you have chips geared at different applications today, you'd have chips geared at even more different applications. In case you haven't noticed, we're moving away from the general purpose computing device anyhow and increasingly into more specialized devices to meet specialized needs."
Not really. PC, Laptops, and even smart phones are staying or becoming more general purpose devices. But if you are right then the statement stands. They already have it. They are called microcontrollers. Right now they are 8-bit or 16-bit devices that often have flash and ram on board. And yes they are growing bigger all the time. A single chip version of the original IBM PC is very possible today.
See my blog http://ilovecookes.blogspot.com/ for light hearted technical information.
I'm using a C3 system as a DVR and I'm also using it as an inexpensive way to learn linux. I have to say, I'm impressed with what Via is bringing to the tables, I'll upgrade to this new processor when a multi core version comes out.
We don't necessarily even have to have that much RAM on board the CPU, not initially anyway. Take a look inside your RAM right now, what's in there? Program memory, stack, but the biggest parts of it is media - that snazzy shiny icon on your desktop chews up a fair bit, so do all your nice gradient title bars and windows.
If we stick program memory and stack on board the CPU, we will already be RIDICULOUSLY fast. Recursive operations, heavy math, will all explode in performance. Put actual media on on-board memory, things that are not accessed nearly as often.
So how much memory do we need for program memory and stack? 32MB? 64MB? Might be doable. I don't think we need have things on the GB scale to see major improvements.
Not full-bore, all the shinies turned on Crysis, but it can run it.
http://enthusiast.hardocp.com/article.html?art=MTQ1MCwxLCxoZW50aHVzaWFzdA==
All opinions expressed herein are my own, and not those of my employers, who are appalled.
Okay well get back to me when you finish with that little detail. I wouldn't say that such a technology is that far off, based on what other developments I've seen in the industry.
My blog
A CPU like this new VIA CPU might be slow, but if you had sufficient memory integrated right on the CPU die, it would blow the pants off your latest 4+GHz Core 2 Duo.
I sincerely doubt it. If you actually analyze what a processor does, most of it just isn't spent chewing through gobs of memory. There's a reason why on-chip cache doesn't increase performance past a certain point.
As far as everything being on-chip, and running at processor speed, you might just have to wait a long time, and then hope no one takes advantage of the more and more memory available off-cache. The largest on-chip cache I've seen lately is 12 megs. If we assume capacities double every 18 months, that's about 10 years to get up to a gig of memory, (about standard to run most software well today). Looking back 10 years, do you think you can run modern software on a machine with the memory that came standard in 1998? I can't.
AccountKiller
i was reading about it,and its said on the web that it can decode like 20 gb/s of raw data on RSA thing.
can it be used to break in encrypted data too?
like.. imagine a gigantic 100.000 Via C7 cluster hidden somewhere in china,chewing on stolen network packets of a bank or something like that.
If you go to this link: http://www.intel.com/design/intarch/celeronm/celeronm.htm there are PDFs at the bottom of the page that detail two 1GHZ processors one at 90nm / 512K Cache and one at 65nm / 1MB Cache both have 5.5W TDP. I have the 90nm version infront of me on a Kontron ETX board, ie something shipping right now. However once you factor in the chipset and memory power requirements its more like 13W at load. 5.5W for 2010 is most likely for a cpu+chipset+memory combo.
I've run dozens of VIA chipset motherboards and peripheral devices at the job and the hardware is so unstable it makes it not worth even trying, even adding a VIA peripheral card will cause an otherwise stable Intel chip system to break down. Constant BSODs, random freezing etc.. I've never had as many problems running pure Intel systems, where only rarely does a hardware fault occur. I admit it's been about 3 or 4 years since I've written off VIA, but why is everyone so happy about VIA when, in my experience, they have made nothing even approaching a reliable computer part?
Okay well get back to me when you finish with that little detail. First off, being a smartass in response to a cogent point doesn't make me listen to you. It makes you look like you care more about appearing smart than being smart. Secondly, in response to the above mentioned point, you say that recent developments have made you think that's where things are headed- could you elaborate?
Sun took out a patent for this about five years ago, as I vaguely recall (it was about five years ago) that basically described a prototype with ... multiple cores on a single die, with a system memory like element for each chip, also on the same die. Obviously they were looking at it from the server market down to the PC market kind of angle.
... about five years ... as I recall (vaguely)
They said it wouldn't be in production for
thx e
Correction: running, presumably, at the speed of L2 cache.
"First off, being a smartass in response to a cogent point doesn't make me listen to you. It makes you look like you care more about appearing smart than being smart."
You have a point but that mystery memory break through is right up with the 100 mpg carburetor or any other mystery tech. Until it is in production it doesn't exist. The math is pretty ugly no matter how you look at it. If you developed a memory tech that is as fast as SRAM and managed to get it down to one transistor per bit well do the math.
1 GB of ram would add 8 billion tranistors to the die.
Four GB pushes that to 32 billion transistors...
Even then in a few years four GBs will be standard on PC at least those that are 64 bit.
See my blog http://ilovecookes.blogspot.com/ for light hearted technical information.
No. NUMA.
But we won't need any more than 64K of memory.
Sure, it'd take either a huge breakthrough in how we make memory work or a lot of progress on how we manufacture CPUs, but that's why I'm asking what makes him think that, rather than simply discounting it as hogwash. I'd love to see some evidence that this could be viable.
When somebody makes outlandish claims involving huge breakthroughs in what is a pretty mature technology and offers no documentation then one should instantly got into hogwash mode.
Your also a little off. It wouldn't take either a huge breakthrough in memory or a lot of progress in CPU manufacturing.
It would take both.
I doubt that it is possible to store more than a single bit with a single transistor or mythical memory cell. To do it and have it work at CPU speed would be a huge leap. To have it work at CPU speed and be be static "not loose information when powered down which he also mentioned" would be a quantum leap in memory technology. To have it be work at CPU speed, be static, and cheap enough to put on the cpu die is into the realm of Santa Claus.
To put well over eight billion of them and a CPU on a single die? That would also be a huge leap. Even if we cut it back to 512 MB you have a well over four billion and a CPU on a die.
So without documentation of some kind that even the memory technology is right around the corner this is into Yeti, UFO, and Black Helicopter territory.
Will we see a gigabyte of super fast ram on the die? I am sure we will someday but it will probably be a cache for the 512 gigabytes of DDR-X that we are using with our 64 core CPUs.
Of course when Windows IX is released you will have to get and extra 512 gigabytes to keep it running. Of course if you are using Linux you will brag about how Linux with KDE-7 works just great with only 128 gigabytes.
So yes while some may see "Get back to me when it happens" it is actually just healthy skepticism. Not impossible by any stretch of the imagination heck if you told me I would have a one gigabyte storage device that cost under $20 ten years ago I would have said yea right. However I don't see on die memory ever replacing main memory. Huge cache's that rival to days main memory spaces but replacing main memory? Only in embedded devices where systems like that are already in existence. But then we could have some really cool embedded devices by then as well.
See my blog http://ilovecookes.blogspot.com/ for light hearted technical information.
Yup, the USB and Unichrome support suck. I was happily running Ubuntu 7.04 but wanted to upgrade to 7.10. After that I've never had my video drivers working again, so downgrading seems the only option. Not the biggest problem, I wanted to go to flash based (no mechanical parts at all) OS anyway. But having no Unichrome support and having suspend not work because of the USB drivers is not fun.
I've been in it so deep that I don't even remember which drivers actually worked on 7.04.
All spelling and grammar errors are intentional. Spelling Nazi's need entertainment.
"Nazis".