TR Forums

Hey gerbils. I was looking at AMD's FX model comparison page and some things caught my eye. I think AMD tends to get sloppy with their tables.

1. I was kinda surprised to see that the FX-4130 and FX-4300 have only 4MB of L3 cache. I used to think 8MB is standard on all their FX chips. A quick look at CPU-World also indicates 4MB of L3 for the FX-4300. Ok, perhaps they're having higher L3 defect rates than expected on those quads.

2. The trickiest part where AMD always seems to have trouble is cache sizes, since they, in the past, seem to get things mixed up a little bit. For example, I remember seeing tables on their site indicating L2 cache for the Phenom II to be 512KB, but it didn't say whether it was just for one core or all the cores. They've fixed it now, though. But look at the FX table, specifically, 'L1 Cache (Instruction + Data) per core'. It says each core has '128KB (64KB + 64KB)'. That's a bit off, isn't it? There's 64KB of L1I per module, shared by two cores, and each core gets its own 16KB of L1D. Perhaps you can split the 64KB into two and say each core gets 32KB L1I and 16KB L1D, for a total of 48KB. Not clean cut, but at least you have a better sense of what's in there. However, it's not 64KB + 64KB like the table says.

3. It says HyperTransport runs at 4GT/sec., or 16GB/sec. A quick check with CPU-Z, however, tells me my FX-8350's HT runs at 5.2GT/sec, not 4.0GT/sec. Not sure if AMD is being conservative here or if my MSI 990FXA-GD65 is pushing my HT beyond spec, but if some of their models run HT at 4.0GT, they should've at least indicated it the way they indicated that different models have different L3 cache sizes.

I don't know about you guys, but AMD just always seems to be kinda sloppy with their tables and their whole website in general. Like I've said before, they need to clean up their website. Who's in charge of it, anyway?

ronch wrote:

1. I was kinda surprised to see that the FX-4130 and FX-4300 have only 4MB of L3 cache. I used to think 8MB is standard on all their FX chips. A quick look at CPU-World also indicates 4MB of L3 for the FX-4300. Ok, perhaps they're having higher L3 defect rates than expected on those quads.

They're lower-end processors; obviously they'll have less cache. This has been standard practice since L2 and L3 cache were included on CPUs.

ronch wrote:

2. The trickiest part where AMD always seems to have trouble is cache sizes, since they, in the past, seem to get things mixed up a little bit. For example, I remember seeing tables on their site indicating L2 cache for the Phenom II to be 512KB, but it didn't say whether it was just for one core or all the cores. They've fixed it now, though. But look at the FX table, specifically, 'L1 Cache (Instruction + Data) per core'. It says each core has '128KB (64KB + 64KB)'. That's a bit off, isn't it? There's 64KB of L1I per module, shared by two cores, and each core gets its own 16KB of L1D. Perhaps you can split the 64KB into two and say each core gets 32KB L1I and 16KB L1D, for a total of 48KB. Not clean cut, but at least you have a better sense of what's in there. However, it's not 64KB + 64KB like the table says.

Piledriver CPUs have 64Kbytes L1 I-cache that is shared across the module. Each ALU in the module has 32Kbytes of D-cache. Thus, 64KB + 64KB per module. Piledriver has double the L1 D-cache per ALU versus Bulldozer.

ronch wrote:

3. It says HyperTransport runs at 4GT/sec., or 16GB/sec. A quick check with CPU-Z, however, tells me my FX-8350's HT runs at 5.2GT/sec, not 4.0GT/sec. Not sure if AMD is being conservative here or if my MSI 990FXA-GD65 is pushing my HT beyond spec, but if some of their models run HT at 4.0GT, they should've at least indicated it the way they indicated that different models have different L3 cache sizes.

HT has never made a great amount of sense to me, but CPU-Z could also be reporting incorrectly. Not sure on this one.

Hi, Auxy,

Yes, I understand that lower end CPUs tend to have less resources, particularly cache. I was just a little bit surprised to learn that the FX-4300 has just 4MB of L3. The FX-4100 had 8MB of L3, and so did the FX-6100. 8MB L3 was standard with Zambezi. L2 cache gets lopped off everytime AMD turns off a module, but L3 pretty much has remained intact back with Zambezi. So with Vishera, things have changed quite a bit, which is kinda disappointing. Since all Vishera chips are the same, it wouldn't make sense for AMD to turn off parts of the L3 on a whim unless they're having yield issues. Product differentiation? Nah. Just look at Zambezi.

As for the L1 caches, I understand that each pair of cores shares 64KB of L1 Instruction cache. Each 'core' (or integer cluster) has its own 16KB of L1 Data cache. Under CPU-Z L1 Data is listed as 16KB x 8, L1 Instruction is listed as 64KB x 4. IIRC Bulldozer presentations also indicate this. Cache sizes remain untouched from Zambezi to Vishera, I reckon.

ronch wrote:

Yes, I understand that lower end CPUs tend to have less resources, particularly cache. I was just a little bit surprised to learn that the FX-4300 has just 4MB of L3. The FX-4100 had 8MB of L3, and so did the FX-6100. 8MB L3 was standard with Zambezi. L2 cache gets lopped off everytime AMD turns off a module, but L3 pretty much has remained intact back with Zambezi. So with Vishera, things have changed quite a bit, which is kinda disappointing. Since all Vishera chips are the same, it wouldn't make sense for AMD to turn off parts of the L3 on a whim unless they're having yield issues. Product differentiation? Nah. Just look at Zambezi.

