Choosing between 3000+, 3200+, and 3500+ winchester and venice CPU’s
I have seen a lot of questions posted about which CPU to purchase – the 3000+, 3200+, or 3500+ winchester (and venice). Most responses are of the following nature:
“Get the higher speed CPU (if the price difference isn’t substantial) due to the higher multiplier.”
While this may be true in some cases, it isn’t always true. The important thing is to purchase the CPU that works well with the rest of your hardware – most importantly, the memory and the motherboard.
Before I launch into an example, I would like to point out a couple of very important facts. These have all been mentioned before but they are fundamental elements that must be understood and it can never hurt to quickly review them.
The Athlon 64 chips have their memory controller located on the chip die. There is a separate bus connecting the processor and the memory called the memory bus. It is COMPLETELY independent from the hyper transport bus (the bus used to transfer information from the rest of the systems components to the CPU). The results of these architectural changes are quite dramatic:
First, “access time for memory read/write is reduced compared to having the memory controller on the chipset. With a separate memory bus and system bus, there is less bus conflict and the effective memory latency between the CPU (after L2 miss) and the memory (L3) is reduced.” -hitechjb
Second is a very important point mentioned by Guantum in his “A64 101” sticky:
“Traditionally, the memory speed is derived off of the front side bus, and can be manipulated by FSB/memory ratios. In contrast, in the A64, memory speed is derived off of the CPU speed in CPU/memory ratios. This is why it’s rather inaccurate to say that the memory is ever running “synchronously.” The memory is always running asynchronously with respect to the CPU speed, off of which it’s derived. How fast it’s running with respect to the HTT does not matter at all. There is no latency hit in running the memory slower than the HTT.”
REMEMBER: There is no latency hit in running the memory slower than the HTT.
-----------------------------------------------------------------------------------------------------
Now for an example:
The motherboard of choice for a S939 system is currently the DFI NForce4 Ultra-D. Why? – because it has a wealth of BIOS options (numerous DRAM:FSB ratios, integrated memtest86, 4V vdimm option to name a few), it can allows for high HTT speeds, it looks cool and its relatively cheap. Thus, I will use the DFI NF4 motherboard in my example.
The DFI NF4 has the following memory dividers available:
100:200 (DRAM:FSB 1:2)
120:200 (DRAM:FSB 3:5)
133:200 (DRAM:FSB 2:3)
140:200 (DRAM:FSB 7:10)
150:200 (DRAM:FSB 3:4)
166:200 (DRAM:FSB 5:6)
180:200 (DRAM:FSB 9:10)
200:200 (DRAM:FSB 1:1)
It is important to understand how memory speed is calculated in an Athlon 64 based system. The following formulas describe the relationship between memory bus frequency, CPU frequency and HTT.
CPU_frequency = HTT x CPU_multiplier
memory_bus_frequency = CPU_frequency / CPU_memory_divider
or
memory_bus_frequency = CPU_multiplier x HTT / CPU_memory_divider
where
CPU_memory_divider = ceiling(CPU_multiplier / memory_HTT_ratio)
-taken from Hitechjb1’s “A64 CPUs, chipsets, motherboards” post.
The Winchester and Venice chips have their multipliers top locked. The following is a list of available multipliers for the 3000+, 3200+ and 3500+ CPU’s.
CPU ------------- Multipliers
3000+ ----------- 9,8,7,6,5…..
3200+ ----------- 10,9,8,7,6,5…..
3500+ ----------- 11, 10,9,8,7,6,5…..
It is useful to determine the CPU_memory_divider at each available multiplier and motherboard setting. Thus, we will systematically calculate all available CPU_memory_dividers available in the DFI NF4 motherboard and 3000+, 3200+, and 3500+ CPU combinations.
