Was Prescott really that bad?

[sirmo]

I didnt say they bough a firm with K7 design. What a strawman.
K7 is build using technology from Dec and the alpha processors with Dirk Meyers as lead engineer.

Who cares? Everything is built using technology from somewhere else? You think Zen isn't going to have some best practices from Apple's Swift core? People "borrow" ideas all the time. Intel borrows ideas from AMD.. and vice versa.

 

[Phynaz]

Who cares? Everything is built using technology from somewhere else? You think Zen isn't going to have some best practices from Apple's Swift core? People "borrow" ideas all the time. Intel borrows ideas from AMD.. and vice versa.

Intel and AMD have a patent cross-licensing agreement in place. There's no "borrowing" taking place.

If AMD doesn't have a licensing agreement with Apple they aren't borrowing anything from Swift.

 

[sirmo]

Intel and AMD have a patent cross-licensing agreement in place. There's no "borrowing" taking place.

If AMD doesn't have a licensing agreement with Apple they aren't borrowing anything from Swift.

Licensing agreement for specific instructions, but general design ideas like in any industry come in vogue or fall out of it from time to time. For instance when Hammer moved the memory controller to the CPU.. Intel followed suit on Nehalem.

You are somehow trying to frame this discussion as AMD only buys tech and doesn't design it themselves, which is absolutely wrong. K7 was an AMD design. Influenced by DEC Alpha but an AMD design through and through.

AMD has excellent engineers, Intel would be the first to admit it. https://www.reddit.com/r/IAmA/comments/15iaet/iama_cpu_architect_and_designer_at_intel_ama/c7mpa4c/

 

[krumme]

Who cares? Everything is built using technology from somewhere else? You think Zen isn't going to have some best practices from Apple's Swift core? People "borrow" ideas all the time. Intel borrows ideas from AMD.. and vice versa.

Strawman in spades again.
If thinking K7 is similar to Zen fit your beauty narrative about AMD be my guest.

 

[sirmo]

Strawman in spades again.
If thinking K7 is similar to Zen fit your beauty narrative about AMD be my guest.

Nothing strawman about my argument. Most design has external influences. It's how design works. No one comes up with a from scratch design as complex as a CPU without getting there on shoulders of giants.

This isn't just true in microprocessors.. this is true for every design. From art, to mechanical engineering, software engineering.. period.

 

[Headfoot]

Nothing strawman about my argument. Most design has external influences. It's how design works. No one comes up with a from scratch design as complex as a CPU without getting there on shoulders of giants.

This isn't just true in microprocessors.. this is true for every design. From art, to mechanical engineering, software engineering.. period.

Fundamentally redesigning the microprocessor without reusing design would be a life's work for a large number of people. The whole VISC thing is about as close as we've seen to something of that magnitude and it takes hundreds of millions of dollars to billions of dollars and the combined intellectual output of a small army of PHDs

 

[ninaholic37]

If zen is good then its the first successfull inhouse ground up design in their history...

Was Bobcat an in-house design? I thought my C-50 netbook competed pretty well against my Atoms N570 netbook at the time, for graphics it was getting 6fps on 3d video when my Atom was getting less than 1fps, and my 10.1" Bobcat had 1280x720 resolution while my Atom had 1024x600. I guess buying ATi helped them while Intel was catching up on iGPU.

 

[IntelUser2000]

Well, that target keeps going up. Back in the Nehalem days, chips would typically see drastic power increases going above ~3.5GHz. SB/IB would do 4GHz pretty easy, spike in power above that point. Now we see 4.5GHz easily, pushing toward 5GHz. Given enough time and process improvements, I think clock speeds will continue climbing, albeit not as fast as during the 90's - mid 00's.

The truth is though, no chip has really gone over 5GHz. If you look at overclocking records, you'll see that its held by CPUs with high number of pipeline stages. Prescotts, Bulldozers. Yes, they go over 5GHz by quite a bit, but records do not need full stability to be accepted. Sandy Bridge overclocks weren't that much different from Kabylake. They regressed seriously with IVB, further with HSW, gained slightly with HSW Refresh, and caught up with SKL. What Skylake/Kabylake seem to allow is to have greater masses reach those clock speeds. But for those that were on the edge anyway, no it hasn't progressed.