The FX-4130 also had 4mbytes. I don't find it that unusual. Maybe the larger L3 isn't of benefit with only two modules, or is of minimal benefit. Or, maybe they're having yield issues. I don't find it terribly likely since 32nm is quite mature, however.

ronch wrote:

As for the L1 caches, I understand that each module shares 64KB of L1 Instruction cache. Each 'core' (or integer cluster) has its own 16KB of L1 Data cache. Under CPU-Z L1 Data is listed as 16KB x 8, L1 Instruction is listed as 64KB x 4. IIRC Bulldozer presentations also indicate this. Cache sizes remain untouched from Zambezi to Vishera, I reckon.

No, per AMD's charts I do believe Vishera has double the L1 D-cache from Zambezi. The only sources I can find on this (google vishera double L1) all talk specifically about doubling the Translation Lookaside Buffer, but I believe that resides in the D-cache, so there you go.

TLBs can be arranged in a hierarchy too, hence 'Level 1' TLB and 'Level 2' TLB, but that doesn't imply they are the instruction or data caches at those levels, as the purpose of the TLBs is to map virtual addresses to physical.

So I think the Piledriver L1 TLB increased, but the L1 data cache and L1 instruction cache didn't. Steamroller will increase the L1 instruction cache to 96kb according to leaked documents.

It is 2mb per module. that 2 megs is shared per core on module for the FX series for L2 cache. l3 is static and trimmed or left alone depending on cost.

According to CPU-World, L1D is 16KB per cluster, so that's 16KB x 8 per Vishera die. If you look at the die shot of Vishera and compare it to Zambezi, there's no way the L1D could've doubled. I remember reading somewhere before that they plan to increase data structures in future iterations, perhaps Steamroller, but not Piledriver. All they did with Piledriver is tweak the core, improve this and that, but no larger data structures.

Auxy, isn't it a bit strange that the FX-4100 has 8MB of L3 while the higher model FX-4130 has 4MB? We're not jumping product segments, just a slight bump in clock speed. So increasing clock speed but at the same time halving the L3 just doesn't seem very logical to me unless they're trying to pull back power consumption. I dunno... Between the FX-4100 and 4130, I'd get the 4100 and run it at 4130 speeds. I don't know if you can re-enable the 4MB of 'hidden' L3 in the 4130.

Edit - I checked with CPU-World.. the FX-4100 is a 95W part and the FX-4130 is 125W. The 4100 runs at 3.6GHz, the 4130 at 3.8GHz. So you see, a very small 200MHz (or 5.5%) clock speed increase raises TDP from 95W to 125W, and that's with half the L3. Now I know TDP is not the most accurate measure of power draw but it still gives you a good idea how much power a CPU requires, especially at load. I have a feeling these 4130 chips are somewhat hot samples that get too hot if you clock them at 3.8GHz with the full 8MB of L3.

They're binned differently. Just because there is a correlation, it does not imply a direct casual relationship. Chips themselves are all over the board and are binned accordingly. So they may have loser restrictions for the higher clocked variant.

In other words the higher quality chips are actually the 4100s and the 4130s are just the 4100s that don't meet the criteria for them, including a bigger cache and power envelope. The 4170s don't even appear to be in stock at Newegg.


I'm actually more baffled as to why they're still producing Zambezis. That includes the 8120 and 8150 as well, this is really quite curious and I've noticed this for some time. I can't imagine this still being overstock from more then six months ago... I agree their whole CPU segment is really quite messy. They should've cleaned out the Zambezi variants with the introduction of Vishera. I can't imagine it being a whole lot cheaper to continue to produce Zambezi as Vishera was a straight up replacement.

Bensam, I think it's not unlikely that current Zambezi-based products are old stock. Isn't this similar to Llano where AMD currently still has warehouses full of them because OEMs don't want them because Trinity is better and everyone wants it? I think AMD is finding it hard to sell all its Zambezi chips.

Another possibility is that perhaps part of their Wafer Supply Agreement forces AMD to use a certain amount of GF's capacity else they pay more of those quarterly 'one-time charges'. Perhaps they haven't fully ramped up Vishera and are letting GF make Zambezi chips instead to meet this minimum capacity requirement?

ronch wrote:

Bensam, I think it's not unlikely that current Zambezi-based products are old stock. Isn't this similar to Llano where AMD currently still has warehouses full of them because OEMs don't want them because Trinity is better and everyone wants it? I think AMD is finding it hard to sell all its Zambezi chips.

I find this likely as well.

Think about how clueless AMD upper management was until recently? I'm sure they were drinking the company kool-aid about the performance of Bulldozer and bought massive orders of those chips -- and NOBODY bought them. Sales of Bulldozer were pathetic. I don't find it unlikely at all that AMD still has bucketloads of them hanging around.

They spent millions to get out of the GF contract, I doubt they'd now be relegated to still being stuck to one (or having quotas, they could just produce more Vishera chips)...

I don't know. This could be normal work for them where they sell older chips at cheaper prices, but they really impede on their product lineup and completely muddle the waters. They have slower variants of Vishera for that...

It could be overstock... but that's a LOT of overstock for the last six months and they're regularly on sale. You can usually get a 8150 for less then the price of a 8320, sometimes a bit more. A bit more ridiculous then that is the 8120, which is hovering around the price of a i3... Newegg has one on sale for $145 right now.

I really wish TR would update their price comparisons every few months like they were doing with hard driver prices after the floods. A 8120 for the same price as a i3 is pretty killer.

Bensam123 wrote:

I really wish TR would update their price comparisons every few months like they were doing with hard driver prices after the floods. A 8120 for the same price as a i3 is pretty killer.

Depends on what you're doing! I'd sure rather have the i3, since I mostly play games. First-gen FX is AWFUL at games.