The Rise, Fall and Revival of AMD (2020) | Hacker News Reader

The Rise, Fall and Revival of AMD (2020)

techspot.com

92 pointsneogodless44 comments

It's easy to underestimate just how difficult processor design is and just how quickly it has advanced over the last few decades.

First you have to design an incredibly complex custom hardware processor to implement your ISA of choice with various speed optimisations. Then you have to convert it to a gate-level model and simulate it to make sure it's reliable.

Then you have to produce a physical/electromagnetic model which juggles capacitance, loading, switching transients, transmission line reflections, and interconnect length, and somehow still works for all possible conditions with good tolerances to maximise yield.

Then you have to handle all the visual/UV/soft X-ray optics, diffusion chemistry, and process management to build the wafers at scale.

It's not a particularly clean or sustainable industry. But even so - the finished chips are the absolute pinnacle of many simultaneous engineering and management disciplines.

You also have to be able to verify (at least the important parts) of the chip

It is the coolest shit we have, right? Do we have anything else that is as advanced/complex/refined/whatever-you-wanna-call-it?

It's certainly the highest precision manufacturing you are going to find in a consumer widget. It's hard to think of anything manufactured to sub-nanometer tolerances. Maybe some scientific instruments like large telescope lenses?

There is software for the synthesis and verification parts.

For the fabrication, you contact your fab of choice.

Yes, this is expensive, but designing a processor is not unlike designing a large distributed software system. It's all software, and the rest is already taken care of mostly.

I look forward to having the time to read this in full. My first PC was an AMD 286/12 with 1mb memory. Later on, the best CPU I'd owned to date was the AMD K6-450. In 2010, I purchased a Phenom II X2 BE, unlocked it to X4 and it served me well for 8 years! I finally replaced it in 2018.

I avoided the bulldozer/piledriver series chips, and got a first gen Ryzen 5 1400. It's a fantastic chip, very fast especially at the price, and will last me years I'm sure.

In the late 90's I also did good business selling computers based on the AMD 486 DX4/120's and 133's. Those things were so fast for the time.

Got myself AMD FX-8350 a decade ago, used it daily with minor overclocking and replaced with Ryzen 3950X last year. This old FX-8350 went into dedicated Linux machine for things I weren’t able to do with virtual one. I am curious how long will it run:-)

Those FX series had a really bad reputation at the time, but it seems the extra cores have helped them age better than might have been expected. My old FX 6300 is also in a Linux box now hosting a Jellyfin server and it does quite well!

Ha, I was just about to mention the FX-6300 I have running in a server that hosts 3 Minecraft instances. It's not the fastest thing ever but it gets the job done. I was going to replace it when it got too slow, but that hasn't happened yet.

My first computer was also an AMD? But that wasn't because I loved AMD or knew anything about processors. I believe my parents bought me that one because it was cheaper than Intel.

Fast forward to today, my current desktop is an AMD Ryzen 3600X.

Did you happen to own any AMD stocks back then?

My first computer was also purchased by my parents. Nope, no stock.

Offtopic: Most might be thinking fully FLOSS is something that's only happened in software, or powerless hardware. But the thing is server and desktop grade powerful and fully FLOSS hardware with PCIe 4.0 and DDR4 ECC memory are available right now to customers and businesses. Not as powerful as AMD. AMD will never be open though. It's not x86. POWER9. FSF RYF certified.

https://www.raptorcs.com/

#talos-workstation on freenode IRC to chat with users.

I like the idea of FSF-certified hardware, but aren't POWER9 machines ten times the price of the average desktop?

Maybe not 10x but they’re certainly more than a similarly spec’d dev workstation. I guess the idea is that developers are the target market and they have a bit more disposable income on average and are more likely to want to pay extra for something like this.

What 'smcl' commented, also all prices are in this link, scroll down:

https://www.raptorcs.com/

Thanks. Looks like their cheapest system goes for $3,404.59. It has VGA output, but no DVI/HDMI/DisplayPort.

Strange: the price on their systems seems to have gone up since the last time I saw them.

I wonder if shipping issues have caused their prices to skyrocket recently?

Doing a quick spec out of a Talos II that is similar[0] to my Threadripper 3970x workstation puts the price at ~2x.

And a typical TR workstation is 2x the price of an "average desktop" so maybe a 4x multiplier?

