390
Former Intel CPU engineer details how internal x86-64 efforts were suppressed prior to AMD64's success
(www.tomshardware.com)
this post was submitted on 19 Oct 2024
390 points (99.0% liked)
Technology
59366 readers
3609 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Everybody in the know, knows that x86 64 bit was held back to push Itanium, Intel was all about market segmentation, which is also why Celeron was amputated on for instance RAM compared to Pentium.
Market segmentation has a profit maximization motive. You are not allowed to use cheap parts for things that you are supposed to buy expensive parts for. Itanium was supposed to be the only viable CPU for servers, and keeping x86 32 bit was part of that strategy.
That AMD was successful with 64 bit, and Itanium failed was Karma as deserved for Intel.
Today it's obvious how moronic Intel's policy back then was, because even phones got 64 bit CPU's too back around 2009.
32 bits is simply too much of a limitation for many even pretty trivial tasks. And modern X86 chips are in fact NOT 64 bit anymore, but hybrids that handle tasks with 256 bits routinely, and some even with 512 bits, with instruction extensions that have become standard on both Intel and AMD
When AMD came with Ryzen Threadripper and Epyc, and prices scaled very proportionally to performance, and none were artificially hampered, it was such a nice breath of fresh air.
On a note of technical correctness: That's not what the bitwidth of a CPU is about.
By your account a 386DX would be an 80-bit CPU because it could handle 80-bit floats natively, and the MOS6502 (of C64 fame) a 16-bit processor because it could add two 16-bit integers. Or maybe 32 bits because it could multiply two 16-bit numbers into a 32-bit result?
In reality the MOS6502 is considered an 8-bit CPU, and the 386 a 32-bit one. The "why" gets more complicated, though: The 6502 had a 16 bit address bus and 8 bit data bus, the 368DX a 32 bit address and data bus, the 368SX a 32 bit address bus and 16 bit external data bus.
Or, differently put: Somewhere around the time of the fall of the 8 bit home computer the common understanding of "x-bit CPU" switched from data bus width to address bus width.
...as, not to make this too easy, understood by the instruction set, not the CPU itself: Modern 64 bit processors use pointers which are 64 bit wide, but their address buses usually are narrower. x86_64 only requires 48 bits to be actually usable, the left-over bits are required to be either all ones or all zeroes (enforced by hardware to keep people from bit-hacking and causing forwards compatibility issues, 1/0 IIRC distinguishes between user vs. kernel memory mappings it's been a while since I read the architecture manual). Addressable physical memory might even be lower, again IIRC. 2^48^B are 256TiB no desktop system can fit that much, and I doubt the processors in there could address it.
And how do you figure that? The Intel 80386DX did NOT have any 80 bit instructions at all, the built in math co-processor came with i486. The only instructions on a 80386DX system that would be 80 bit would be to add a 80387 math co-processor.
But you obviously don't count by a few extended instructions, but by the architecture of the CPU as a whole. And in that regard, the Databus is a very significant part, that directly influence the speed and number of clocks of almost everything the CPU does.
You're right, I misremembered.
For those old processors, yes, that's why the 6502 was 8-bit, for modern processors, though? You don't even see it listed on spec sheets. Instead, for the external stuff, you see number of memory controllers and PCIe lanes, while everything internal gets mushed up in IPC. "It's wide enough to not stall the pipeline what more do you want" kind of attitude.
Go look at anything post-2000: 64 bit means that pointers take up 64 bits. 32 bits means that pointers take up 32 bits. 8-bit and 16-bit are completely relegated to microcontrollers, I think keeping the data bus terminology, and soonish they're going to be gone because everything at that scale will be RISC-V, where "RV32I" means... pointers. So does "RV64I" and "RV128I". RV16E was proposed as an April Fool's joke and it's not completely out of the question that it'll happen. In any case there won't be RV8 because CPUs with an 8-bit address bus are pointlessly small, and "the number refers to pointer width" is the terminology of . An RV16 CPU might have a 16 bit data bus, it might have an 8 bit data bus, heck it might have a 256bit data bus because it's actually a DSP and has vector instructions. Sounds like a rare beast but not entirely nonsensical.
Doesn't mean it's any less important, it's just not a good marketing measure,because average people wouldn't understand it anyway, and it wouldn't be correct to measure by the Databus alone.
