Article Another "greatest CPU of all times" article.

[mikegg]

Amen. 4004 is the beggining of everything. M1 is more on the Conroe tier.

M1 is well above Conroe tier. It's a paradigm shift in both speed, efficiency, and overall chip design for PCs.

Conroe was definitively faster than what AMD had at the time but not multiple times better like the M1 was in inefficiency.

Even today, M5 can be 3-4x more efficient than AMD chips.

 

[SPBHM]

I remember also the Apple A7 being very very impressive, specially because it was very early in the designed by Apple CPU days,

 

[511]

M1 is well above Conroe tier. It's a paradigm shift in both speed, efficiency, and overall chip design for PCs.

Conroe was definitively faster than what AMD had at the time but not multiple times better like the M1 was in inefficiency.

You are comparing a dated 14nm CPU to a 5nm process that's like 3-4 gens of TSMC Node (12nm,10nm,7nm and than 5nm ). It was the biggest improvement in SoC for the last few years yes I agree but not Conroe tier.

Nothing beats 4004 in this list we have had more revolutionary designs than M1 throughout the history what about the design that did OoO or that did Superscaler those are far more impressive.

Even today, M5 can be 3-4x more efficient than AMD chips.

I Doubt this big time it's not true maybe in st efficiency but that's it but not 3-4X

 

[Thibsie]

Yeah, I had one - from 300 MHz to 450 MHz in a blink of an eye 😉

And dual CPU capable thanks to Abit.

 

[Hulk]

And dual CPU capable thanks to Abit.

"Abit." Now that is a name I have not heard in a long time. A long time.

 

[gdansk]

The 1800X stands out as a particularly bad CPU on this list. I suppose the idea was Ryzen bringing cheap MT, but that was the 1700/1600 not the 1800X.

 

[Thibsie]

"Abit." Now that is a name I have not heard in a long time. A long time.

I still remember my KT7-raid, KT7A-raid and KG7. Lovely times 🙂

 

[SteveGrabowski]

Anand had an article about the G3258 and basically said it was better to spend a few bucks more on a locked i3 because it had HT. Surely it aged better for that reason too. And if I recall correctly with the G3258 you still need a Z series chipset to OC making it a worse value than you would otherwise think.

Not even aged better. The i3 curbstomped it in gaming from the beginning. You didn't need a Z-series board to OC on it though. I bought a G3258 back when I didn't GAF about PC gaming and could run it stably at 4.6GHz on the stock cooler on a dinky H81 board (Gigabyte GA-H81M-DS2V if I remember right).

But back to i3 vs G3258, when the GTX 970 launched I got interested in PC gaming again and ordered one with a Xeon E3-1231v3 (basically a locked i7-4790 minus 200 MHz and marketed as a low end server processor instead of highish end desktop cpu and $70 cheaper). The 970 came right away, the Xeon took 2.5 weeks to come, so I got the chance to play G3258 + GTX 970 at 1080p and it was an insane bottleneck. Tomb Raider 2013 would have this periodic oscillation from 20 fps to 100 fps that made it completely unplayable and that was a cpu light game of the time. Shadow of Mordor was similarly terrible on that G3258 whether at 4.6GHz, 4.2 GHz, or stock. Far Cry 4 (which came with the gpu) wouldn't even start. So played Bioshock Infinite and some old COD games that worked great on that cpu until my Xeon came in. Then for fun I disabled two cores of my Xeon in the BIOS to see how much better a simulated i3 was than the G3258 and it was night and day, most games showed no noticeable difference running the Xeon 4C/8T vs that Xeon at 2C/4T with my GTX 970.

 

[SteveGrabowski]

i5-2500k at ninth place seems kinda low. I remember people were running those forever, and didn't seem like many games other than Crysis 3 started taking advantage of the extra threads the i7 had until 2017 or so. Was there even a no-brainer upgrade from the 2500k until say the Ryzen 5 3600?

 

[Thunder 57]

I still remember my KT7-raid, KT7A-raid and KG7. Lovely times 🙂

I was very fond of Epox back in the Socket A days. I miss the variety of motherboards. DFI was another popular one. Someone I knew was loyal to Soyo.

Not even aged better. The i3 curbstomped it in gaming from the beginning. You didn't need a Z-series board to OC on it though. I bought a G3258 back when I didn't GAF about PC gaming and could run it stably at 4.6GHz on the stock cooler on a dinky H81 board (Gigabyte GA-H81M-DS2V if I remember right).

