Intel's New Pentium 4 Chipsets Reviewed

RainDog writes "Intel has released its 845PE and GE chipsets for the Pentium 4 processor, and reviews are hitting the web. The new chipsets officially support DDR333, but are stuck with AGP 4X and ATA/100 support. What's most interesting about these new chipsets is that they're faster than VIA and SiS' latest Pentium 4 offerings, both of which support faster AGP 8X and ATA/133 graphics and disk interfaces. As if that weren't enough, Intel's new "Blue Mountain" motherboard comes on a black PCB with all sorts of multimedia ports and memory timing options. Not bad for the traditionally conservative Intel."

Comments

Is it me....

[trims]

... or don't we see chipset manufacturers avoiding the hard problems completely? I realize that cost is an issue, but for the most part, we're talking about high-performance workstation and server boards, which cost $500+ or more.

The biggest issues these days are:

• Data Starved Processors - and this is all about latency (and, to a lesser extent, bandwidth) to main memory. I don't care if there is DDR400 memory support, what I want to kow is why isn't there a L3 cache? I mean, even the high-end Xeons these days have a max of 1MB or so on-die L2. Sure that's great, but do you know how many common datasets blow right through that? It's often dozens (if not hundreds) of cycles to access main RAM. The alpha architecture did L3 on the motherboard way back in 1994 or so. Why don't these modern server chipsets support 16MB or so of SRAM for L3 cache? Hell, they should probably support 64MB or so.
• Improved hyperthreading support - go check out the Ars Technica article on this. Hyperthreading can potentially really help performance, but it's being held back by (among other things) problems with cache coherency and loading. While much of this is on the CPU (and thus, a chipset can't help), there are a bunch of stuff that could be moved into the chipset for help.
• Useful shit in the Chipset - ATA/133 isn't that useful (vs ATA/100). Firewire is OK, as is USB 2.0, but what I want to know is where are nice stuff like block data copy between video and RAM (like the SGI chipsets for the Indy/O2 had) for high-performance video processing? AGP is a joke for this (as anyone doing video processing will tell you). These chipsets are aimed at workstations, after all.
• Standard interfaces for custom silicon - no, I'm not talking PCI-X or crap like that. There should be a standard interface directly to the chipset for people who want to do custom silicon ASICs and have them have direct access to the high-bandwidth internals of the chipset. I mean, even in the low end, why should a FCAL controller chip have to pass the PCI bus? Or a hard-core encryption coprocessor? Or a hardware routing ASIC? All need several GB of bandwidth directly to memory (or each other), and I can't see any reason not to have them surface mounted next to the north bridge with a dedicated interface.

Unfortunately, there seems to be little innovation going on in chipsets these days. The high end looks very, very, very depressingly identical to the cheap consumer crap. WTF folks?

-Erik

Re:Is it me....

[Pulzar]

Standard interfaces for custom silicon - no, I'm not talking PCI-X or crap like that. There should be a standard interface directly to the chipset for people who want to do custom silicon ASICs and have them have direct access to the high-bandwidth internals of the chipset. I mean, even in the low end, why should a FCAL controller chip have to pass the PCI bus? Or a hard-core encryption coprocessor? Or a hardware routing ASIC? All need several GB of bandwidth directly to memory (or each other), and I can't see any reason not to have them surface mounted next to the north bridge with a dedicated interface.

Expect something like this in the early 2004, when 3GIO chipsets come out to production.. most will have 4+ side-ports directly to the northbridge to used as you please. The plan is to use them for peripherals, but you'd be free to attach anything that talks 3GIO. It probably won't be quite "a few GB of bandwidth", but that really depends on the chipset designer, and not a protocol/interface limitation.

Improved hyperthreading support - go check out the Ars Technica [arstechnica.com] article on this. Hyperthreading can potentially really help performance, but it's being held back by (among other things) problems with cache coherency and loading. While much of this is on the CPU (and thus, a chipset can't help), there are a bunch of stuff that could be moved into the chipset for help.

What usefull stuff can the chipset do for hyperthreading? I'd love to hear some ideas.

The high end looks very, very, very depressingly identical to the cheap consumer crap.

"Cheap consumer crap" is what sells the most, and most companies do not have the resources to do work on more than a couple of chipsets at a time, so most of the R&D time is spent on implementing the new standards and getting things to work at the new frequencies that CPUs and RAM require. Maybe things will get better when the economy picks up and high end becomes more profitable once again.