Hacker News Reader

Comment by rowanG077 | Hacker News Reader

I doubt this will happen. Cache is way too important. I would be very surprised to see chips with less cache in the future. I guess we will see in 10 years where the industry has moved. Currently a LOT of the transistor budget is blown on things we don't really need.

View on HN

Yeah, I don't think we'll see reductions in cache, that's always a hard thing because software comes to expect it and can be hurt by reductions, it's kind of a one-way ratchet. But we'll stop seeing the huge increases gen-over-gen, at least on the primary die. Additional transistors spent will probably mostly be spent on logic rather than cache.

N7 basically was the era of "let's throw cache on everything, even products that traditionally haven't had caches". GPUs never had L3 cache before, for example, but that became advantageous, even in GPUs, which are focused around logic/computation rather than deep cache structures. And now you are seeing that train grind to a halt - RDNA3 did not expand the cache further, although they did increase the bandwidth of the cache (2.7x higher, although bear in mind that RDNA3's memory subsystem is 50% wider which means cache bandwidth is effectively 1.8x higher in a relative sense).

Similarly Intel went completely nuts once they finally got to a 7nm-tier node (Intel 7 aka 10ESF). Raptor Lake in particular is just caches all the way down...

Since cache no longer shrinks at 5nm and 3nm, but logic does, it makes sense to do some logic-intensive things rather than just throwing all your area at cache like 7nm.

On the flip side though, since cache can be pulled out to a separate cache die fairly effectively (AMD v-cache), you can continue to scale cache there. RDNA4 and NVIDIA Blackwell are both rumored to be coming fairly quickly which suggests a potential respin. And both AMD and NVIDIA have things they need to work on, NVIDIA doesn't have DP2.0 and AMD seems to have screwed up RDNA3 fairly badly, so a "similar but improved" quick refresh makes sense. Rumors mentioned "[NVIDIA] Ada Lovelace with a stacked cache die" at once point and imo that is very plausible for the next-gen Blackwell chips as a potential quick-fix improvement to keep scaling Ada.

But I think we're going to see an overall trend towards "the logic die is for logic and L3 cache gets stacked on top" for now. That seems to be a formula that works without too much trouble. L1 and L2 on the logic die is unavoidable, there is too much incentive for proximity/latency improvements, but big stacked L3s seem very effective and doesn't cause MCM-style problems.

Yes if cache can be stacked I can see it happening. But I thought this is AMD exclusive tech and other aren't allowed to replicate it.

I don't think there's anything AMD-exclusive about cache stacking. It's made possible with TSV (through-silicon via) and direct bonding (ultrasonic or Cu-Cu bonding) to make the movement energy very low. You just bond the second die into some via you've left in your main die, it's relatively straightforward, just an evolution of the same HBM/HBM2 principles but with cache instead.

Going forward there will also be non-cache (and non-memory!) things bonded as well.

Any specifics on what you think transistors are being wasted on?

Mostly SoC related features. Look at the M1 for example. The budget allocated to a DP controller is HUGE. That doesn't have to be on the SoC chip and can be externalized.

Reply on news.ycombinator.com