DFI NF4 w/ 3000+ CPU
Start with the 9x multi:
With a 1:1 ratio (DRAM:FSB) we have:
CPU_memory_divider = ceiling(CPU_multiplier / memory_HTT_ratio)
CPU_memory divider = ceiling(9 / (1/1))
CPU_memory_divider = 9
With a 9:10 ratio (DRAM:FSB) we have:
CPU_memory_divider = ceiling(CPU_multiplier / memory_HTT_ratio)
CPU_memory divider = ceiling(9 / (9/10)) = 10
CPU_memory_divider = 10
With a 5:6 ratio (DRAM:FSB) we have:
CPU_memory_divider = ceiling(CPU_multiplier / memory_HTT_ratio)
CPU_memory divider = ceiling(9 / (5/6)) = 10.8
CPU_memory_divider = 11
……. Continuing this process for the 9x multi and the rest of the available memory dividers yields the following CPU_memory_dividers:
DRAM:FSB ----------------- CPU_memory_divider
1:1 --------------------------------- 9
9:10 ------------------------------- 10
5:6 -------------------------------- 11
3:4 -------------------------------- 12
7:10 ------------------------------- 13
2:3 -------------------------------- 14
3:5 -------------------------------- 15
1:2 -------------------------------- 18
Now if we continue to do the same things w/ the 8x and 7x multipliers (I wont go lower than this because very few people will be using a multiplier lower than 7x. Most people will end up using 8x or 9x) we get the following results:
With the 8x multi
DRAM:FSB ----------------- CPU_memory_divider
1:1 --------------------------------- 8
9:10 -------------------------------- 9
5:6 -------------------------------- 10
3:4 -------------------------------- 11
7:10 ------------------------------- 12
2:3 -------------------------------- 12
3:5 -------------------------------- 14
1:2 -------------------------------- 16
With the 7x multi
DRAM:FSB ----------------- CPU_memory_divider
1:1 --------------------------------- 7
9:10 -------------------------------- 8
5:6 --------------------------------- 9
3:4 -------------------------------- 10
7:10 ------------------------------- 10
2:3 -------------------------------- 11
3:5 -------------------------------- 12
1:2 -------------------------------- 14
To sum up, the 3000+ CPU has the following CPU_memory_dividers available:
7, 8, 9, 10, 11, 12, 13, 14, 15, 16 and 18
We can use the same process to find the available CPU_memory_dividers for the 3200+ and 3500+ CPU’s. The following tables are a summary of the results.
CPU -------- Available CPU_memory_dividers
3000+ ------ 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 18
3200+ ------ 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 18, 17, 20
3500+ ------ 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 18, 17, 20, 19, 22
BOLD = additional dividers resulting from the 10x multi (3200+) and the 11x multi (3500+)
What does this mean?
Before I come to any conclusions I would like to remind you of the following:
1. There is no latency hit in running the memory slower than the HTT. (asynchronously)
2. DFI NF4 motherboards allow for very high HTT speeds
3. DFI NF4 motherboards have MANY DRAM:FSB ratios to allow for a wealth of CPU/Memory dividers (as listed above).
With these facts in mind, we can see that there is no real advantage to having the 3200+ or 3500+ over the 3000+. The 3200+ gives you the additional 17 and 20 CPU_memory_dividers while the 3500+ gives the additional 17, 20, 19, and 22 dividers. These dividers are utterly useless as nobody will be running their ram below 200mhz. (Eg… if you manage to get your CPU to 3000mhz, using the 17, 19, 20, and 22 dividers would make the memory run at 176mhz, 158mhz, 150mhz, 136mhz respectively).
The only reason you would want to purchase a 3200+ or 3500+ over a 3000+ is if the CPU’s were speed binned. However, Winchester overclocking results show that the 3000+ CPU holds its own against both the 3200+ and the 3500+. It is generally accepted that all three Winchester core based CPUs (3000+, 3200+, 3500+) top out at around the same speed. (Note that some steppings are better than others)
So what exactly is wrong with the “get the higher speed CPU because of the higher multiplier” comments?
Well, if you are running a motherboard that is capable of higher HTT speeds and has the right CPU / memory dividers available, then the higher multipliers don’t give you an advantage.
