diff --git a/ethernet/Makefile b/ethernet/Makefile
index 16f9a796..b8f956ba 100644
--- a/ethernet/Makefile
+++ b/ethernet/Makefile
@@ -2,7 +2,10 @@ DOS33 = ../dos33fs-utils/dos33
 TXT2BAS = ../asoft_basic-utils/tokenize_asoft
 MAKEB = ../dos33fs-utils/make_b
 
-all:	ethernet.dsk memcpy.o
+all:	ethernet.dsk
+
+memcpy:	memcpy.o
+	ld65 -o memcpy memcpy.o -t none
 
 memcpy.o:	memcpy.s
 	ca65 -o memcpy.o memcpy.s -l memcpy.lst
@@ -29,4 +32,5 @@ ethernet.dsk:	SETUP.BAS \
 	$(DOS33) -y ethernet.dsk BSAVE -a 0x4000 ./c/vmw_logo.png
 
 clean:	
-	rm -f *~ *.BAS R.TXT *.o *.lst
+	rm -f *~ *.BAS R.TXT *.o *.lst memcpy
+
diff --git a/ethernet/ethernet.dsk b/ethernet/ethernet.dsk
index e5591fe8..7a680598 100644
Binary files a/ethernet/ethernet.dsk and b/ethernet/ethernet.dsk differ
diff --git a/ethernet/memcpy.s b/ethernet/memcpy.s
index 03990e76..492f9752 100644
--- a/ethernet/memcpy.s
+++ b/ethernet/memcpy.s
@@ -1,20 +1,66 @@
 .define EQU =
 
 PTR	EQU	$06
+PTRH	EQU	$07
 
-	lda	#0
+WRAPL	EQU	$08
+WRAPH	EQU	$09
+
+SIZEL	EQU	$0A
+SIZEH	EQU	$0B
+
+tx_copy:
+
+	lda	#0		; always copying from 0x4000
 	sta	PTR
 	lda	#$40
 	sta	PTR+1
 
-	ldx	#8
+	ldx	#SIZEH		; number of 256-byte blocks
+	beq	copy_remainder	; if none, skip ahead
+
 	ldy	#0
-copy_loop:
+copy256:
 	lda	(PTR),y
-	sta	$5000
+	sta	$C0B7		; change based on uthernet slot
+
+	cmp	WRAPH,x
+	bne	nowrap256
+
+	cmp	WRAPL,y
+	bne	nowrap256
+
+	lda	#$40
+	sta	$C0B5
+	lda	#$00
+	sta	$C0B6		; wrap tx buffer address to 0x4000
+
+nowrap256:
+	iny
+	bne	copy256
+
+	inc	PTR+1		; update 16-bit pointer
+	dex			; finish a 256 byte block
+	bne	copy256
+
+	ldx	#SIZEL
+copy_remainder:
+	lda	(PTR),y
+	sta	$C0B7		; change based on uthernet slot
+
+	cmp	WRAPL,y
+	bne	nowrap_r
+
+	lda	#$40
+	sta	$C0B5
+	lda	#$00
+	sta	$C0B6		; wrap tx buffer address to 0x4000
+
+nowrap_r:
 	iny
-	bne	copy_loop
 	dex
-	bne	copy_loop
+	bne	copy_remainder
 
 	rts
+
+
diff --git a/ethernet/webserver.bas b/ethernet/webserver.bas
index 652479f3..150f6b3f 100644
--- a/ethernet/webserver.bas
+++ b/ethernet/webserver.bas
@@ -28,6 +28,15 @@
 33 POKE DP,192:POKE DP,168:POKE DP,8:POKE DP,15
 40 PRINT "UTHERNET II READY: 192.168.8.15"
 '
+' Setup Memcpy routine
+'
+50 FOR I=0 TO 72: READ X: POKE 768+I,X:NEXT I
+51 DATA 169,0,133,6,169,64,133,7,162,11,240,36,160,0,177,6
+52 DATA 141,183,192,213,9,208,15,217,8,0,208,10,169,64,141,181
+53 DATA 192,169,0,141,182,192,200,208,229,230,7,202,208,224,162,10
+54 DATA 177,6,141,183,192,217,8,0,208,10,169,64,141,181,192,169
+55 DATA 0,141,182,192,200,202,208,232,96
+'
 ' Setup Socket 0
 '
 100 REM *** Setup Socket 0
@@ -210,10 +219,11 @@
 '
 ' Check for buffer wraparound
 '
-1942 BW=0
+1942 BW=0:BO=0
 1945 IF (SI+TA>=24576) THEN BW=1:BO=24576-TA:PRINT "TX BUFFER WRAPAROUND IN ";BO
 '
 ' Write data to TX buffer
+' First write header
 '
 2000 T%=TA/256
 2005 POKE HA,T%: POKE LA,TA-(T%*256)
@@ -222,6 +232,9 @@
 2017 IF BW=0 THEN GOTO 2020
 2018 BO=BO-1: IF BO=0 THEN POKE HA,64:POKE LA,0:BW=0
 2020 NEXT I
+'
+' Write disk part
+'
 2025 FOR I=1 TO FS
 2026 C=C+1: IF C=50 THEN PRINT ".";:C=0
 2030 POKE DP,PEEK(16383+I)
@@ -230,6 +243,13 @@
 2035 NEXT I
 2040 PRINT
 '
+' The above is slow
+' Intead use our machine language routine
+'
+'2025 B%=BO/256:POKE 9,B%:POKE 8,BO-(B%*256)
+'2027 B%=FS/256:POKE 11,B%:POKE 10,FS-(B%*256)
+'2030 CALL 768
+'
 ' Update TX write ptr
 '
 2050 REM ** UPDATE TX WRITE PTR