Reply to post: Re: The worlds fastest cpu

Oracle teases 'easy-to-absorb' platform updates, wants 'all' your infrastructure biz

Mad Mike

Re: The worlds fastest cpu

@PlinkerTind

Oh dear, oh dear. I like your explanation of the T5/M5 observation I posed, but unfortunately, this link to Oracle shows you're completely wrong.

https://blogs.oracle.com/orasysat/entry/sparc_t4_t5_m5_m10

This shows they both use the S3 core. Indeed, so did the T4!! Essentially, Oracle lowered the core count in the M5 (compared to T5) and added a heap more cache. So, they're absolutely the same design and technology. They just added cache to solve problems they'd found in earlier T? servers and needed to reduce the core count to get the die space.

To count as generations, you really have to look at core designs and Oracle has produced nowhere near that many generations of core designs. They've used similar methods as IBM (and others) to produce new CPUs without changing the basic cores. Drop the fabrication size and increase speed for instance. Upgrade from PCIe2 to PCIe3 on the chip. However, the same core design is present.

In terms of performance benchmarking, let's consider that. Benchmarks are a highly convoluted and bad way of trying to demonstrate actual CPU performance. The configurations are contrived and the software is normally specifically configured and optimised to the exact layout of the server and CPU. Factors outside the CPU are often used to provide what looks like enhanced performance (e.g. using flash based I/O on TPC-C). Also, nobody benchmarks against virtualised infrastructure. I'm not going to deny that doing the above makes the Oracle chips look fantastic. However, in the REAL world, where virtualisation and mixed workloads are the norm, these advantages drop away. For instance, Oracle claims you can have a LDOM per thread, but you'd have to be insane to do it. I've used T3 servers which ran like a dog under virtualisation as any search of the web will show. So, benchmarks are fine to a point, but need to be taken with a very large pinch of salt.

You've also completely missed my point about encryption and compression. Oracle (in general) writes their software (such as Oracle DB) to specifically prevent it using competitors encryption and compression accelerators. So, Power has for years now had encryption acceleration within the processor. Whether you turn this on or off (hardware level), Oracle DB will simply not use it. They've made a decision on the coding of their software to avoid using a perfectly acceptable and suitable acceleration tool. Why? Because they want their hardware to look better. This is why the benchmarks show such a difference. If Oracle software actually used the available features on competitors hardware, who knows what the story would be. Using hardware accelerators often requires more than just a hardware turn on or off. Just think of the various SSE enhancements to x86 architecture. You can have them turned on or off, but if you never use those instructions, you won't get any benefit regardless. So, my statement is absolutely correct. Oracle actually prevents use of the Power 8 on-chip encryption accelerator whether it is turned on or off at the hardware level.

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Biting the hand that feeds IT © 1998–2022