For example, say you have the choice between the 3000+ and the 3200+ Winchester and that you already own the DFI NF4 Ultra-D (or SLI-DR). Both CPU’s are multiplier top locked at 9x and 10x respectively. For explanatory purposes, say both chips max out at 2700mhz and that we have OCZ VX RAM that maxes out at 265mhz (2-2-2-x).
Using the 3000+
CPU Max = 2700mhz
OCZ VX Max = 265mhz
Ideal CPU/memory ratio = 2700mhz / 265mhz = 10.18
If we use the /10 CPU_memory_divider we get:
CPU = 2700mhz
OCZ VX = 270mhz (too high for our memory)
Thus, we drop the CPU speed until the memory is running at its max.
CPU = 2650mhz
OCZ VX = 265mhz
(Note: The /11 divider is available but using it would result a memory speed of 2700/11 = 245mhz. Thus you have to choose between CPU, MEM of 2650,265 or 2700,245) The 20mhz memory speed increase will probably be better than a 50mhz CPU speed increase).
Note that the /10 CPU_memory_divider is available by using the 9x multiplier and the 9:10 (DRAM:FSB) ratio. The /11 CPU_memory_divider is available by using the 9x multiplier and the 5:6 (DRAM:FSB) ratio. Also note that the HTT is running at 294mhz… which is no problem for the DFI NF4.
Using the 3200+ (everything is essentially the same)
CPU Max = 2700mhz
OCZ VX Max = 265mhz
Ideal CPU/memory ratio = 2700mhz / 265mhz = 10.18
Using the /10 CPU_memory_divider we get:
CPU = 2700mhz
OCZ VX = 270mhz (too high for our memory)
Dropping the CPU speed we get
CPU = 2650mhz
OCZ VX = 265mhz
So what’s different between the 3000+ and the 3200+?
Well, the 3200+ CPU has the 10x multiplier available. Thus, we don’t have to use a ratio to attain a memory speed of 265mhz (we can run the memory 1:1 w/ the HTT). BUT – as mentioned before, There is no latency hit in running the memory slower than the HTT (asynchronously or with a DRAM:FSB ratio). There is no evidence (that I know of) to suggest that running the memory asynchronously w/ the HTT results in a performance hit. If there is a performance hit it probably wouldn’t be noticeable.
A note on HTT speeds
Using a higher multiplier will result in a lower HTT speed. This can be beneficial for people with motherboards that don’t like high HTT speeds. For example. Say we have 3 CPU’s – a 3000+, a 3200+, and a 3500+ Venice or Winchester. All max out at 3000mhz (ya… their all amazing CPU’s )
the 3000+ would be running at 333x9
the 3200+ would be running at 300x10
the 3500+ would be running at 273x11
The DFI NF4 should have no problems running with HTT at 333mhz but other boards may have some trouble. (MSI K8N Neo 2 939 board for example).
Conclusion
Which CPU you buy should be dependant on the rest of your hardware – primarily your motherboard. DO YOUR HOMEWORK!! I can’t stress this enough. Find out which motherboards allow for high HTT speeds. Research the various bios options on certain motherboards and find out what DRAM:FSB ratios are available. Get some feedback from people who own the equipment you are looking at purchasing. The forums are an excellent resource for ALL of this information. Once you have this information, choose a CPU accordingly.
The DFI NF4 is an amazing piece of technology. If you own this motherboard and are purchasing a Winchester based CPU then your best bet is to choose the 3000+ version. If you still don’t know why, go back and reread this post. If you are purchasing a Venice in the near future the same logic applies. It is important to note, however, that we do not know if the Venice CPU’s will be speed binned. If so, then a 3500+ or 3800+ will most definitely OC higher than the 3000+ (on average). I have a hunch that this is not the case…… but to be on the safe side I chose something in the middle - the 3200+. Can’t back out now…. already wired the money to NCIX. I guess we will have to wait and see.
-Adrayic
I have seen a lot of questions posted about which CPU to purchase – the 3000+, 3200+, or 3500+ winchester (and venice). Most responses are of the following nature:
“Get the higher speed CPU (if the price difference isn’t substantial) due to the higher multiplier.”
While this may be true in some cases, it isn’t always true. The important thing is to purchase the CPU that works well with the rest of your hardware – most importantly, the memory and the motherboard.
Before I launch into an example, I would like to point out a couple of very important facts. These have all been mentioned before but they are fundamental elements that must be understood and it can never hurt to quickly review them.
The Athlon 64 chips have their memory controller located on the chip die. There is a separate bus connecting the processor and the memory called the memory bus. It is COMPLETELY independent from the hyper transport bus (the bus used to transfer information from the rest of the systems components to the CPU). The results of these architectural changes are quite dramatic:
First, “access time for memory read/write is reduced compared to having the memory controller on the chipset. With a separate memory bus and system bus, there is less bus conflict and the effective memory latency between the CPU (after L2 miss) and the memory (L3) is reduced.” -hitechjb
Second is a very important point mentioned by Guantum in his “A64 101” sticky:
“Traditionally, the memory speed is derived off of the front side bus, and can be manipulated by FSB/memory ratios. In contrast, in the A64, memory speed is derived off of the CPU speed in CPU/memory ratios. This is why it’s rather inaccurate to say that the memory is ever running “synchronously.” The memory is always running asynchronously with respect to the CPU speed, off of which it’s derived. How fast it’s running with respect to the HTT does not matter at all. There is no latency hit in running the memory slower than the HTT.”
REMEMBER: There is no latency hit in running the memory slower than the HTT.
-----------------------------------------------------------------------------------------------------
Now for an example:
The motherboard of choice for a S939 system is currently the DFI NForce4 Ultra-D. Why? – because it has a wealth of BIOS options (numerous DRAM:FSB ratios, integrated memtest86, 4V vdimm option to name a few), it can allows for high HTT speeds, it looks cool and its relatively cheap. Thus, I will use the DFI NF4 motherboard in my example.
The DFI NF4 has the following memory dividers available:
100:200 (DRAM:FSB 1:2)
120:200 (DRAM:FSB 3:5)
133:200 (DRAM:FSB 2:3)
140:200 (DRAM:FSB 7:10)
150:200 (DRAM:FSB 3:4)
166:200 (DRAM:FSB 5:6)
180:200 (DRAM:FSB 9:10)
200:200 (DRAM:FSB 1:1)
It is important to understand how memory speed is calculated in an Athlon 64 based system. The following formulas describe the relationship between memory bus frequency, CPU frequency and HTT.
CPU_frequency = HTT x CPU_multiplier
memory_bus_frequency = CPU_frequency / CPU_memory_divider
or
memory_bus_frequency = CPU_multiplier x HTT / CPU_memory_divider
where
CPU_memory_divider = ceiling(CPU_multiplier / memory_HTT_ratio)
-taken from Hitechjb1’s “A64 CPUs, chipsets, motherboards” post.
The Winchester and Venice chips have their multipliers top locked. The following is a list of available multipliers for the 3000+, 3200+ and 3500+ CPU’s.
CPU ------------- Multipliers
3000+ ----------- 9,8,7,6,5…..
3200+ ----------- 10,9,8,7,6,5…..
3500+ ----------- 11, 10,9,8,7,6,5…..
It is useful to determine the CPU_memory_divider at each available multiplier and motherboard setting. Thus, we will systematically calculate all available CPU_memory_dividers available in the DFI NF4 motherboard and 3000+, 3200+, and 3500+ CPU combinations.
DFI NF4 w/ 3000+ CPU
Start with the 9x multi:
With a 1:1 ratio (DRAM:FSB) we have:
CPU_memory_divider = ceiling(CPU_multiplier / memory_HTT_ratio)
CPU_memory divider = ceiling(9 / (1/1))
CPU_memory_divider = 9
With a 9:10 ratio (DRAM:FSB) we have:
CPU_memory_divider = ceiling(CPU_multiplier / memory_HTT_ratio)
CPU_memory divider = ceiling(9 / (9/10)) = 10
CPU_memory_divider = 10
With a 5:6 ratio (DRAM:FSB) we have:
CPU_memory_divider = ceiling(CPU_multiplier / memory_HTT_ratio)
CPU_memory divider = ceiling(9 / (5/6)) = 10.8
CPU_memory_divider = 11
……. Continuing this process for the 9x multi and the rest of the available memory dividers yields the following CPU_memory_dividers:
DRAM:FSB ----------------- CPU_memory_divider
1:1 --------------------------------- 9
9:10 ------------------------------- 10
5:6 -------------------------------- 11
3:4 -------------------------------- 12
7:10 ------------------------------- 13
2:3 -------------------------------- 14
3:5 -------------------------------- 15
1:2 -------------------------------- 18
Now if we continue to do the same things w/ the 8x and 7x multipliers (I wont go lower than this because very few people will be using a multiplier lower than 7x. Most people will end up using 8x or 9x) we get the following results:
With the 8x multi
DRAM:FSB ----------------- CPU_memory_divider
1:1 --------------------------------- 8
9:10 -------------------------------- 9
5:6 -------------------------------- 10
3:4 -------------------------------- 11
7:10 ------------------------------- 12
2:3 -------------------------------- 12
3:5 -------------------------------- 14
1:2 -------------------------------- 16
With the 7x multi
DRAM:FSB ----------------- CPU_memory_divider
1:1 --------------------------------- 7
9:10 -------------------------------- 8
5:6 --------------------------------- 9
3:4 -------------------------------- 10
7:10 ------------------------------- 10
2:3 -------------------------------- 11
3:5 -------------------------------- 12
1:2 -------------------------------- 14
To sum up, the 3000+ CPU has the following CPU_memory_dividers available:
7, 8, 9, 10, 11, 12, 13, 14, 15, 16 and 18
We can use the same process to find the available CPU_memory_dividers for the 3200+ and 3500+ CPU’s. The following tables are a summary of the results.
CPU -------- Available CPU_memory_dividers
3000+ ------ 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 18
3200+ ------ 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 18, 17, 20
3500+ ------ 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 18, 17, 20, 19, 22
BOLD = additional dividers resulting from the 10x multi (3200+) and the 11x multi (3500+)
What does this mean?
Before I come to any conclusions I would like to remind you of the following:
1. There is no latency hit in running the memory slower than the HTT. (asynchronously)
2. DFI NF4 motherboards allow for very high HTT speeds
3. DFI NF4 motherboards have MANY DRAM:FSB ratios to allow for a wealth of CPU/Memory dividers (as listed above).
With these facts in mind, we can see that there is no real advantage to having the 3200+ or 3500+ over the 3000+. The 3200+ gives you the additional 17 and 20 CPU_memory_dividers while the 3500+ gives the additional 17, 20, 19, and 22 dividers. These dividers are utterly useless as nobody will be running their ram below 200mhz. (Eg… if you manage to get your CPU to 3000mhz, using the 17, 19, 20, and 22 dividers would make the memory run at 176mhz, 158mhz, 150mhz, 136mhz respectively).
The only reason you would want to purchase a 3200+ or 3500+ over a 3000+ is if the CPU’s were speed binned. However, Winchester overclocking results show that the 3000+ CPU holds its own against both the 3200+ and the 3500+. It is generally accepted that all three Winchester core based CPUs (3000+, 3200+, 3500+) top out at around the same speed. (Note that some steppings are better than others)
So what exactly is wrong with the “get the higher speed CPU because of the higher multiplier” comments?
Well, if you are running a motherboard that is capable of higher HTT speeds and has the right CPU / memory dividers available, then the higher multipliers don’t give you an advantage.