We shouldn't allow predetermined notion to rule out logical conclusions. At some point with all technology, there's just a practical barrier that makes little sense to go any further. Regarding clock speeds, the lower clocked and lower power chips still increase quite rapidly. The transistor level improvements for 14+ reflected quite well with KBL 15W chips. The 4.5W chip has a max speed of 3.6GHz!

 

[UncleCrusty]

I hope we will someday get some details about the canceled Tejas.

Indeed, at 40-50 pipeline stages it would have had an FO4 delay of 8~10. This would be encroaching on the FO4 delay of the double frequency ALU block of Wilamette / Northwood (~8?) and Prescott (~6?), so they would have likely had to cut the double frequency logic out from Tejas. I wonder if they would have just left it at 2 integer ops/cycle, it's not as if the P4 could make good use of 4 anyways...

 

[krumme]

Was Bobcat an in-house design? I thought my C-50 netbook competed pretty well against my Atoms N570 netbook at the time, for graphics it was getting 6fps on 3d video when my Atom was getting less than 1fps, and my 10.1" Bobcat had 1280x720 resolution while my Atom had 1024x600. I guess buying ATi helped them while Intel was catching up on iGPU.

Yep. Cpu bobcat core and gpu ip from Austin. Then a small team of aprox 70 in India Hyderabad collected all the ip into zakate e350. It was a highly synthesized design using tsmc 40nm.
And yeaa it was leaps and bonds better than Atom. I have had all the stuff and experienced the same as you.
But bobcat is a far simpler design than a big core. And notice i write big core in one of the post for that reason.
Some of the bobcat architects went to samsung. And we see the result in their new exynos in s7 and ofcource it shows.

 

[krumme]

Nothing strawman about my argument. Most design has external influences. It's how design works. No one comes up with a from scratch design as complex as a CPU without getting there on shoulders of giants.

This isn't just true in microprocessors.. this is true for every design. From art, to mechanical engineering, software engineering.. period.

"The K7 design team was led by Dirk Meyer, who had worked as a lead engineer at DEC on multiple Alpha microprocessors during his employment at DEC. When DEC was sold to Compaq Corporation in 1998, Compaq discontinued Alpha processor development. Sanders approached many of the Alpha engineering staff as Compaq/DEC wound down their semiconductor business, and was able to bring in nearly all of the Alpha design team. The K7 engineering design team was thus now consisted of both the previously acquired NexGen K6 team (already including engineers such as Vinod Dham) and the nearly complete Alpha design team."

+licensed ev6

https://en.m.wikipedia.org/wiki/Athlon

Now go look at the organization designing zen and all the fuzz about Keller - a single person 😉 It is a huge difference. It takes years to bring an organization to such a level you dont have to rely your entire design on new externals hires. Amd just didnt have the competences in the mid 90ties for good reasons.

 

[sirmo]

"The K7 design team was led by Dirk Meyer, who had worked as a lead engineer at DEC on multiple Alpha microprocessors during his employment at DEC. When DEC was sold to Compaq Corporation in 1998, Compaq discontinued Alpha processor development. Sanders approached many of the Alpha engineering staff as Compaq/DEC wound down their semiconductor business, and was able to bring in nearly all of the Alpha design team. The K7 engineering design team was thus now consisted of both the previously acquired NexGen K6 team (already including engineers such as Vinod Dham) and the nearly complete Alpha design team."

+licensed ev6

https://en.m.wikipedia.org/wiki/Athlon

Now go look at the organization designing zen and all the fuzz about Keller - a single person 😉 It is a huge difference. It takes years to bring an organization to such a level you dont have to rely your entire design on new externals hires. Amd just didnt have the competences in the mid 90ties for good reasons.

Sorry I don't see what you're driving at. All this is well known. It still doesn't mean K7 wasn't developed at AMD. Alpha was not even x86 for crying out loud.

 

[nismotigerwvu]

Yeah, I mean by that logic no CPU has been made "completely in house" by anyone in that time frame.

The Core line was hatched by Intel's Israeli team, which I'm sure there were quite a few engineers there that weren't hired directly by Intel fresh out of college without EVER having worked for anyone else. Even then, there's no way the P6 architecture they started from was completely cast by lifers either.

