I had problem for boot gPXE with my testing computer. It keeps reboot(crash) right after gPXE relocats itself to high memory location.

I've traced partial codes of src/arch/i386/prefix/*prefix.S And dumped the e820 memory maps as stefanha suggestes. The addresses all looks normal, no overlapps, and the gPXE loading address seems reasonable.

NVIDIA Boot Agent 215.0503
Copyright (C) 2001-2005 NVIDIA Corporation
Copyright (C) 1997-2000 Intel Corporation

CLIENT MAC ADDR: 00 15 F2 3A BE 48  GUID: 00501D51-548A-0D10-A3D8-AB7EEAB94556
CLIENT IP: 192.168.201.208  MASK: 255.255.255.0  DHCP IP: 192.168.201.186
GATEWAY IP: 192.168.201.186
PXE->EB: !PXE at 99AB:0080, entry point at 99AB:0E57
         UNDI code segment 99AB:5BD3, data segment 98B3:0F80 (610-638kB)
         UNDI device is PCI 00:0A.0, type DIX+802.3
         610kB free base memory after PXE unload
Fetching system memory map
FBMS base memory size 596 kB [0,95000)
INT 15,e801 extended memory size 15360+64*32511=2096064 kB [100000,7fff0000)
INT 15,e820 region [0,9f800) type 1
INT 15,e820 region [f0000,100000) type 2
INT 15,e820 region [fec00000,100000000) type 2
INT 15,e820 region [e0000000,f0000000) type 2
INT 15,e820 region [7fff3000,80000000) type 3
INT 15,e820 region [7fff0000,7fff3000) type 4
INT 15,e820 region [9f800,a0000) type 2
INT 15,e820 region [100000,7fff0000) type 1
Relocate: currently at [400000,435c98)
...need 35ca7 bytes for 16-byte alignment
Considering [0,9f800)
...usable portion is [0,9f800)
Considering [100000,7fff0000)
...end truncated to 7ff00000 (avoid ending in odd megabyte)
...usable portion is [100000,7ff00000)
...new best block found.
Relocating from [400000,435c98) to [7feca360,7feffff8)

After about a week's struggle, I removed all DIMMs and add-on cards, and cleaned all pin connectors. Fount that it boots fine without any software patch.