3 ms·

I assume it's almost zero migration for 99% of use cases from an Intel instance to an Epyc Rome instance. But if they figured out performance with Graviton2, w

I assume it's almost zero migration for 99% of use cases from an Intel instance to an Epyc Rome instance. But if they figured out performance with Graviton2, what makes you think ARM isn't the clear future for server performance per dollar?

kllrnohj 6y ago
What makes you think ARM is the clear future? It's easier to do a wide decoder thanks to ARMv8's fixed-length instructions, but that's more or less the only difference that results from the ISA. And only Apple is even taking advantage of that currently - all the ARM CPU designs are 4-wide decode. Same as pretty much every modern x86 CPU. Do you think Amazon will be better at design & fabrication than Intel (when they get their shit together) or AMD? Do you expect Amazon to actually throw Apple's level of R&D money & acquisitions at the problem? And when/if they do, why would you expect any meaningful difference at the end of the day on anything? People love to talk about supposed ARM power efficiency, but the 64-core Epyc Rome at 200w is around 3w per CPU core. Which is half the power draw of an M1 firestorm core. The M1's firestorm is also faster in such a comparison, but the point is power isn't a magical ARM advantage, and x86 isn't inherently some crazy power hog. Phoronix put Graviton2 to the test vs. Epyc Rome and the results ain't pretty: https://www.phoronix.com/scan.php?page=article&item=epyc-vs-graviton2&num=8 https://www.phoronix.com/scan.php?page=article&item=epyc-vs-... "When taking the geometric mean of all these benchmarks, the EPYC 7742 without SMT enabled was about 46% faster than the Graviton2 bare metal performance. The EPYC 7742 with SMT (128 threads) increased the lead to about 51%, due to not all of the benchmarks being multi-thread focused." Graviton2 being a bit cheaper doesn't mean much when you need 50% more instances. But right now there's only a single company that makes ARM look good and that's Apple. And Apple hasn't been in the server game for a long, long time now. Everyone else's ARM CPU cores pretty much suck. Maybe Nvidia's recent acquisition will change things, who knows. But at the end of the day if AMD keeps things up, or Intel gets back on track, there really doesn't look to be a bright future for ARM outside of Apple's ecosystem and the existing ARM markets.
- EricE 6y ago
  I think the M1 will wake Amazon up (if they aren't already) to the incredible advantage of being able to customize a CPU to your needs. See the discussion about speed gains Apple got from accelerating common calls in their OS. You will never convince a bin part/least common denominator manufacturer like Intel or AMD to do that. Just because other ARM cores suck today doesn't mean they have to forever. Apple's don't. They took it seriously. Perhaps Amazon is too and we are just at the start of their journey. It took Apple over 10 years to get to where we are with the M1. You cited one of the significant contributors to performance - the 8-wide decode. x86 is hamstrung to 4 because of legacy. We aren't at the beginning of the story with ARM for performance, but ARM certainly isn't nearly as hamstrung out the gate by the legacy of x86 either. Heck, is there anyone making a pure 64 bit x64 chip? There's a bunch of overhead right there. Your right that ARM isn't magical - but ARM does have the potential for significantly more runway and headroom. The trade off is backwards compatibility. As most code continues to move further and further away from direct hardware interaction, is that backwards compatibility as valuable overall as it was 20 years ago? 10 years ago? I guess we will find out :)
  - theevilsharpie 6y ago
    > I think the M1 will wake Amazon up (if they aren't already) to the incredible advantage of being able to customize a CPU to your needs. See the discussion about speed gains Apple got from accelerating common calls in their OS. You will never convince a bin part/least common denominator manufacturer like Intel or AMD to do that. AMD has a semi-custom division that will do exactly that for anyone capable of paying: https://www.amd.com/en/products/semi-custom-solutions https://www.amd.com/en/products/semi-custom-solutions
  - kllrnohj 6y ago
    > See the discussion about speed gains Apple got from accelerating common calls in their OS. You will never convince a bin part/least common denominator manufacturer like Intel or AMD to do that. Of course you will because that happens all the time. How do you think we ended up with VT-x and friends? Intel took a use case that reached enough usage and added specialized instructions for it. This has happened a ton over the years on x86. See also AES-NI for a more application-specific addition. In addition to obviously the huge amount of SIMD experimentation. This is not fruit Apple discovered. AMD & Intel haven't been leaving this performance on the table the last 20+ years. Hell the constant addition of instructions for certain workloads is a major reason x86 is as huge & complex as it is.

