AMD's 128-core Epycs could spell trouble for Ampere Computing With the unveiling of its 128-core Epycs, codenamed Bergamo, AMD has put forward a challenge to Ampere Computing's tentative footing in the cloud and hyperscale arena. Despite the impending threat of another cloud-native chip, Ampere Computing chief product officer Jeff Wittich isn't too concerned. "I remain confident that we …
The power consumption advantages of Ampere are not insignificant.
Not anymore.
Porting software isn't generally as bad between the two architectures as it used to be either. Ultimately it (should) come down to a TCO decision that looks like it'll favour ARM servers. (Maybe).
Porting and supporting multiple architectures is still a pain-point. Intel and AMD don't actually need to outperform Ampere, they just need to come close enough that it isn't worthwhile for most to put in that effort.
Such purity in marketing, it's almost unbelievable. Such beauty can be found everywhere.
What's happening in datacenter? Cloud-native, scale-out workloads.
What's the future of computing? Cloud-native, scale-out workloads.
What's your shirt colort? Cloud-native, scale-out workloads.
I gathered the below info and in terms of cores and GHz, if they are both under steady state maximum load they are comparable.
AmpereOne 3GHz*192 cores ~= 576
AMD EPYC 9754 2.25GHz*256 threads ~=576
(turbo only works when the CPU is initially powered on or mostly idle. In a DC that will really only happen at initial power up, or if the load is migrated to more idle servers, which might give the AMD chips a 37% faster boot uptime time, could be enough of an advantage to be considered useful.)
For power usage the 192 core AmpereOne has a miniscule advantage of 350W vs 360W for the 128 core/256 thread AMD EPYC 9754 "Bergamo".
Where the AMD EPYC 9754 will probably win is for cache, it has roughly 50% more and that will make a big difference to overall performance. But that does depend on the size of the L1 cache and I can't find any numbers for that. If AMD have a tiny L1 cache, they could end up being extremely comparable.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
AMD EPYC 9754 "Bergamo", Zen 4c CPUs come with up to 128 cores spread over eight core complex die (or CCD) chiplets with 16 cores each.
Base Clock 2.25GHz
Max. Boost Clock up to 3.1GHz
Cores/Threads 128/256
L2 Cache 128MB (~1MB cache per thread)
L3 Cache 256MB (~2MB cache per thread)
Memory 12 Channels DDR5 System Memory Specification Up to 4800MHz (Per Socket Mem BW 460.8 GB/s)
Connectivity 128 Lanes PCIe Gen5
TDP 360W
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
AmpereOne (136 Cores | 144 cores | 160 cores | 176 cores | 192 cores)
Consistent Freq up to 3.0 GHz
L1: 16 KB Instruction and 64 KB Data per Core
Cores/Threads 192/192
L2 Cache 384MB (2MB of private L2 cache per core)
Memory 8 Channel DDR5 with ECC Up to 16 DIMMs (2DPC) and 8 TB total memory
Connectivity 128 Lanes PCIe Gen5 (optional 64 lane CCIX for multi-socket support) 32 Controllers Bifurcation to x4
200-350W per socket (~1.8 watts per core peak)
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
SMT doesn’t really work like that. It’s 128x clock rate, the second thread per core just helps to keep it busy. However the IPC is also not comparable between totally different architectures, so it isn’t useful to compare based on simple multiplication.
Overall performance boost from SMT varies tremendously on workload, but it would be unusual to exceed 30%, and on a pure integer benchmark is probably well approximated by 0.
Also you absolutely can run a whole chip in a turbo mode for an extended period, if the power and cooling can support it. Max all-core frequency is likely to be significantly below max single-core, but it is almost certainly above the base clock.
The way most turbo works is that it boosts performance until thermal overload, then throttles performance until the temperature inside the chip has cooled enough to avoid permanent damage. As chips/chiplets feature sets get smaller and smaller the watts per square millimetre is going up. There is an exponential relationship between clock frequency and power used (almost all power ends up as heat/thermal energy that needs to be removed at least as fast as it is being generated). The real problem is that we do not have any "cheap" and "extremely efficient in terms of energy required" cooling solutions that can cool the watts per square millimetre inside a chip that are close to the inside a rocket engines exhaust. So the solution is to under clock them and for marketing to push the turbo speed hard to shift more product fast.
And I do agree that my approximation is extremely crude, but without access to physical hardware for benchmarks, it is a first step approximation. There are so many factors I'm ignoring just to get a ball park figure, it is not good, but it is better than nothing.
So Intel are doing so badly at the moment their processors aren't considered worth comparing against by either of these two companies.
Much as I never liked their business practices, and the way they held computing back because they had no real competition, things like this really do need to be a 3 way fight; but it looks like another 2 to 3 years before Intel will be back on solid competitive ground and it also looks like they will be the little guy in the fight.
But where Ampere once filled a hole in the market for workloads that prioritized core density over all else, the fledgling chipmaker now has Intel, AMD, and the full momentum of the x86 architecture to contend with.
I'll be honest, I didn't realised we were that far along the transition.
A few years ago this story would've been 'Upstart Ampere set to challenge x86', but already the framing is shifting to the two parties being 'equal'.
I remember this shift from 20-25 years ago, it started with comments like 'x86 is a toy, no one would use it in serious servers', then it ended up with more comments like this about how 'x86 was making inroads but the traditional big iron was fighting back'....
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
