22 May 2020

Budget high-tier CPUs?! 3750X &3850X rumours... [UPDATED]

People were really amazed and surprised by AMD's release of the 3100 and 3300X 4/8 C/T processors recently. However, these two processors, powered by Zen 2 chiplets really caught the imagination and passion of enthusiasts and tech reviewers alike for their value to performance ratio. 

In fact, those two processors, are currently "out of stock" across various etailers, or have huge mark-ups on their (M)RRP from third party sellers... and, the processor can't even be found in some territories!

Which leads me to one conclusion - these were limited clearance stock.

The very existence of the 3100 and 3300X showed that AMD had some inventory where the core complexes they used to manufacture the 3700/3800 had defects that disabled 2 cores per chiplet and 4 cores per chiplet. Some of those chiplets with 2 unusable cores gave rise to the 3600 and 3600X, however 4 disabled cores on a CCX could not be utilised within AMD's existing Ryzen 3000 series product stack and neither could 6 disabled cores.

For the longest while, I presume this was because many believed that the yield on the chiplet production wafer was relatively high, with minimal defects. This meant that relatively few chiplets could be binned towards the 3100 and 3300X parts. This assumption appears to have been correct, if the apparent limited commercial supply of those two processors is true (from their apparent sell-out and limited geographical release I noted above).

However, the 3100 and 3300X are not the only possible configuration for those chiplets and, quite frankly, the actual financial return on those parts must be quite minimal once physical material costs (BOM) are taken into account. Looking again at that product stack, it's very clear that AMD very cleverly binned parts into two segments: 6* and 8 operational cores per CCD. Now we have the 4 operational cores per CCD parts**, we have the logical conclusion that doubling the number of chiplets per die would provide another, more profitable product for AMD to produce.
*Consisting of 2x3 or 4+2 cores per CCD. **Consisting of 1x4 or 2x2 cores per CCX (at this point in time 3+1 doesn't appear to have been utilised, if they exist)
You can see the gap in product: the 3900X used two chiplets from the 3600 series, the 3950X used two chiplets from the 3700/3800 series and the 3300/3100 series utilised the chiplets that failed to fail into these categories... 2 and 4 cores per CCX.

Hence the rumour of the 3750X/3850X. WCCFTech posits that these parts are Matisse refreshes meant to compete with the freshly released 10700K and 10900K parts... but that makes no sense to me. In fact, as many commenters point out, these parts are coming very close to the Ryzen 4000 series release... too close, for some to believe. However, I don't find it very hard to believe at all. Instead of the suggested 10 core parts suggested by WCCFTech, I would very much see this as AMD inventory cleaning their defective chiplets now that enough have been amassed through their efficient chiplet manufacturing process.

So what might these theoretical parts be? As I said above, these are likely to be an R7 3750X with two chiplets with 2 cores active per CCX (core complex die explanation) and an R7 3850X with 2x4 cores per CCD on the same CCX (see this explanation of the differences between the 3100 and 3300X). It's likely that these parts might be able to clock a little higher than the 3700X and 3800X due to their more spread-out cores (thus less electrical interference and localised heat generation) so I'd expect 3.9 to 4.0 GHz base clocks for the 3850X and 3.8 GHz for the 3750X. 

I also expect a £50/€50 decrease in price over the existing 3700X/3800X parts.

So what's the catch?

Looking at the above layouts, there's an obvious, more lucrative part waiting in the wings...

The catch is that, just like the 3300X and 3100, I expect these parts to have extremely limited availability and I expect price-gouging after the fact by third party sellers.

If, and when, these parts release... if you want one, snatch it up quickly because I don't think they'll be around for long!! Saying that, I would have expected the lion's share of these defective chiplets to be given to these two products given the increased margins that AMD will be able to reap from their sale.

It also stands to reason that AMD held these parts back - one in order to reduce cannibalism of their more prestigious parts and to reduce confusion from buyers over similarly performing products (inter-CCX/CCD latency being a factor). Secondly, to squeeze every last possible cent from the chiplet design as possible and to wait for enough inventory to build up for anything more than a paper release - and make no mistake, these products only exist because Ryzen 5 3600  and Ryzen 7 3700X were such huge sellers. Without the number of wafers ordered from TSMC by AMD, there would likely not be enough chiplets to be able to manufacture any reasonable amount of these products.

Anyway, I was very pleased by AMD's announcement of the potential for MSI to give me compatibility for Ryzen 4000 series CPUs on my B450 motherboard. However, depending on the price/performance ratio of these chips, a 3850X might be on the cards for me instead...

It seems that if these processors are coming to market, it doesn't seem likely with the incoming releases of the leaked 3600XT, 3800XT and 3900XT - which offer supposedly higher clocks on the 7nm+ process.

No comments: