32 core EPYC v1 is inferior to 32 core TR in some respects, superior in others. TRv2 has the ryzen+ dies, EPYC v1 does not, but it has 8 channel memory and direct access to memory (some of it at least) from each die. Sure some memory channels aren’t accessible directly from all dies like TR, but in TRv2 2 complete dies can’t access any RAM directly at all (if i understand correctly).
Still, if the bulk of the running code/frequently accessed data fits in CPU cache, TRv2 should be superior to EPYC in those situations.
Didn’t think it needed to be mentioned to be honest. This is very clearly intel being “ME TOO!!!” and overclocking the absolute snot out of a $10,000 CPU just to “show” that they have something in the ball-park. It is in no way a real product.
AMD can likely put out 32 core TRv2 for $1500-$2000 US and still make a profit on that… as it is just 4x Ryzen dies with some “glue” (to quote intel ) and they’re retailing at what… $350 each for 2700X? Including a good box cooler with RGB??
Even if intel could run 5Ghz on air on that thing… there’s just no way they will be able to compete on price or price:perf with monolithic dies of the size required. Even at a very good defect rate, the yields on dies as big as their 28 core chip are horrible. And they can’t move it to their 10nm process to help, because the yields will be even worse.
Those slides from last year where they were poo-pooing the idea of multi-die packages in reference to EPYC really is going to come back and bite them in the arse when they need to do it themselves. And they will.
AdoredTV has a very interesting video put up recently on yields vs. die size, if you guys haven’t checked it out, i’d highly recommend it.
Can’t remember if it was already linked in thread but here…
Also, AMD should one-up intel with TRv2…
V1 threadripper hits 5.2 Ghz on LN2. I reckon with enough LN2 the dies are spaced apart far enough to maybe do this with 32 cores/4 dies as well