As I stated it's MORE complex today, not less, as the downvoters of my posts seem to refuse to acknowledge. The first Pentium had a 64 bit Databus for a 32 bit CPU. Exactly because data transfer is extremely important. The first Arm CPU was designed around as fast RAM access/management as possible, and it beat the 386 by several factors, with a tenth the transistors.
Although true, this is a very simplistic way to view it, and not relevant to the actual overall bitwith of the CPU, as I've tried to demonstrate, but people apparently refuse to acknowledge.
But bit width of the Databus is very important, and it was debated heavily weather it was even legal to market the M68008 Sinclair QL as a 32 bit computer, because it only had an 8 bit databus.
But as I stated other factors are equally important, and the decoder is way more important than the core instruction set, and modern higher end decoders operate at 256 bit or more, allowing them to decode multiple ( 4 ) instructions per cycle, again allowing each core to execute multiple instructions per clock, in 2 threads. Without that capability, each core would only be about a third as fast.
To claim that the instruction set determines bit wdth is simplistic, and also you yourself argued against it, because that would mean an i486 would be an 80 bit CPU. And obviously todays CPU's would be 512 bit, because they have 512 bit instructions.
Calling it 64 bit is exclusively meant to distinguish newer CPU's from older 32 bit CPUS, and we've done that since the 90's, claiming that new CPU architectures haven't increased in bit width for 30 years is simply naive and false, because they have in many more significant ways than the base instruction set.
Still I acknowledge that an AARCH64 or AMD64 or i64 CPU are generally called 64 bit, it was never the point to refute that. Only that it's a gross simplification of what modern CPU's have become, and that it's not technically correct.
Let me finish with a question:
With a multi-core CPU where each core is let's just say 64 bit, how many bits is the whole CPU package? Which is what we call the "CPU" today, when saying CPU we are not generally talking about the individual cores, because then it would have to be plural.
The reason you're getting downvoted is because you're saying that "64-bit CPU" means something different than is universally acknowledged that it means. It means pointer width.
Yes, other numbers are important. Yes, other numbers can be listed in places. No, it's not what people mean when they say "X-bit CPU".
RV128 exists. It refers to pointer width. Crays existed, by your account they were gazillion-bit machines because they had quite chunky vector lengths. Your Ryzen does not have a larger "databus" than a Cray1 which had 4096 bit (you read that right) vector registers. They were never called 4096 bit machines, they Cray1 has a 64-bit architecture because that's the pointer width.
Yes, the terminology differs when it comes to 8 vs. 16-bit microcontrollers. But just because data bus is that important there (and 8-bit pointers don't make any practical sense) doesn't mean that anyone is calling a Cray a 4096 bit architecture. You might call them 4096 bit vector machines, and you're free to call anything with AVX2 a 256-bit SIMD machine (though you might actually be looking at 2x 128-bit ALUs), but neither makes them 64-bit architectures. Why? Because language is meant for communication and you don't get to have your own private definition of terms: Unless otherwise specified, the number stated is the number of bits in a pointer.
https://en.wikipedia.org/wiki/64-bit_computing
It also states Address bus, but as I mentioned before, that doesn't exist. So it boils down to instruction set as a whole requiring 64 bit processor registers and Databus.
Obviously 64 bits means registers are 64 bit, the addresses are therefore also 64 bit, otherwise it would require type casting every time you need to make calculations on them. But it's the ability to handle 64 bit registers in general that counts, not the address registers. which is merely a byproduct.
You were arguing the definition of "X-bit CPU". We're not talking about "X-bit ALU". It's also not up to contention that "A 64-bit integer is 64 bit wide". So, to the statement:
This does not say which of "processor register, address buses, or data buses" applies to CPU and which to ALU.
Having 64 bit registers doesn't necessitate that you have 64 bit addresses. It's common, incredibly common, for the integer registers to match the pointer width but there's no hard requirement in theory or practice. It's about as arbitrary a rule as "Instruction length must be wider than the register size", so that immediate constants fit into the instruction stream, makes sense doesn't it... and then along come RISC architectures and split load immediate instructions into two.
Processors don't typecast. Please stop talking.
Which is why it's such a pain, because you have to do it manually:
https://lemire.me/blog/2021/10/21/converting-binary-floating-point-numbers-to-integers/
I'm sorry are we somehow assuming floating-point pointers, now, of course you need to convert there. "casting" is a specific thing you do in C which may or may not involve conversion of actual data. Processors don't speak C. Processors don't have a type system.