But back to i3 vs G3258, when the GTX 970 launched I got interested in PC gaming again and ordered one with a Xeon E3-1231v3 (basically a locked i7-4790 minus 200 MHz and marketed as a low end server processor instead of highish end desktop cpu and $70 cheaper). The 970 came right away, the Xeon took 2.5 weeks to come, so I got the chance to play G3258 + GTX 970 at 1080p and it was an insane bottleneck. Tomb Raider 2013 would have this periodic oscillation from 20 fps to 100 fps that made it completely unplayable and that was a cpu light game of the time. Shadow of Mordor was similarly terrible on that G3258 whether at 4.6GHz, 4.2 GHz, or stock. Far Cry 4 (which came with the gpu) wouldn't even start. So played Bioshock Infinite and some old COD games that worked great on that cpu until my Xeon came in. Then for fun I disabled two cores of my Xeon in the BIOS to see how much better a simulated i3 was than the G3258 and it was night and day, most games showed no noticeable difference running the Xeon 4C/8T vs that Xeon at 2C/4T with my GTX 970.

It was a joke because of classic Intel segmentation. I watched the video @DAPUNISHER included and forgot that that Pentium was stripped of AVX and AVX2! That might be why some games didn't start. It was Intel trying to look like they were throwing a bone to users during the crappy AMD days but it was just garbage for an unlocked CPU.

 

[Thunder 57]

i5-2500k at ninth place seems kinda low. I remember people were running those forever, and didn't seem like many games other than Crysis 3 started taking advantage of the extra threads the i7 had until 2017 or so. Was there even a no-brainer upgrade from the 2500k until say the Ryzen 5 3600?

Depended on use case and AVX2 I suppose. Personally I went from a 3570k to a 2600X because threads became more important. So many people jumped on the 2500k train that only stragglers like me got the Ivy Bridge equivalent instead. It ran a bit cooler and couldn't OC as high but was good enough for a long while.

 

[SteveGrabowski]

It's overclocking potential was the equivalent of polishing a turd. Iceberg did a revisit of it a few years ago, and it's comedy gold.

Can't believe the article in the first post saying it was better than an i3 in gaming. No way in hell, the writer clearly never gamed on one. I had one in 2014 and it was great for what I originally bought it for, just an internet browsing machine with the occasional round of Minecraft or xmame. But when I paired it with a GTX 970 (don't laugh, newegg took forever delivering the Xeon E3-1231v3 I ordered with it) it was a painful experience for a lot of 2014 games. It ran amazing in old games like COD Modern Warfare 1 & 2 and in Bioshock Infinite, but the other games I tried on it were awful. Shadow of Mordor and Tomb Raider 2013 had horrible framerate oscillations, I mean think how crappy a periodic oscillation between 20 fps and 120 fps would feel. Turn on VSync and it was just an oscillation between 20 fps and 60 fps, also unplayable. Far Cry 4 wouldn't start on it because it hardcoded startup on Core 3 if I remember right. Crysis 3 was a mess even with half refresh VSync to 30. I'd drop to 25fps all the time and it felt gross. Meanwhile when my Xeon finally came in I couldn't find a game that wasn't fun to play and a great experience with my GTX 970 after disabling two cores in the BIOS to simulate a 3.8GHz i3. Even Crysis 3.

 

[Tuna-Fish]

Both the Athlon and Athlon 64 were huge steps forward. AMD’s missteps afterward were baffling to me. Had they focused on strong IPC increases rather than new architectures, I do wonder if they would’ve been better off.

The execution engine of Athlon64 was just really dated, and they had sort of painted themselves into a corner where it wasn't easy to improve in an incremental way. It was a neat, efficient design for a shallow OoO CPU when it was designed for K7, but trying to increase the size of any of its queues would have ballooned power use. They knew it was time to redesign it, but they just failed repeatedly.

 

[Abwx]

And dual CPU capable thanks to Abit.

IIRC that was the BP6 MB, Abit after publicly stating, under Intel pressure, that they wouldnt release MB for AMD Athlon/Duron abrutly changed their behviour and released the KT7 133 Raid for socketed versions of these CPU, i bought one for a Duron 700, legendary ultra cheap high perfs CPU that was better than my Pentium 3 1GHz, the latter had theoricaly better perfs but the Duron was so much more reactive.

