Semiconductor world - CPU/GPU Wars

Great interview with a server engineer with 20 years experience

  1. Biz application lifetime is about 25 years!! servers still run applications go back to windows 2000. These will not work in AMD at all.
  2. Server engineers overall have moved to 3 year cycle compared to 7 year cycle earlier in their agreement with Intel. Ideally server engineers want to move immediately. Lesser cycle gives them better choices.
  3. Current EPYC CPUs are being validated. Server folks validate new component before bringing it online. But covid has pushed that out.
  4. Older applications may not move to AMD.
  5. The cost of switching is huge - unless AMD can give 4x perf do not expect any switching from intel to AMD. The cost is the down time in servers. 10% 20% better perf is not enough.
  6. Risk is premium - Money is not the problem. The cost is the down time. In healthcare, a down server can cause death.
  7. Older applications will not be switched to AMD
  8. AMD cannot go above 50% - Intel is well entrenched and whatever applications current work with intel based servers
  9. Enterprise servers - At current application counts - Even if Intel kept selling the same thing for 10 years AMD cannot go above 50% market share!! Steady perf is premium. No one is going to move applications to a new platform now- supply chain + stability + security is important.
  10. Most AMD can do is 35%!!
  11. If AMD has to go above 50% - they have to rely on amazon/microsoft etc. That too even 10 years is optimistic. Not to mention need for new applications in enterprise market to bring this change. Even for ARM. All things same - 30-40 years.
  12. When he was running the show in a healthcare server - Moving from intel 32 bit xeon to 64 bit xeon took them 8 years.
  13. Intel security issues caused the move to zen3 but delayed due to covid. Otherwise, they would have waited for things to settle down. May be zen 6 in 2024!!
  14. AMD is very competitive in 2020 than before with their support + new hardware etc.
  15. Now that cycles are 3 years - Cost of purchase starts to matter now compared to Total Cost of Operations
  16. Enterprise buy Intel because they have to - Old applications. Just keep things running. 14nm works too. But the margins keep reducing for intel.
    GPU
  17. NVDA CUDA has 10 year head start - Very difficult to switch to AMD - Platform is sticky.
  18. Radeons have a shot in supercomputing space.
  19. One thing going for AMD is openness. In general the industry prefers open standards compliant stack. NVDA tries to tie users to its stack + applications - A problem with NVDA GPU software. AMD is open friendly.

    Much more. I am next going to see how much money does 35% mean.
4 Likes