You can use 32-bit pointers in x86_64 long mode, no issue. You don't even need to bit-fiddle:
mov rax, [esi]is perfectly legal. Opcode0x67488B06. Dereferencingrsiwould be0x488B06."floating-point pointers" is not a thing:
No it's not:
https://en.wikipedia.org/wiki/Type_conversion
You don't even have a clue, you are just talking trash.
In assembly you don't generally talk about pointers, but address modes. Like register, immediate or memory (indirect).
Have you ever actually been programming any serious assembly? Because you sure don't sound like it.
Great! Now please explain how opcodes are expressions. Also, what processor instruction a cast from one pointer type to another pointer type corresponds to.
You are way out of your depth here. Have you even implemented a compiler.
EDIT:
Oh cute edit to make to make my response look bad retroactively.
But as you wanted to get pedantic: A pointer is a value which is intended to be dereferenced, that (hopefully) corresponds to a valid memory address. "address", "pointer", "reference", it's a matter of taste which one you use. It exists "in assembly" just as "an index" exists in C: Not because it's a language feature, but because it's a concept you use when writing in the language. And yes I speak pretty fluent x86, at least the non-SIMD part. Did I mention that I was there, at ground zero "why is is thing not compiling in 64 bit mode" times, fixing code?
Now, back to my question:
Figuring out the answer to that will tell you everything you need to know about where you went wrong. Where you went from talking about actual concepts to arguing semantics.
Where did you get that from? Because that's false, please show me dokumentation for that.
64 bit always meant the ability to handle 64 bit wide instructions, and because the architecture is 64 bit, the pointers INTERNALLY are 64 bit, but effectively they are only for instance 40 bit when accessing data.
Your claim about pointer width simply doesn't make any sense.
That the CPU should be called by a single aspect they can't actually handle!!! That's moronic.
People literally use the word "literally" to mean figuratively. It doesn't make any sense. One might even call it moronic.
But it's the way it's done. Deal with it.
No that's not true, it's way way more complex than that, some consider the data bus the best measure, another could be decoder. I could also have called a normal CPU bitwidth as depending on how many cores it has, each core handling up to 4 instructions per cycle, could be 256 bit, with an average 8 core CPU that would be 2048 bit.
There are several ways to evaluate like Databus, ALU, Decoder etc, but most ways to measure it reasonably hover around the 256 bit, and none below 128 bit.
There is simply no reasonable way to argue a modern Ryzen CPU or Intel equivalent is below 128 bit.
There absolutely is, and the person you responded to made it incredibly clear: address width. Yeah, we only use 48-bit addresses, but addresses are 64-bit, and that's the key difference that the majority of the market understands between 32-bit and 64-bit processors. The discussion around "32-bit compatibility" is all about address size.
And there's also instruction size. Yes, the data it operates on may be bigger than 64-bit, but the instructions are capped at 64-bit. With either definition, current CPUs are clearly 64-bit.
But perhaps the most important piece here is consumer marketing. Modern CPUs are marketed as 64-bit (based on both of the above), and that's what the vast majority of people understand the term to mean. There's no point in coming up with another number, because that's not what the industry means when they say a CPU is 64-bit or 32-bit.
Edited for clarity
You are stating the register width, which is irrelevant to the width of the address bus. But that doesn't make a shred of sense. it's like claiming a road is 40000 km long around the globe, it's just not finished, so you can only drive on a few km of it. The registers are 64 bit, but "only" 40 can be used. Enough to address 1 Terabyte of RAM.
If you want to measure by Address width we don't have a single 64 bit CPU, because there doesn't exist a 64 bit CPU that has a 64 bit Address bus.
Yes they have, and that's what the vast majority of people mean when they say a CPU is 32-bit or 64-bit. It was especially important in the transition from 32-bit to 64-bit because of all the SW changes that needed to be made to support 64-bit addresses. It was a huge thing in the early 2000s, and that is where the nomenclature comes from.
Before that big switch, it was a bit more marketing than anything else and frequently referred to the size of the data the CPU operated on. But during and after that switch, it shifted to address sizes, and instructions (not including the data) are also 64-bit. The main difference w/ AVX vs a "normal" instruction is the size of the registers used, which can be up to 512-bit, vs a "normal" 64-bit register. But the instruction remains 64-bit, at least as far as the rest of the system is concerned.
Hence why CPUs are 64-bit, all of the interface between the CPU and the rest of the system is with 64-bit instructions and 64-bit addresses. Whether the CPU does something fancy under the hood w/ more than 64-bits (i.e. registers and parallel processing) is entirely irrelevant, the interface is 64-bit, therefore it's 64-bit.
Nobody ever called the purely 8 bit Motorola M6800, MOSTech 6502, Zilog Z80, ot the Intel 8080 16 bit computers for having a 16 bit address bus. They were 8 bit instruction and data bus, and were called 8 bit chips. The purely 16 bit Intel 8086 wasn't called a 20 bit CPU for having a 20 bit Address bus, it was called a 16 bit CPU for having 16 bit instruction set and databus. Or the Motorola M68000 a 24 bit CPU for having a 24 bit adress bus, it was a 32 bit CPU for having a 32 bit instruction set.
I have no idea how you are upvoted, because your claim tha CPUs are called by their address bus bit length is decidedly false.
The most common is to use the DATA-bus or instruction set, and now also the instruction decoder and other things, because the complexity has evolved. But no 64 bit CPU has a 64 bit address bus, because that would be ridiculous.
Back in the day, it was mostly instruction set, then it became instruction set / DATA-bus. Today it's way way more complex, and we may call it x86-64, but that's the instruction set, the modern x86-64 CPU is not 64 bit anymore. They are hybrids of many bit widths.
Show me just ONE example of a CPU that was called by its address bus.
https://people.ece.ubc.ca/edc/379.jan2000/lectures/lec2.pdf
Tell me when 8086 and 8088 were called 20 bit CPU's!!
https://www.alldatasheet.com/datasheet-pdf/view/82483/MOTOROLA/MC6800.html
The 6800 was an 8 bit CPU with 16 bit Adress bus as was the 6502/6510.
https://en.wikipedia.org/wiki/Motorola_68000
The 68000 is here correctly called 16/32 because it'a a 16 bit DATAbus and 32 bit instruction set.
The Address bus is 24 bit, but never has a CPU been called 20 ot 24 bit because of their address bus, despite many 16 bit CPU's have had address busses of that length.
Incidentally, the MOS 6510 in the Commodore 64, had an extra 17th address bit, enabling it to use ROM and cartridges together with the 64 KB RAM. It would be absolutely ridiculous to call it either a 16 or 17 bit computer, and by no accepted standard would it be called that.
I guess you know more about hardware nomenclature Linux kernel developers, because they call modern Intel/AMD and ARM CPUs amd64 and aarch64, respectively.
AMD64 is the name of the instruction set they program to, it has nothing to do with how many bit the CPU is. Obviously the core instruction set is 64 bit, but as I've tried to explain, a chips bit width is not realistically determined by instruction set alone anymore.
Although they are almost identical, the equivalent Intel to AMD64 is called i64.
AArch64 Is the Arm Architecture family 64, again the instruction set you program for, and not the bit width of the CPU.
None of those describe the address bus width either. i64 and AMD64 and AARCH64 come will all sorts of different address bus widths, all of which are less than 64 bit wide.
https://www.tomshardware.com/reviews/processor-cpu-apu-specifications-upgrade,3566-2.html
Although this is a bit dated, the latest I heard was 48 bit address bus, which would surpass the above from 2013 by a factor of 256.
Obviously none of these 64 bit architecture CPU's are called neither 40 or 48 bit.
Sure, but that was a long time ago. Lithography marketing also used to make sense when it was actually based on real measurements, but times change.
All those chips you're talking about were from >40 years ago. Times change.
Sure, yet when someone describes a CPU, we talk about the instruction set, so we talk about 32-bit vs 64-bit instructions. That's how the terminology works.
I never denied that, what I denied was the ridiculous idea that Address bus was a meaningful measure. AMD64 is a 64 bit instruction set by definition, but a modern Ryzen CPU is so much more than just AMD64. And the same is true for the competition.
Originally an AMD64 CPU was single core single threaded. This is far from true today, so obviously since the CPU can handle multiple instructions on multiple cores, the "CPU Package" is also necessarily wider.