kllrnohj 6y ago

What makes you think ARM is the clear future? It's easier to do a wide decoder thanks to ARMv8's fixed-length instructions, but that's more or less the only difference that results from the ISA. And only Apple is even taking advantage of that currently - all the ARM CPU designs are 4-wide decode. Same as pretty much every modern x86 CPU. Do you think Amazon will be better at design & fabrication than Intel (when they get their shit together) or AMD? Do you expect Amazon to actually throw Apple's level of R&D money & acquisitions at the problem? And when/if they do, why would you expect any meaningful difference at the end of the day on anything? People love to talk about supposed ARM power efficiency, but the 64-core Epyc Rome at 200w is around 3w per CPU core. Which is half the power draw of an M1 firestorm core. The M1's firestorm is also faster in such a comparison, but the point is power isn't a magical ARM advantage, and x86 isn't inherently some crazy power hog. Phoronix put Graviton2 to the test vs. Epyc Rome and the results ain't pretty: https://www.phoronix.com/scan.php?page=article&item=epyc-vs-graviton2&num=8 https://www.phoronix.com/scan.php?page=article&item=epyc-vs-... "When taking the geometric mean of all these benchmarks, the EPYC 7742 without SMT enabled was about 46% faster than the Graviton2 bare metal performance. The EPYC 7742 with SMT (128 threads) increased the lead to about 51%, due to not all of the benchmarks being multi-thread focused." Graviton2 being a bit cheaper doesn't mean much when you need 50% more instances. But right now there's only a single company that makes ARM look good and that's Apple. And Apple hasn't been in the server game for a long, long time now. Everyone else's ARM CPU cores pretty much suck. Maybe Nvidia's recent acquisition will change things, who knows. But at the end of the day if AMD keeps things up, or Intel gets back on track, there really doesn't look to be a bright future for ARM outside of Apple's ecosystem and the existing ARM markets.

EricE 6y ago
I think the M1 will wake Amazon up (if they aren't already) to the incredible advantage of being able to customize a CPU to your needs. See the discussion about speed gains Apple got from accelerating common calls in their OS. You will never convince a bin part/least common denominator manufacturer like Intel or AMD to do that. Just because other ARM cores suck today doesn't mean they have to forever. Apple's don't. They took it seriously. Perhaps Amazon is too and we are just at the start of their journey. It took Apple over 10 years to get to where we are with the M1. You cited one of the significant contributors to performance - the 8-wide decode. x86 is hamstrung to 4 because of legacy. We aren't at the beginning of the story with ARM for performance, but ARM certainly isn't nearly as hamstrung out the gate by the legacy of x86 either. Heck, is there anyone making a pure 64 bit x64 chip? There's a bunch of overhead right there. Your right that ARM isn't magical - but ARM does have the potential for significantly more runway and headroom. The trade off is backwards compatibility. As most code continues to move further and further away from direct hardware interaction, is that backwards compatibility as valuable overall as it was 20 years ago? 10 years ago? I guess we will find out :)
- theevilsharpie 6y ago
  > I think the M1 will wake Amazon up (if they aren't already) to the incredible advantage of being able to customize a CPU to your needs. See the discussion about speed gains Apple got from accelerating common calls in their OS. You will never convince a bin part/least common denominator manufacturer like Intel or AMD to do that. AMD has a semi-custom division that will do exactly that for anyone capable of paying: https://www.amd.com/en/products/semi-custom-solutions https://www.amd.com/en/products/semi-custom-solutions
- kllrnohj 6y ago
  > See the discussion about speed gains Apple got from accelerating common calls in their OS. You will never convince a bin part/least common denominator manufacturer like Intel or AMD to do that. Of course you will because that happens all the time. How do you think we ended up with VT-x and friends? Intel took a use case that reached enough usage and added specialized instructions for it. This has happened a ton over the years on x86. See also AES-NI for a more application-specific addition. In addition to obviously the huge amount of SIMD experimentation. This is not fruit Apple discovered. AMD & Intel haven't been leaving this performance on the table the last 20+ years. Hell the constant addition of instructions for certain workloads is a major reason x86 is as huge & complex as it is.