No business works that way. Engineers aren't like budding yeast, you don't don't just shore up the numbers by sitting and waiting for them to undergo binary fission. AMD was a rapidly growing company in the 90's as they transitioned from a 2nd source company to a full fledged CPU designer. It isn't shocking that they took a few lumps in their first design with the K5, they were working form the ground up in a new field for them. Even then the K5 had some notable strengths, it trounced in the Pentium in integer IPC. However, it only reached parity in FPU tasks at best and lagged behind significantly at worst. If they had been able to hit their intended clocks (roughly equal to the Pentiums of the era) the story would have been much different.

The only chip you could make a strong case for not being an AMD design was the K6 and there's some caveats there too. The original K6 was only on the market for about a year before it's immediate successor was released in the K6-2, so it isn't like AMD was just sitting on their hands with the design. The entire K6 line itself was only on the market for ~2 years before the Athlon came around as well, so it really was more of a transitional architecture than a technological cornerstone even if it was a major hit economically.

 

[JoeRambo]

Indeed, at 40-50 pipeline stages it would have had an FO4 delay of 8~10. This would be encroaching on the FO4 delay of the double frequency ALU block of Wilamette / Northwood (~8?) and Prescott (~6?), so they would have likely had to cut the double frequency logic out from Tejas.

Wasn't dual pumped ALU logic cut from Prescott already? I vaguelly remember it beeing the case.

 

[DrMrLordX]

Wasn't dual pumped ALU logic cut from Prescott already? I vaguelly remember it beeing the case.

http://www.chip-architect.com/news/2003_04_20_Looking_at_Intels_Prescott_part2.html

These units do fully support the double pumped ALU's but they do so by doing things in parallel and not by operating at a double frequency. Now this is of course OK, but a number of articles I wrote based on this just don't make much sense. Such as in my first Prescott article from a year ago where I said that a 4GHz Prescott equipped with a double speed Data Cache should be called an 8 GHz processor.....

Thanks Hans! From 13 years ago!

 

[PingSpike]

Prescott was bad. Let's face it, the strategy there was "People can't recognize performance, but they do recognize big Mhz numbers as a crude proxy for performance". I think in this case the old axiom about no one losing money underestimating the intelligence of the public didn't work. Or maybe it would have if they could have just scaled those things up forever.

Imagine AMD didn't exist during the prescott years. And now remember what has happened as they have lacked competitive products in the SandyBridge+ era.

 

[crashtech]

Did a lot of people run Pentium-M on the desktop? I honestly don't remember seeing a lot of them, but at that point I think I was pretty heavy into AMD.

 

[Denithor]

Those who knew were into AMD.

Those who didn't, bought into the MHz wars, and/or bought their PC from Dell, etc. And those guys were taking BIG payoffs from Intel to not use AMD chips, pushing "normal" consumers onto inferior Intel processors.

Imagine how the landscape might look today if Intel had played fair and AMD had been able to properly capitalize on their superior position.

 

[nismotigerwvu]

Did a lot of people run Pentium-M on the desktop? I honestly don't remember seeing a lot of them, but at that point I think I was pretty heavy into AMD.

They existed, but were limited to the builders market. Even then, for gamers AMD still made way more sense. http://www.anandtech.com/show/1610/15

 

[UncleCrusty]

Wasn't dual pumped ALU logic cut from Prescott already? I vaguelly remember it beeing the case.

I was under the impression it remained, but the 1/2 cycle bypass was removed to make room for the 64 bit addition on the critical path.

http://www.chip-architect.com/news/2003_04_20_Looking_at_Intels_Prescott_part2.html



Thanks Hans! From 13 years ago!

Interesting, although I wonder where he got his information from. The lack of the fast bypass would make it indistinguishable to an outside observer. Scheduling would be simpler in the fast frequency domain though, which is why I thought they would just leave it at 2 ALUs upon its removal.

 

[Dribble]

Prescott was the first Intel dual core desktop chip (for home users at least). That was its one big step forward over Northwood and probably it's only positive claim to fame.

 

[daxzy]

They existed, but were limited to the builders market. Even then, for gamers AMD still made way more sense. http://www.anandtech.com/show/1610/15

Yes at stock they were just ok. But lets be honest, if you buy a CT-479 adapter at the time, you'd be getting the easy 2.4-2.5 GHz overclocks. That's when they shine. The main problem with the Anandtech review is that they use the mobile Intel 855GM chipset. Folks with CT-479 would use the higher performing (and cheaper and more readily available) desktop 865/875/915 chipsets.

From the Anandtech you linked, the Pentium-M 735/745 at $245/299 retail was the sweet spot. You could easily FSB overclock to 133, to hit 2.26/2.4 GHz. Raising to 150 FSB would get an attainable 2.55 GHz for the PM-735, but 2.7 GHz on the PM-745 was quite rare.

http://www.anandtech.com/show/1650/12

With the CT-479, ASUS has effectively demolished all other desktop Pentium M solutions. There's no reason to even consider an 855GME motherboard from AOpen or DFI; the ASUS solution is cheaper, better performing and is even a much more stable overclocker.

As a gaming platform and as a general purpose/office machine, the Pentium M does fairly well, but it is in content creation, workstation and media encoding applications that the Pentium M continues to fall behind. Part of the problem is that the Pentium M needs clock speed to compete, which we saw when we overclocked it up to 2.56GHz. But even at 2.56GHz, the Pentium M wasn't a competitive CPU when it came to tasks like media encoding, indicating that if the Pentium M is to succeed on the desktop, it's going to need some architectural improvements.

https://www.pcper.com/reviews/Proce...pgrade-Kit-Review/Gaming-Quake-III-and-UT2003

At that level we saw the P-M out performing the Athlon 64 FX-55 processor in gaming and the Pentium 4 in some media tests. The power of this little mobile processor continues to impress me, and I am eagerly awaiting for Intel to adopt this basic architecture and expand on it for their entire line.

http://www.tomshardware.com/reviews/dothan-netburst,1041-12.html

Ironically, Tomshardware didn't do it optimally. You were supposed to take a 100/400 FSB Dothan to overclock, not the 133/533 FSB SKU, as you'd hit an FSB wall unless you are doing some exotic cooling.

Additionally, we were able to raise the FSB from 133 to 160 MHz without any trouble at all. The result was that our 2.13 GHz Pentium M 770 ended up running at 2.56 GHz! At this clock speed, our two year old platform was able to beat the processor heavyweights Athlon 64 FX and Intel Pentium 4 Extreme Edition in all 3D games!

 

[DrMrLordX]

Prescott was the first Intel dual core desktop chip (for home users at least). That was its one big step forward over Northwood and probably it's only positive claim to fame.

Eh not really. You're thinking of Smithfield. Not quite the same thing as Prescott which launched a year earlier.

 

[daxzy]

Eh not really. You're thinking of Smithfield. Not quite the same thing as Prescott which launched a year earlier.

Smithfield was just two Prescott cores glued together (quite literally, actually). Thus you can image the massive increase in power consumption and the clock speed wall.

http://www.anandtech.com/show/1986

All Smithfield processors were made of two 90 nm Prescott cores, next to each other on a single die with 1 MB of Level 2 (L2) cache per core. Hyper-threading was disabled in all Pentium D 8xx-series Smithfields with the exception of the Pentium Extreme Edition 840.

The one saving grace of Smithfield, IMO, was the low price. The Pentium-D 805 was quite popular because it was dirt cheap at $133 (the cheapest X2-3800+ was $350). On air, you could FSB O/C to 166 (3.3 GHz) and have a pretty cheap DC solution.

 

[UncleCrusty]

I was under the impression it remained, but the 1/2 cycle bypass was removed to make room for the 64 bit addition on the critical path.

Was still curious about this, and found some Intel papers confirming the presence of double frequency ALUs in Prescott and Cedar Mill. The removal of support for dependent operations in consecutive 1/2 cycles rather defeats the point though. If they it down from 4 to 2 or even 1 ALU ops/cycle, ALU throughput would likely remain quite far down on the hierarchy of the P4's bottlenecks in most applications. Given how marketing-driven the P4's design was, I'm astonished they designed such a ridiculously overkill, cool-sounding, and ultimately pointless thing as the "Rapid Execution Engine" and didn't market it in any capacity.