I have no idea what has gone wrong here? I'm not denying that a modern Intel or AMD or Arm CPU generally is called a 64 bit CPU.
I'm just stating that if they had to be measured by their actual capabilities, a modern Ryzen CPU for instance, is actually closer to being a 256 bit CPU, and that's per core!. In part due to technologies that make them able to execute several instructions in a single clock cycle, that operate on way wider busses than older CPU's, that encoded only a single thread per core.
But there can be absolutely no doubt that Address bus was NEVER used to determine the bit width of a CPU, that would simply be ridiculous, as it ONLY determines addressable RAM and nothing else.
Those easy to understand examples were only to show how claiming address bus can be a meaningful measure for the bit width of a CPU is ridiculous.
Also the AMD64 is only part of the instruction set of a modern Ryzen CPU, so although AMD64 definitely is a 64 bit instruction set, it only describes one part of the CPU. It also supports: x87, MMX, SSE, SSE2, SSE3, SSSE3, SSE4. 1, SSE4. 2, AES, CLMUL, AVX, AVX2, FMA3, CVT16/F16C, ABM, BMI1, BMI2, SHA.
Many of which have way wider instructions than 64 bit, AVX2 for instance supports 512 bit math.
That seems to be exactly what you're arguing about, unless I have misread this entire thread.
If we want to highlight other capabilities, we should use different terminology than "X-bit" because that has been pretty much universally agreed upon to refer to instruction sizes and addresses, not data pipelines. And we do that, product spec sheets refer to extensions to point out the unique capabilities they offer (e.g. Intel was pretty famous for supporting AVX-512 almost 10 years before AMD).
That said, now that 32-bit is essentially dead, the "X-bit" marker is essentially dead, and saying something is 256-bit or whatever today is just going to confuse people. People have gotten into the habit if talking about specific capabilities if it's relevant (which it isn't for most people, who just care about "IPC").
That was kind of the point, it's ridiculous to think a modern CPU hasn't evolved dramatically since the introduction of mainstream 64 bit in 2003.
It's still called 64 bit, but there are so many developments.
Exactly, and that is achieved by a modern core operating at about 256 bit internally, to achieve faster execution.
I'm not arguing it's wrong to call it 64 bit, because there is no "true" bit width to call it. So we might as well still call it 64 bit, because it describes the core instruction set. (not just pointers as was claimed by someone else) My point was just that it doesn't really describe the dramatic development of the CPU as a whole, and even the individual cores are more complex in hardware, despite the main instruction set remains the same.
See also: ECC memory.
Sometimes for some reason, there's no limit. Like the cheap i3-8100 can use ECC memory
AMD allowed procssors to use ECC memory since Ryzen so the the jig was up.
It was also a big surprise when Intel just gave up. The industry was getting settled in for a David v Goliath battle, and then Goliath said this David kid was right.
Yes, I absolutely thought Intel would make their own, and AMD would lose the fight.
But maybe Intel couldn't do that, because AMD had patented it already, and whatever Intel did, it could be called a copy of that.
Anyways it's great to see AMD finally is doing well and finally is profitable. I just never expected Intel to fail as badly as they are? So unless they fight their way to profitability again, we may be in the same boat again as we were when Intel was solo on X86?
But then again, maybe x86 is becoming obsolete, as Arm is getting ever more competitive.
Right, I think the future isn't Intel v AMD, it's AMD v ARM v RISC-V. Might be hard to break into the desktop and laptop space, but Linux servers don't have the same backwards compatibility issues with x86. That's a huge market.
Intel as a company isn't going anywhere any time soon; they're just too big, with too many resources, not to do at least OK.
They have serious challenges in their approach and performance to engineering, but short of merging with someone else they'll find their niche. For as long as x86-derived architectures remain current (i.e. if AMD is still chugging along with them) they'll continue to put out their own chips, and occasionally they'll manage to get an edge.
The real question would be what happens if x86 finally ceases to be viable. In theory there's nothing stopping Intel (or AMD) pivoting to ARM or RISC-V (or fucking POWER for that matter) if that's where the market goes. Losing the patent/licensing edge would sting, though.
I hated that you had to choose, virtualization or overclocking so much. Among a lot of other forced limitation crap from intel.
A bit like cheap mobile phones had a too small ssd and buying one at least "normal" sized bumped everything else (camera, cpu, etc) up too, including price ofc.