Introduction
In this third part about Intel’s indispensable future products you’ll find information about the future desktop chipsets as well as about the pretty impressive stuff that’s going to happen in the notebook arena.
Desktop Chipsets
We are running our high performance systems on BX-platforms for about a year now and it will still take a few months until there is going to be a change. In the low cost sector there has recently been the launch of the i810-chipset, but nobody in the business seems too amazed about it, as I could find out when I spoke to motherboard manufacturers at Computex. The future doesn’t look too ‘intoxicating’ as well really.
i820 alias Camino
Haven’t we heard really tons about ‘Camino’ already? The intoxicating miracle chipset, the first to support RDRAM or ‘direct Rambus RAM’, the first to support the great AGP4x, and almost the first to support 133 MHz front side bus!! Well, we’ve also heard about the downsides, particularly about the missing PC133 SDRAM support. At first there was no sensible SDRAM support altogether and then there were problems with the RDRAM at higher Rambus speeds. It seems a matter of fact that hardly any Intel chipset produced as much press as i820 ever before. Now what’s the real deal?
Camino and RDRAM
Well, Camino will use RDRAM RIMMs in the first place. This memory type has got a very highly clocked serial interface, that can only transfer 16 bit of data at once, but it does that at 300 to 400 MHz with transfers at the rising as well as the falling edge of the clock. The high frequency of the Rambus interface enables very fast burst-accesses, but the serialization of the commands increases latency. This has the following effect. Software that needs to read large amounts of continuos data out of memory will benefit from Rambus, applications that only require small portions from different areas won’t benefit from Rambus, the L2-cache might actually cover the fact that the (memory) performance of those applications will even decrease slightly. Unfortunately there’s only a minority of applications that do indeed benefit from fast burst-accesses, the majority (3D games, business apps) won’t. If you remember my Computex-article, you’ll also remember what I was told by the few motherboard makers who had their Camino-platforms up and running at 400 MHz RDRAM-interface speed. “It’s running fine, but the performance is identical to BX” was what I heard from everybody that I asked.
What will be the benefit of RDRAM?
Well, the CPU itself won’t be able to benefit from Camino and the RDRAM much. If the Camino-platform should run at only 100 MHz front side bus, as e.g. with PIII 600/100 or PIII 650/100, the CPU won’t be able to transfer more than 800 MB/s, regardless how fast the memory may be, because of the 64-bit (=8 byte) wide memory bus running at 100 MHz FSB. Please remember that 800 MB/s could already be supplied by PC100 SDRAM. Running the FSB at 133 MHz, as e.g. with PIII 533/133, 600/133 or 667/133, gives a CPU peak data bus bandwidth of still 1.066 GB/s only. RDRAM at 400 MHz can supply a peak bandwidth of 1.6 GB/s, PC600 RDRAM can already supply up to 1.2 GB/s. Thus CPU depending applications won’t benefit much from RDRAM at all and they’ll only benefit if you run the FSB at 133 MHz.
RDRAM plus AGP4x does make sense
The story looks a bit different when you consider that Intel was smart enough to equip Camino with AGP 4x as well. They pretty much had to, so the RDRAM was halfway worth our while. Applications that make heavy usage of the AGP, which are mainly 3D-games, could often enough run into arbitration problems when accessing main memory over the AGP whilst the CPU was trying to access main memory as well. AGP2x could transfer data at peaks of 533 MB/s, whilst the CPU at 100 MHz FSB could transfer data at up to 800 MB/s. PC100 only offered 800 MB/s bandwidth though, and so arbitration problems were more than likely. AGP 4x is now able to transfer data at up to 1.066 GB/s, just as a CPU at 133 MHz FSB, so that the 1.6 GB/s supplied by PC800 RDRAM are indeed useful in case an application takes advantage of AGP4x. We’ll have to see how much of a difference AGP4x is going to make with the games though. After all we shouldn’t forget that one of the currently fastest graphics chips, the Voodoo3 from 3Dfx, is not even taking real advantage of AGP2x right now (no AGP-texturing support) and it’s still performing very respectable. Compared to the bandwidth of memory mounted on the graphics-card, which ranges from 2.3 GB/s (3Dfx Voodoo3 2000) to over 4.5 GB/s (Matrox G400), the AGP4x bandwidth is still pretty pathetic.
What the roadmap says about RDRAM
There seems to be a little problem in the production process of RDRAM. Currently, it’s very difficult to manufacture RDRAM that runs at 400 MHz. This was the one type of RDRAM that Intel was talking about in the past. Now it’s only going to be available in small quantity at most likely very high prices. The result is that the parts that wouldn’t pass the tests at 400 MHz are either ‘downgraded’ to run at 356 MHz or if even this test fails the ‘scrap’ can still be sold as PC600 RDRAM, running at 300 MHz. The PC700 = 356 MHz version is nothing else but the confession of RAM-makers that they can’t supply the 400 MHz product. It was quickly invented to have something that can at least be sold more expensive than the PC600 RDRAM. That’s why ‘Carmel’ or i840 can’t run it, the PC700 standard came through too late.
The above picture simply states my comments, PC700 was only just ‘created’ and it’s of course important to explain that PC700 is even intoxicatingly faster than it sounds!!! 22 MB/s more than 1.4 GB/s makes a difference of even 1.6 % !!! Thank you Intel for pointing out this highly important issue, I’m sure I would have noticed that when running my system. I would have said “well, I guess my system is running even faster than PC700, it feels like some 1-2% really!!”