 

[Thunder 57]

IIRC that was the BP6 MB, Abit after publicly stating, under Intel pressure, that they wouldnt release MB for AMD Athlon/Duron abrutly changed their behviour and released the KT7 133 Raid for socketed versions of these CPU, i bought one for a Duron 700, legendary ultra cheap high perfs CPU that was better than my Pentium 3 1GHz, the latter had theoricaly better perfs but the Duron was so much more reactive.

Ah yes Duron, an excellent CPU that is the only one I know of that has a smaller L2 cache than L1. It could get away with that because the L2 on K7 was, well, crap. Excellent L1 though.

 

[Abwx]

Ah yes Duron, an excellent CPU that is the only one I know of that has a smaller L2 cache than L1. It could get away with that because the L2 on K7 was, well, crap. Excellent L1 though.

Cache was exclusive, so that was 64k + 128k = 192k effective cache, Celeron had 128k + 32k inclusive so only 128k actual cache, in perfs the Duron was very close to a similarly clocked Athlon and noticeably above the Celeron, not counting its higher FSB, ie, 100MHz vs 66MHz and 133MHz RAM support vs 100MHz.

 

[Thunder 57]

Cache was exclusive, so that was 64k + 128k = 192k effective cache, Celeron had 128k + 32k inclusive so only 128k actual cache, in perfs the Duron was very close to a similarly clocked Athlon and noticeably above the Celeron, not counting its higher FSB, ie, 100MHz vs 66MHz and 133MHz RAM support vs 100MHz.

That doesn't mean that K7's L2 was shit. It was dog slow connected by a 64 bit bus compared to the P3's 256 IIRC. Why else do you think Barton didn't do much for K7?

 

[Abwx]

That doesn't mean that K7's L2 was shit. It was dog slow connected by a 64 bit bus compared to the P3's 256 IIRC. Why else do you think Barton didn't do much for K7?

You are likely talking of the Slot A version of the Athlon that had SRAM for L2 cache,
the Duron only existed in the socket A version that was released later along with the socketed Athlon, on those CPUs they transited to an on die L2, that tbeing said it was 20% faster in games than the Pentium 3, so seems that it wasnt a limitation.

Slot A had 0.5x to 0.33x CPU frequency L2 SRAM on 64b bus but with double data rate (EV6 bus) while the socketed Athlon Thundebird and Duron had full speed L2 .

Btw the first Pentium 3 also had SRAM for L2.

 

[Jan Olšan]

Btw the first Pentium 3 also had SRAM for L2.

TFW you remember when SSE was Katmai New Instructions

 

[Abwx]

TFW you remember when SSE was Katmai New Instructions

SSE was nothing more than Intel s response to 3DNow wich was implemented in previous AMD CPUs like the K5 and K6, AMD later also implemented SSE because Intel was dominant in the X86 market, beside it s not known but K6 III was better than the Pentium 3 in Integer, it s just that the latter was quite better in FP and at the time games were massively using FP, this was reversed by the Athlon wich was better in both INT and FP thanks to its 3 ALUs design and triple issue FP pipeline while the P3 was 2 issues FP and still, only conditionaly.

 

[Jan Olšan]

SSE was nothing more than Intel s response to 3DNow wich was implemented in previous AMD CPUs like the K5 and K6, AMD later also implemented SSE because Intel was dominant in the X86 market, beside it s not known but K6 III was better than the Pentium 3 in Integer, it s just that the latter was quite better in FP and at the time games were massively using FP, this was reversed by the Athlon wich was better in both INT and FP thanks to its 3 ALUs design and triple issue FP pipeline while the P3 was 2 issues FP and still, only conditionaly.

I don't think I can agree with that, even as an AMD fan. SSE was (is) clean architectural move to 128bit SIMD width using new registers (obviously later extended with more ops, arguably the first SSE set was too limited, lacking integer SIMD for multimedia, which was sorely needed back then for video, and even SSE2 was usable but limited until SSSE3). It was obviously architected and planned in advance.
If anything, 3DNow! was likely reactionary - AMD saw opportunity to make a floating-point complement to the (integer-only) MMX set that they could do quickly before Intel makes one.

3DNow! may have been the first x86 floating-point SIMD but it was a hack - same hack as MMX, of course. Those extensions were better than nothing and were quickly available but their use of x87 registers was rather messy (and complicated the x86 ISA and cores) and limited the performance because just 64bit width. Less of an issue for MMX (you could process 8 INT8 values in one op), but quite a bit worse for floats (FP32 datatype so just two values processed per op).
What you say is simply incorrect.

Had Intel not made SSEx, someone else would have to. But certainly not like MMX/3DNow.

BTW, 3DNow! was not present in K5 and not even in the K6. It was one of the selling points of K6-2 and later chips. I wish K5 had MMX, in the first place...

 

[Abwx]

I don't think I can agree with that, even as an AMD fan. SSE was (is) clean architectural move to 128bit SIMD width using new registers (obviously later extended with more ops, arguably the first SSE set was too limited, lacking integer SIMD for multimedia, which was sorely needed back then for video, and even SSE2 was usable but limited until SSSE3). It was obviously architected and planned in advance.
If anything, 3DNow! was likely reactionary - AMD saw opportunity to make a floating-point complement to the (integer-only) MMX set that they could do quickly before Intel makes one.

3DNow! may have been the first x86 floating-point SIMD but it was a hack - same hack as MMX, of course. Those extensions were better than nothing and were quickly available but their use of x87 registers was rather messy (and complicated the x86 ISA and cores) and limited the performance because just 64bit width. What you say is simply incorrect.

Had Intel not made SSEx, someone else would have to. But certainly not like MMX/3DNow.

BTW, 3DNow! was not present in K5 and not even in the K6. It was one of the selling points of K6-II and later chips. I wish K5 had MMX, in the first place...

That s not as simple as this, there s a good summary at wiki, 3DNow had more advantages and was more advanced than you think, among others :

One advantage of 3DNow! is that it is possible to add or multiply the two numbers that are stored in the same register. With SSE, each number can only be combined with a number in the same position in another register. This capability, known as horizontal in Intel terminology, was the major addition to the SSE3 instruction set.

3DNow! - Wikipedia

en.wikipedia.org

 

[Jan Olšan]

That's quite a minor feature that is much less important. And not something that couldn't be added to the SSE/XMM instructions, as you see.

 

[Thunder 57]

You are likely talking of the Slot A version of the Athlon that had SRAM for L2 cache,
the Duron only existed in the socket A version that was released later along with the socketed Athlon, on those CPUs they transited to an on die L2, that tbeing said it was 20% faster in games than the Pentium 3, so seems that it wasnt a limitation.

Slot A had 0.5x to 0.33x CPU frequency L2 SRAM on 64b bus but with double data rate (EV6 bus) while the socketed Athlon Thundebird and Duron had full speed L2 .

Btw the first Pentium 3 also had SRAM for L2.

I haven't lost my memory yet. Unless I just scalped it on eBay. Yes the Slot A Athlon had off die cache that ran at variable fractions of the CPU speed. That is NOT what I am talking about. If anything you aren't sure what you are talking about.

The EV6 bus vs the Intel FSB is not what I am talking about. The EV6 buss at 200/266/333/400 was quite great once DDR RAM came along.

As far as L2 (P3) cache goes:

It is eight-way set-associative and is accessed via a Double Quad Word Wide 256-bit bus, four times as wide as Katmai's.

Link

The AMD Athlon processor’s high-performance cache architecture includes an
integrated dual-ported 128KB split-L1 cache with separate snoop port; an integrated fullspeed, 16-way set-associative, 256KB L2 cache using a 72-bit (64-bit data + 8-bit ECC) interface

Link

AMD has had crap cache other than L1 until Zen. It was a big reason they went with an IMC with K8. Don't get me started on Phenom II's L3 or any of Bulldozer.

That's why the Duron was so close to the Athlon and why Barton with 512KB of L2 didn't do nearly the same as P4 Northwood with 512KB. That and the P4 needed all the cache it could with that crappy decode and long pipeline.

 

[Abwx]

That's why the Duron was so close to the Athlon and why Barton with 512KB of L2 didn't do nearly the same as P4 Northwood with 512KB. That and the P4 needed all the cache it could with that crappy decode and long pipeline.

Yet they were both better than P 3 and Celeron, so that didnt keep them from being the best chips, i had a P3 1GHz and compared to the Duron 700 it couldnt even multitask without the screen freezing until a given task was finished, nothing of the sort with the Duron, the P3 and Celeron did look decent only on a single bench at one time basis and still, on a heavy apps they were both behind a Duron, just look at 3DSMax rendering time on the link below.

Anyway you are focusing on some technical charcteristics while what matter is the end result, that is the final perfs, extending a bus width is useless if the bottleneck is the execution engine.

CC Winstone 2000 & 3D Studio Max - AMD Duron - HardWare.fr

Vous en avez rêvé, AMD l'a fait : le Duron, une version

www.hardware.fr