AMD Ryzen CPU Reviews and Previews (3000 series)

Hi-res chiplet shot:
Is it me, or the L1i cache array is way too large for 32KB? The L1d cache looks fine, though.
Another thing, the two FMA/SIMD pipes are now clearly distinguishable. Looks like AMD made a new layout for that part of the architecture, not just extend the data paths to 256-bit.
Perhaps the section labelled L1 I$ for Zen included the L1 and uop cache. For both cores, maybe the uop cache is the narrow rectangular bank that juts out from the rest of the L1. For the Zen 2 shot, the top left core might have it run along the entire upper side instead of yielding space to the L1.

For Zen 2, this section is bigger. At the same time, the bottom side has something jutting out a little as well from the L1 block, unlike in Zen 1. However in Zen 1 there are other rectangular regions that are either predictors or a thin rectangular region that was in a nebulous decode block. Perhaps some of the lower parts of that array aren't L1 I$ either?
It might be clearer with a more detailed die shot that actually etches down into the silicon rather than using infrared light. The notes for the shot on Flickr indicate the corner of the Zen 2 chiplet was cracked anyway.

There does seem to be some PHY in the right side of the mid-line of the chiplet, with some blocks that have some resemblance to the IO die's presumed IF PHY. They might be smaller on the CPU die, and I'm not sure what the bright area around it might be. There is a lesser number of such blocks on the left side. From the initial view of this layer of the chiplet, perhaps the IO die is equipped with more IFOP lanes than the current Zen 2 chiplets have.

I'll be curious if profiling of Zen 2's 256-bit operations shows a warmup period similar to what Intel's cores have had since Sandy Bridge. Before, Zen's cracking of operations meant the always-active 128-bit units didn't need to be woken up. The new pipeline might be gating off the unused high side, which might introduce the need for warmup.
 
hD69eHo.jpg


The supposed uCode cache location is definitely enlarged in Zen2, it is just that the Zen1 uCode cache array has odd number of SRAM banks that doesn't exactly maps to the stated 2048 entries.
 
hD69eHo.jpg


The supposed uCode cache location is definitely enlarged in Zen2, it is just that the Zen1 uCode cache array has odd number of SRAM banks that doesn't exactly maps to the stated 2048 entries.
its interesting how the FPU of zen2 looks like a almost literal doubling of Zen1.
 
Is it technically still splitting the 256bit instructions then?
And why don't they get double-rate 128bit?
 
Is it technically still splitting the 256bit instructions then?
No. 256bit instructions generate a single uop.

And why don't they get double-rate 128bit?

While the EUs are duplicated, the execution paths to it still has just 4 pipes. The upper half is attached to it's own register file, so you cannot operate on the lower bits with it.
 
The supposed uCode cache location is definitely enlarged in Zen2, it is just that the Zen1 uCode cache array has odd number of SRAM banks that doesn't exactly maps to the stated 2048 entries.
I didn't see a specific description of the layout of ops in the cache. Trace and uop caches can have different policies on how they map operations that don't fit cleanly into a cache line, like how the P4 split things differently for large immediates or some 64-bit operations. Perhaps Zen 1's cache had some variation in how it packed operations into a line, or had extra space per line it could use for metadata, immediates, or an occasional op that had a less compact decoder output.

Zen 2's larger cache size might mean a change in layout, perhaps a more generous allocation means the arrays come out even again, or area pressure might have meant a trade-off of a larger cache with less forgiving layout.
 
I'm having terrible time choosing a mobo. Is great color scheme which would fit with my aio-watercooling, better vrm and intel nic worth 50€+? (Trying to pick between MSI Gaming Edge Wifi (WiFi is useless for me) and Asus Prime Pro)
 
What I will never understand is AMD choice with the X models that offers exactly 0% improvement over the non X models...All that delay to "get the clocks right" to offer a 25% higher price for a letter in the box?

This video shows that the only difference between the 3600 and 3600X is the cooler and if you put on a good aftermarket cooler on them then there is no difference between the 3600 and 3600X.

The takeaway is just purchase a 3600 and get a good aftermarket cooler and presto changeo you now have a 3600X

 
Last edited:
This video shows that the only difference between the 3600 and 3600X is the cooler and if put on a good aftermarket cooler on them then there is no difference between the 3600 and 3600X.

The takeaway is just purchase a 3600 and get a good aftermarket cooler and presto changeo you now have a 3600X

That is exactly my point, there is no point in getting the more expensive one.
 
Isn't the X binned for a highter clock, instead of hoping to be lucky with the regulars?

No the only difference is that the 3600 has a 65 watt cooler and the 3600X has a 95 watt cooler. If you swap coolers then the 3600 performs like a 3600X and a 3600X performs like a 3600.

The video above clearly points out that there is no difference (and no binning) between the 3600 and the 3600X and all the $50 higher cost for the 3600X is for the 95 Watt cooler and not worth the price increase.
 
The video above clearly points out that there is no difference (and no binning) between the 3600 and the 3600X and all the $50 higher cost for the 3600X is for the 95 Watt cooler and not worth the price increase.
There's not much to go on with a sample size of 1 so there may still be some binning but likely even binning won't make much difference.
 
So far is been the same with every review not only with the 3600 but also with the 3700 vs 3800.


@Malo

This video pretty much puts to rest the idea that AMD is binning for more performance with the same core parts.

3700X vs 3800X hit the same wall at 4.3 GHz.

The two videos pretty much say the same thing. In the 3000 series if you need a hex core CPU pick the 3600 not the 3600X and if you need an octal core CPU pick the 3700X not the 3800X.
 
Seems like all they're doing is having different out-of-the-box versions so anyone who goes and buys a 3800X or a 3600X from Microcenter will get a slightly faster CPU than the 3700X or 3600 due to the cooler. Some silly consumer product segmentation to break up the pricingm gain some profits and provide more than a few models to choose from but anyone who does 10 minutes of research and willing to put their own 3rd party cooler on it (basically 95% of custom PC builders?) will just buy the lower version.
 
Back
Top