Camino or i820 will support PC100 SDRAM as well now, as already mentioned in my Computex-article. This solution was not planned in the first instance, the only plan to make SDRAM run with Camino was supposed to be a RIMM-riser card in which you could plug your old SDRAM. This riser card won’t only be pretty expensive, it will also be a complete performance killer, because it translates the serial Rambus interface back into the parallel SDRAM interface, increasing latency even more. A comparison of RDRAM vs. SDRAM on a riser card would have a clear outcome, SDRAM would lose out big time with this rather cheesy solution. Now motherboard makers can include an ‘MTH’-chip (Memory Translation Hub ??) onto the board, which talks directly to maximal two DIMM slots. We’ll see what kind of performance will be achieved with this complicated solution, I doubt that it will be any better than our good old BX, it might be even slower.
i840 or ‘Carmel’
Carmel will be the new performance chipset for workstation and servers, running with Slot1 or Slot2, Pentium III and Pentium III Xeon. It will also offer AGP4x, RDRAM-support and 133 MHz FSB, but it will add the following:
- Quad processor support
- Dual-RDRAM interface, offering a peak bandwidth of 3.2 GB/s
- 64-bit PCI (‘normal’ PCI is only 32-bit wide)
- up to 8 GB of memory (Camino only offers 1 GB)
- MRH-S (Memory-Repeater-Hub for SDRAM) offers SDRAM-support at high bandwidth through up to 4-way interleaving
At the same time there won’t be any support of PC700 RDRAM.
i810e
Wow, this one is my favorite out of them all! Are you also wondering what the little ‘e’ means behind the now well known ‘i810’? Well, it doesn’t mean much really, but it makes i810e a new product. The difference over i810 is merely the ‘support of 133 MHz FSB’ and a 133 MHz display-cache bus. Wow! This means you can plug your expensive new Pentium III 533/133 or Pentium III 600/133 into a platform with this low cost integrated chipset. That’s what we all ever dreamt of, hampering the intoxicating performance of a new Intel-CPU with a chipset that offers mediocre graphics performance. This combination will indeed show how ‘indispensable’ SSE is. A 3D-application gets first accelerated by SSE-instructions to be slowed down right afterwards by i810e’s slow 3D-graphics. I really wonder how successful Intel’s marketing will be with this combination, because one thing is for sure, nobody needs it! Even in an office environment where you wouldn’t require 3D, you don’t require 133 MHz FSB and SSE either and can happily work with i810 and Celeron 466 or 500, saving a very decent amount of money.
Solano
This chipset is almost the most boring out of the bunch of new chipset. It’s nothing than a ‘tuned’ i810e-chipset to me. The difference to the lovely i810e is the introduction of ‘Command per Clock’ = CPC ‘for great graphics performance’ (whatever Intel means by that) and the support of AGP4x for an external (!!!) AGP-3D-card. Now what shall we say about that one? Intel realized that their integrated graphics solution isn’t quite good enough! Therefore Solano will still come with integrated i75x-graphics, but it will enable you to add a different 3D-card if you like. This 3D-card can make usage of AGP4x, but unfortunately there isn’t any RDRAM-support. PC100 with CPC has to do it for Solano, I wonder if this will be good enough. As a matter of fact I doubt it, but people who were talked into building or buying a PIII-platform with i810 will certainly love Solano just as much. Long live Intel’s marketing deparment!
Mobile CPU Roadmap
The most different CPUs are currently found on Intel’s Mobile processor roadmap. Don’t get confused though, as a matter of fact it’s the roadmap I honestly like best.
All the nice and actually well performing notebook CPUs we are using right now will be gone pretty soon. Celeron will be available in faster versions and the only young Pentium II / ‘Dixon’ will very soon be replaced by Coppermine for notebooks.
‘Geyserville’
The top four CPUs in this list are marked green for a reason. They include Intel’s upcoming ‘Geyserville’-technology for notebooks. Geyserville is nothing really too spectacular, but it’s a pretty good idea. It makes sure that the CPU gets clocked down, the core voltage drops and thus the power consumption becomes less once the notebook runs from its battery, prolonging its operating time. This means that a ‘Geyserville’-enhanced CPU might run at the same speed as a slower CPU without ‘Geyserville’-tech when operating the notebook from the battery, but as soon as you plug it into the next power-plug, the ‘Geyserville’-CPU can supply you with its full performance.
Mobile Pentium III
Coppermine includes the cache on-die and runs at low-voltage due to its .18µ-core, which is why it’s a perfect CPU for notebooks. All versions above 500 MHz will include ‘Geyserville’-technology and the first three will be launched with Coppermine for desktop and Camino on September 5, 1999. There’s not really that much to say about it, except that this will start the era of ‘indispensable’ SSE-enhancements on notebook-platforms. The only thing I cannot understand is that mobile PIII 550 will be launched after mobile PIII 600, maybe someone from Intel could give me a call and explain that to me.
Mobile Celeron with SSE
Isn’t it cool that the mobile Celeron will be the first to receive the SSE-blessings? The launch of mobile Celeron with SSE 450/100 and 500/100 will be in Q1 next year, whilst the desktop Celeron will have to wait until Q2/2000 to be equipped with the ‘Coppermine-128’-core. Anyway, it shows how important the notebook-segment has become. I’m sure we can expect great performance on the notebooks of the close future. I’m really looking forward to it.
Tomorrow this series will finally be finished with the Server CPU Roadmap, Server Chipset Roadmap and the pricing chart. Then you can read about Merced, Foster, Willamette and their platforms …