Stealth 2001 - My i850 Renovation Project & Sveta SMP 2003 - My i860 Fun Project!

OK slow down a little

You're mixing some things up there.

There's no doubt that Thunderbird and of course Palomino und T-Bred had a much better IPC than all of the P4 Cores.

And it is in deed true that the first P4 chips were totally crushed by the Thunderbirds of the same area although they ran several hundred megahertz slower.

But - it is defenitely wrong, that AMD had the 3200+ out, when Intel launched the 3,06 HT. If I remember correctly AMD had still only T-Bred B on the market (with the 2700+ the fastest available and the 2800+ as paperlaunch CPU); the Top Dog 3200+ Barton launched about half a year or so later; the Athlon 64 launched about a whole year later than the Northwood 3,06 (I don't want to argue about 1 or 2 months or so..).

So - the 3,06 was clearly the fastest CPU on the market for several months.

And still - a 1,67GHz Thunderbird definitely doesn't stand a chance against the 3,06 Northwood. I think 400 - 500 MHz more and then you'll be at the level

No question AMD was better value, no question the P4 was crap overall but we gotta keep the facts right.

Even when overclocking the Thunderbird would never reach the 3.06 GHz Northwood or the Prescott. Never. Sorry for that. An Athlon XP sure, but the old Thunderbird, which is missing a few things? No.

Prescott didn't have L3 Cache, only Gallatin out of the Intel Desktop Line of CPUs.

The Prescott was further desgined/"optimized" for more clockspeed (Intel aimed for 10GHz which we all know failed miserably ) and was therefor even worse IPC-wise than Northwood. Additionally the first Prescotts ran extremely hot, they were just pure crap when they launched, just as the Willamette Core (well not completely comparable, but still a step back at first). The revised 6xx P4s at least fixed the power consumption, ran really cool and overclock extremely well.

The G1 stepping of the Prescott is really nice. Its the last and best stepping you can get, when using socket 478. It overclocks really nice (4.0 GHz is realistic) and runs cool for a P4.

But even with 4 GHz, the P4 will loose against an Athlon64. Neburst was a dead end.

So Intel modified the P3 and created Banias and Dothan and on this base, they created the glory Core-architecture. I still use a Xeon X5470...

After more then 2 years my Blue Thunder got an update:

I droped the memory size from 4x 512 MB PC800-40 to 4x 256 MB PC800-45. But the new RDRAM heatspreader in blue in compose with the CPU Cooler looks mutch better. And there is no different in the performance between 40 ns and 45 ns.

Kingston 4x 512 MB PC800-40	Samsung 4x 256 MB PC800-45

Just curious. Are these 5 ns in difference ever showing any impact like in UT or anywhere else?

Never found any benchmarks refering to different rambus modules.

No. The access time actually depends on the clock speed of the RDRAM. This is not similar to things like CAS latency specs for SDRAM. So if you raise the clock speed, the access time sinks in turn, as it is calculated (partly) from the actual clock. The access time specified for a given RDRAM merely specifies its maximum clock rate. So 45ns parts are for 800MHz data rate. The (small!) differences observed here are likely nothing but the usual measurement uncertainty.

Hence, PC1066 32ns RDRAM isn't "clocked faster *and* lower latency". It's "lower latency *because* it's clocked faster".

This is also why I've complained a few times about CAS/RAS2CAS etc. ratings for SDRAM as they're specified in pretty useless clock cycles instead of real nanosecond values. So, CL2 at PC133 is actually faster than CL2 at PC100, because those "2 clock cycles" pass quicker at 133MHz clock rate. Comparing just CAS latencies without comparing clock speeds in relation to them is essentially useless.

A bit of background story: Real Rambus access time and bandwidth is actually dependent on quite a few more factors, that were actually limited on cheap x86 (Pentium III, Pentium 4) platforms. RDRAM needs lots of power and runs hot, so on the x86 PC, most memory pages were kept closed (no voltage supply) or half-open (partly powered). For actually using a page, you need it in full-open mode. E.g. on i840, 32 pages could be kept full-open at the same time. This was to not overheat the memory subsystem or require active cooling for it. Pages not in use for some clock cycles would be switched to half-open, then later to closed.

Naturally, accessing a full-open page is super fast, but accessing a closed page needs more time, because it needs to be switched on first. So access depends on the amount of full-open and closed pages you'll access over the course of your program working with the memory.

Digital AlphaServers and AlphaStations with the ALPHA EV7 RISC microprocessor on the other hand did not have such a limitation. Aside from running the Rambus memory in quad-channel using four IMC's in the CPU (plus a 5th channel for redundancy), those systems ran in full-open page mode for a whopping 2048 pages (64 times that on Intels!) across the RDRAM at any given time. Hence, they were faster in terms of memory access and effective bandwidth. In turn, the memory ran much hotter and used more power. But those machines were usually well-ventilated.

Just a bit of history. I hope I remember every detail correctly, the excellent RDRAM architecture articles I read maaany years ago have been lost by now.

More interesting information (also about RDRAM) can be gathered from [this document] from the University of Magdeburg. This also has measurements for effective (!) full-open (80ns) vs. closed page (130ns) access times.

Makes for a fun read even if all you're interested in is RDRAM tech (and not ALPHA processors and servers using them).

Thanks for your explanation :).

One interesting point coming up while reading 'Alpha EV7'. Didn't AMD took the protocol for their Athlon processors?

Yes, that one was from the even older ALPHA EV6 (21164) though. AMD sublicensed it for their original Athlon platform and the ones following it. It was later replaced with HyperTransport, like QPI replaced the aged FSB. The insane Rambus stuff happened only later with the EV7 chips. Some time ago you could even find double-height 1 GiB RDRAM modules on eBay, which won't work on any x86 platform. Those are all from/for EV7 systems, which support the larger modules.

Good to know. I was wondering some time ago, if I should give one of those modules a try :).

The next Battle-Mashine for GATs x264 Benchmark.

Colusa is running. At First with two Foster CPUs.

The FX 5200 Testing Card did their best.

Not so impressiv, in comparison to two Tualatin.

But there will be more