For example, say you have the choice between the 3000+ and the 3200+ Winchester and that you already own the DFI NF4 Ultra-D (or SLI-DR). Both CPU’s are multiplier top locked at 9x and 10x respectively. For explanatory purposes, say both chips max out at 2700mhz and that we have OCZ VX RAM that maxes out at 265mhz (2-2-2-x).
Using the 3000+
CPU Max = 2700mhz
OCZ VX Max = 265mhz
Ideal CPU/memory ratio = 2700mhz / 265mhz = 10.18
If we use the /10 CPU_memory_divider we get:
CPU = 2700mhz
OCZ VX = 270mhz (too high for our memory)
Thus, we drop the CPU speed until the memory is running at its max.
CPU = 2650mhz
OCZ VX = 265mhz
(Note: The /11 divider is available but using it would result a memory speed of 2700/11 = 245mhz. Thus you have to choose between CPU, MEM of 2650,265 or 2700,245) The 20mhz memory speed increase will probably be better than a 50mhz CPU speed increase).
Note that the /10 CPU_memory_divider is available by using the 9x multiplier and the 9:10 (DRAM:FSB) ratio. The /11 CPU_memory_divider is available by using the 9x multiplier and the 5:6 (DRAM:FSB) ratio. Also note that the HTT is running at 294mhz… which is no problem for the DFI NF4.
Using the 3200+ (everything is essentially the same)
CPU Max = 2700mhz
OCZ VX Max = 265mhz
Ideal CPU/memory ratio = 2700mhz / 265mhz = 10.18
Using the /10 CPU_memory_divider we get:
CPU = 2700mhz
OCZ VX = 270mhz (too high for our memory)
Dropping the CPU speed we get
CPU = 2650mhz
OCZ VX = 265mhz
So what’s different between the 3000+ and the 3200+?
Well, the 3200+ CPU has the 10x multiplier available. Thus, we don’t have to use a ratio to attain a memory speed of 265mhz (we can run the memory 1:1 w/ the HTT). BUT – as mentioned before, There is no latency hit in running the memory slower than the HTT (asynchronously or with a DRAM:FSB ratio). There is no evidence (that I know of) to suggest that running the memory asynchronously w/ the HTT results in a performance hit. If there is a performance hit it probably wouldn’t be noticeable.
A note on HTT speeds
Using a higher multiplier will result in a lower HTT speed. This can be beneficial for people with motherboards that don’t like high HTT speeds. For example. Say we have 3 CPU’s – a 3000+, a 3200+, and a 3500+ Venice or Winchester. All max out at 3000mhz (ya… their all amazing CPU’s )
the 3000+ would be running at 333x9
the 3200+ would be running at 300x10
the 3500+ would be running at 273x11
The DFI NF4 should have no problems running with HTT at 333mhz but other boards may have some trouble. (MSI K8N Neo 2 939 board for example).
Conclusion
Which CPU you buy should be dependant on the rest of your hardware – primarily your motherboard. DO YOUR HOMEWORK!! I can’t stress this enough. Find out which motherboards allow for high HTT speeds. Research the various bios options on certain motherboards and find out what DRAM:FSB ratios are available. Get some feedback from people who own the equipment you are looking at purchasing. The forums are an excellent resource for ALL of this information. Once you have this information, choose a CPU accordingly.
The DFI NF4 is an amazing piece of technology. If you own this motherboard and are purchasing a Winchester based CPU then your best bet is to choose the 3000+ version. If you still don’t know why, go back and reread this post. If you are purchasing a Venice in the near future the same logic applies. It is important to note, however, that we do not know if the Venice CPU’s will be speed binned. If so, then a 3500+ or 3800+ will most definitely OC higher than the 3000+ (on average). I have a hunch that this is not the case…… but to be on the safe side I chose something in the middle - the 3200+. Can’t back out now…. already wired the money to NCIX. I guess we will have to wait and see.
-Adrayic
Last edited: