1) Pentium 4s and Celerons with L2 cache size of 256K or less choose a length 4M FFT when they could use an FFT between 1600K and 4M in size. Fixed in next release.
2) The on screen message reporting that errors have occurred during the LL test was wrong. The counts for "SUM(INPUTS) != SUM(OUTPUTS)" and "ROUNDOFF > 0.4" were reversed. Fixed in next release.
3) Dual-boot users running a 32-bit and 64-bit executable on the same exponent may experience "Unable to initialize FFT" message. This happens when prime95 writes an FFT size to worktodo.txt and that FFT size is only supported by just one of the two executables. For example, the 2240K FFT length is supported for 64-bit Core 2 but not 32-bit Core 2. Fixed, somewhat inelegantly, in 26.5 -- unimplemented FFT lengths are ignored when worktodo.txt is read.
4) Prime95 will sometimes inexplicably unreserve exponents. I think this happens when prime95 incorrectly calculates the CPU speed. Version 26.5 will only use a new slower CPU speed measurement after several slower CPU measurements. Hopefully, this will resolve the unreserve problem caused by a single erroneous CPU speed measurement.
5) Prime95 would lose the how_far_factored and tests_saved information on PRP= lines in worktodo.txt. Fixed in 26.5.
5) Prime95 did not offer an option to delete P-1 save files when a work unit completes. Fixed in 26.5.
6) Prime95 did not fully understand Sandy Bridge CPUID output. Fixed in 26.5.
7) Prime95 did not accurately report the CPU speed. Fixed in 26.5.
8) At startup, the workers threads do not start up properly until communication with the server completes. If you have a large worktodo.txt file, this will result in wasted CPU time and many "use count" error messages. Fixed in 26.5.
9) Time estimates for trial factoring to 2^79 and above were incorrect. Fixed in 26.5.

-For rare cases where the program cannot figure out the number of cores and hyperthreading,
the NumPhysicalCores option may help.

1) Occasionally, P-1 stage 2 would report 100% complete before stage 2 completed. This bug has been around forever. I've finally found the cause! Fixed in 26.5 build 2.
2) Worktodo.txt was not updated properly when P-1 completed for a PRP test. Fixed in 26.5 build 3.
3) In 26.5 build 3, when benchmarking and measuring CPU speed the program will keep all cores busy by launching auxiliary threads that loop indefinitely. This should prevent Intel Turbo boost from kicking in. This should prevent cases where the CPU speed is reported as 1.87 GHz, yet a benchmark reports very fast timings because the CPU core has been boosted to 3.2 GHz because only one core is active.
4) In 26.5 build 3, the OS's mapping of hyperthreaded logical CPU numbers to physical CPUs is determined automatically at startup. The AffinityScramble setting (see undoc.txt) is no longer supported. It is replaced by the AffinityScramble2 setting.
5) In 26.5 build 3, the message "set affinity to run on any cpu" instead listed all the CPU numbers and as well as CPU #31. Fixed in build 4.
6) In 26.5 build 3 and 4, benchmarking a dual-core non-SSE2 machine will crash. Fixed in build 5.
7) KeepPminus1SaveFiles=0 option described in undoc.txt did not work when a factor was found. Fixed in 26.6.
8) The 32K,64K,80K length FFTs with sumout checking for Pentium 4's with 256K or less cache were not implemented. Symptom is a torture test failure. Fixed in 26.6.
9) Prime95 does not load on Sandy Bridge CPUs running Win XP. I've got a trial fix in 26.6.
10) Prime95 chooses very inefficient FFT implementations for Core 2 Celerons (they have 512K L2 cache). Working on a fix for for 26.6.
11) Logical CPU numbers were sometimes output zero-based and sometimes one-based. In 26.6, they are all one-based.

This version supports 64-bit optimized AVX FFTs. 32-bit AVX FFTs are also a little bit faster. I haven't done full benchmarks so I'm not sure how much faster it is than versions 27.2 or 26.6.

The good/bad news is these FFTs are so fast that they are limited by memory bandwidth -- standard Sandy Bridge CPUs will experience a slow down when running all 4 cores. I'd like to hear from Sandy Bridge-E users to see if they also suffer slow downs when all 4 cores are running.

This version fixes a number of bugs in 27.3 - mostly related to LLR and PFGW issues. There are no speed improvements over 27.3. Hopefully, Bulldozer processors now properly run SSE2 FFTs.