You could price out a desktop for your own comparison. I don't think it's fair to compare something explicitly designed to be a workstation to a desktop-class machine.

It's both fair and not; if you want a desktop, the cheapest POWER desktop is 3k, and at 4 cores it's arguably not competing against TR. Yes it's IO throughput will put a desktop class part to shame, but you can't go any cheaper.

Agreed. There isn't really a viable desktop-class offering from Raptor. This is a shame.

For those who have need for the workstation-ish features Raptor offers in its Talos line, the premium isn't too great over a roughly equivalent x86 workstation-class machine.

Any news or update on POWER10?

POWER10 switched to a serial memory interface (OMI), IIUC Raptor won't touch any of this stuff until there are FOSS memory modules, which there aren't currently.

As much as I enjoy the idea, I can not justify the price, especially considering the huge breadth of software NOT available.

While it would work for my purposes as a server or a dev workstation, you still have to deal with a fairly niche processor architecture.

Ironically, though, the development of software for more modern POWER systems has kept some of the Apple PowerPC systems more up to date (assuming they are running linux, like Void).

Intel are so lucky that their fab issues have come at a time when AMD physically cannot even dream of shipping enough chips to make them bleed.

That being said I think Intel only have to get their act together once to get right back in there - they have dragged 14nm all the way, their uarchs can't be that bad.

Huh, TIL: the K6 came from an acquisition.

> Graphics Core Next (GCN). This design would last for nearly 8 years […] still in use today as the integrated GPU

And in desktop GPUs.

The goddamn power of marketing! Even tech writers seem to assume that "RDNA" is some kind of revolutionary from-scratch change. It's still the same ISA!! Look at the drivers and compilers. It's simply the regular evolution, just with a rebrand.

(upd: even later in this article they do say it's "a significant reworking of GCN"… well, why did the first mention sound kinda like it wasn't acknowledging this?)

> The goddamn power of marketing! Even tech writers seem to assume that "RDNA" is some kind of revolutionary from-scratch change. It's still the same ISA!! Look at the drivers and compilers. It's simply the regular evolution, just with a rebrand.

GCN 1.2 changed the opcodes from GCN 1.0. If the machine code doesn't line up anymore, is it really fare to call it the same ISA?

In the case of RDNA: its all 32-wide SIMD instead of 64-wide SIMD. That dramatic difference completely destroys the bpermute / permute / DPP assembly instructions (https://gpuopen.com/learn/amd-gcn-assembly-cross-lane-operat...), which have gone from 64-way permutes (in GCN) into 32-way permutes (in RDNA).

GCN 1.2 was already looking pretty different from GCN 1.0, I'd say RDNA absolutely deserves the name "new ISA".

After all, GCN 1.2 is about as similar to GCN 1.0 as 8080 was to 8086 (same assembly language and registers, but new opcodes). I think most people are willing to call 8080 and 8086 different ISAs.

RDNA is extremely different at a base level because of that 32x SIMD vs 64x SIMD.

In the gpu world it's common to change instructions in iterative updates, since assembly is not really a popular way to program these. I think it's fair to call it the same as long as the compiler backends and drivers are still the same ones. Still amdgpu in LLVM, still RadeonSI and RADV in Mesa, etc.

It is hard to understand today, but back then every manufacturer of a chip design needed a "second source", a fully independent competitor who could supply a compatible chip. Otherwise it would be far too risky to design their chip into your product. What if they folded? AMD, by cloning the 8080 and other chips, gave Intel what it had to have to be a serious supplier for serious products.

The 6502 succeeded by being 1/10 the price, and was designed into products of companies that embraced extreme risk.

Second sourcing was a requirement imposed by the DoD.

A simple, yet very powerful, weapon for ensuring a healthy economy. Having at minimum two potential sources for any component in the economy, with anything that can't afford that classified as a Utility and regulated as such.

From the single or none fulfillment of broadband options in the US even in dense suburban areas, let alone rural areas, it is clear that there are places in dire need of similar government market regulation.

The DoD requirement enabled commercial customers to specify it, too. The supply-chain nightmares experienced by makers of modern equipment, with microcontrollers only just designed in dropping out of production, generates nostalgia for the power to require a second source.

Portability of embedded-code source, standardized ISAs, and ability to turn around a new circuit board design in a week, and FPGAs all help make lack of second sources survivable.

At what point were 6502 clones everywhere? Long after the Apple II and C64 selected it?

