Hacker News Reader

Comment by llamaLord | Hacker News Reader

To be fair, they should have seen Ryzen coming, any long-term AMD user knew years before Ryzen landed that it was going to be a good core because AMD were very vocal about how badly wrong they bet with Bulldozer (previous core family).

AMD bet BIG on the software industry leaning in heavily on massive thread-counts over high throughput, single-threaded usage... But it never happened so the cores tanked.

It was never a secret WHY that generation of core sucked, and it was relatively clear what AMD needed to do to fix the problem, and they were VERY vocal about "doing the thing" once it became clear their bet wasn't paying off.

View on HN

(I'm curious about this story, as I am unfamiliar with it.)

Why did that generation of core (Bulldozer) suck?

What was it that AMD needed to do to fix the problem?

(Links to relevant stories would be sufficient for me!)

Chips and Cheese has probably the most in depth publicly available dive for the tech reasons why Bulldozer was the way it was:

https://chipsandcheese.com/2023/01/22/bulldozer-amds-crash-m...

https://chipsandcheese.com/2023/01/24/bulldozer-amds-crash-m...

---

From a consumer perspective, Bulldozer and revisions as compared to Skylake and revisions were:

+ comparable on highly multi-threaded loads

+ cheaper

- significantly behind on less multi-threaded loads

- had 1 set of FPUs per 2 cores, so workloads with lots of floating point calculations were also weaker

- Most intensive consumer software was single or a very small number of thread focused still (this was also a problem for Intel in trying to get people to buy more expensive i7s/i9s over i5s in those days)

Bulldozer was contemporary with Sandy Bridge, not Skylake. Piledriver competed with Ivy Bridge and Haswell. The next Construction cores (Steamroller and Excavator) were only found in APUs, and not in desktop FX parts. Around the time of Skylake, AMD didn't have a meaningful presence in the desktop space. All they were selling was quad-core APUs based on minor revisions of Bulldozer, and the highly outdated FX-x3xx Piledrivers.

The Bulldozer design had a few main issues.

1.Bulldozer had a very long pipeline akin to a Pentium 4. This allows for highclocks but comparatively little work being done per cycle vs their competition. Since clocks have a ceiling around 5GHz they could never push the clocks high enough to compete with intel. 2.They used a odd core design with 1 FPU for every 2 integer unit instead of the normal 1:1 that we have seen on every x86 since the i486. This leads to very weak FPU performance needed for many professional applications. Conversely it allowed for very competitive performance on highly threaded integer applications like rendering. This decision was probably under the assumption APUs would integrate their GPUs better and software would be written with it in mind since a GPU easily out does a CPUs FPU but it requires more programming. This didn't come to be. 3. They were stuck using Global Foundries due to previous contracts when they spun it off requiring AMD use GloFlo. This became a anchor as Gloflo fell behind market competitors like TSMC. Leaving AMD stuck on 32nm for a long while, until gloflo got 14nm and eventually AMD got out of the contract between zen 1-2.

bonus: Many IC designers have bemoaned how much of bulldozers design was automated with little hand modifications which tends to lead to a less optimized design. 3. 3.

There's been lots written about this but this is my opinion.

Bulldozer seemed to be designed under the assumption heavy floating point work would be done on the GPU (APU) which all early construction cores had built in. But no one is going to rewrite all of their software to take advantage of the iGPU that isn't present in existing CPUs and isn't present in the majority of CPUs (Intel) so it sort of smelt like Intel's itantic moment, only worse.

I think they were desperate to see some near term return on the money they spent on buying ATI. ATI wasn't a bad idea for a purchase but they seemed to heavily overpay for it which probably really clouded management's judgement.

I thought it was a bad idea when I first read of it. It reminded me of Intel's Netburst (Pentium 4) architecture.

They've seen all that. You don't have to have an MBA or a MIT degree to plot projected performance of your or your competitors' chips.

It was process failures. Their fabs couldn't fab the designs. Tiger lake was what? 4 years late?

Reply on news.ycombinator.com