Probably the most famous one would be the Ricoh CPU in the NES, supposedly as close to a 6502 as they could get without patent infringement, so it's missing a few odd instructions.

I suspect second-sourcing must have been going on fairly early in the 6502's lifecycle though. Commodore OWNED MOS by the time the 6502 was a shipping, finished product, so they're obviously going to dogfood it on the PET/CBM/VIC series.

A 1977 design like the Apple II had a bit of plausible deniability-- they may not have known that Commodore was going to be a major player in the small-computer market when they did their design, and then they were stuck with the 6502 and derivatives for backwards compatibility for the rest of the series.

But then you have machines like the 1979-release Atari 400/800. By then it was obvious you're buying from a potential hostile competitor. There must have been enough confidence in their business practices or the availability of second-sourcing that they weren't afraid of being cut out of the market.

CPU-World shows a bunch of 6502s from different manufacturers with date codes in the 1980s, but I can't comment much more about if they were clones, or license-built second sources, or what.

Aside from the NES, I can't think of any major greenfield 6502 products much after 1981, but that's likely less about second sourcing and more about a deluge of cheap Z80s and 8088s making the 6502's low price less compelling.

I was always intrigued by AMD processors in the 90s, although pretty much stuck with Intel.

I built an Athlon 64 PC in 2004 - it was the first 64-bit consumer processor. Intel was going down the Itanium path, and AMD just added 64-bit extensions to X86, which ended up being the winning strategy. I remember writing an article about it for my job at the time.

That served me well until I upgraded to a Phenom II X6 in 2010. I was super excited by the prospect of having 6 cores, even though I had no real use for that many. That PC lasted me for over 10 years, although I did upgrade the RAM, added an SSD, and upgraded the GPU twice. Was still a competent PC at the end for both gaming and software development, and managed to part it out for a decent amount of money.

Late last year I finally succumbed the the Ryzen itch and upgraded to a Ryzen 5600x. It's a fantastic platform, and hopefully that PC will last me another 10 years with only minor upgrades.

> Imageon, the handheld graphics division of ATI, was sold to Qualcomm in a paltry $65 million deal. That division is now named Adreno, an anagram of "Radeon" and an integral component of the Snapdragon SoC (!).

It would be hard to come up with a corporate story that didn't have such missteps. Now, Samsung has licensed Radeons to be in their mobile SoCs.

Article makes me want to fire up my k6.

My new ryzen 2700 isn't as fast as my 8350... because pcie lanes. My m2 drives are more than 500MB/s slower on the 450 motherboard onboard m2 slots. Should have gone with a threadripper. Still, I'm always a underdog fan and don't like paying the tariff for owning the most popular hardware. I compare specs and teviews. But availability is another issue...

Does anyone know if AMD has any sort of strategy around Thunderbolt for their mobile chips? Are they stuck because of Intel IP issues?

Their desktop lines are basically the best you can get (assuming you can find anything in stock), but even the latest 5000 series laptops conspicuously lack Thunderbolt. This makes them a tough sell for some, as Thunderbolt has become the defacto standard for modern docking stations...

USB4 is essentially a more open Thunderbolt, so I think their strategy will be to implement that instead.

> the top-end version of the Am386DX launched in 1989 at 40 MHz.

I had this CPU. 386dx 40Mhz. What a beast. It was my first PC that I purchased myself. It was also my first online purchase. I bought it off of FidoNet through a BBS. I'm still amazed that worked. Cash-on-delivery (COD). $50. Probably took an entire month getting to me via UPS.

Ever since I got to be the one who chooses the specs, I always tried do buy/build PCs with AMD processors. My logic is simple: always work against the most dominant player. Be it Intel with CPUs, Microsoft with OSs or Putin in russian politics.

I can't get past this sentence:

> From these humble beginnings, and a quick move from Santa Clara to Sunnyvale (Silicon Valley in California)

Is the rest of the article better than this?

For those outside the Bay Area: Santa Clara and Sunnyvale are physically adjacent, and I'm guessing all of the sites where AMD's HQ has been historically located are probably within 0.5 miles of each other near that border.

Meh, people generally get confused by California's municipalities.

"Los Angeles is so sprawling"....I mean yeah, it definitely is; but it's especially so when you erroneously believe Pasadena and Santa Monica to be part of LA. Or worse so, parts of OC, Ventura and the IE.

Reply on news.ycombinator.com