> I could replace my 6-year old Intel i7-3770 desktop with 32GB of RAM, but what's out there that's significantly better at a reasonable cost?
AMD Ryzen 7 1700 for $200. ~10% faster single core and you get 8C/16T. DDR4 RAM is quite expensive, though.
(closest available benchmarks, 1700 is about 3% slower than 2700) [0]: https://www.anandtech.com/bench/product/2111 [1]: https://www.anandtech.com/bench/product/551
The system the i7-3770 replaced was a 4 year old dual-core, with (IIRC) just 1GB of RAM. The i7 was a dramatic improvement, but I don't feel that the Ryzen would be a dramatic improvement.
And that's the issue. It is not just me being a little unimpressed with current offerings, or me just being cheap. The issue is that there is even a valid discussion about replacing a 6 year old desktop. In the past, in the 1990's and 2000's, it wasn't a question. You had to upgrade because you wanted to run Windows XP decently, for example.
In the "old days", there was also demand from upgrades, but we're seeing that tapering off. And that should be a big concern for the investors.
So if you can benefit from 8Cores (ie compiling for one) then it is a nice upgrade.
If you rarely utilize all 4 cores on your 3770 then it is a pointless upgrade.
That is the thing, going from a single core to 2 core pretty much benefited everyone. Heck PentiumHT was a big thing.
Going from 2 core to 4 cores took longer to show benefits and 4 cores to 8 cores is going to be slower still to show benefits for most users.
At the end of the day single core speed has been stagnating for the last 6 years every since Sandy Bridge which was the last big jump.
I mean I plan on buying AMDs ThreadRipper 2 32 core chip since it will be of benefit to me for running 100 copies of the same program.
Still those 32 cores will not help with regular computing.
True, but it is a major upgrade.
The i7-3770 supports two channels of DDR3-1600 memory, or 25.6 GB/s of memory bandwidth. A Ryzen processor (or other recent generation) supports DDR4-3200, which literally doubles your memory bandwidth. Fortunately or unfortunately, we now have much faster PCIe SSDs so memory bandwidth isn't as crucial for normal use, but a 100% improvement is a much better reason to upgrade than a measly 10% bump.
AMD Ryzen 7 2700 Passmark Score (<1yr old, ~$290): 15382 [2]
That's a 65% improvement in benchmark performance. So not the "doubling" trend I was used to growing up, but not insignificant.
[1] https://www.cpubenchmark.net/cpu.php?cpu=Intel+Core+i7-3770+...
[2] https://www.cpubenchmark.net/cpu.php?cpu=AMD+Ryzen+7+2700&id...
The stuff I've read about the new process nodes from the last decade indicate that moving to those hasn't been easy. And the chip designers haven't had it easy either, with each new process comes a longer and longer list of restrictions and other difficulties.
Transitioning an existing CPU design (for example) to a new process node was never easy per se, but it has never been harder than it is these days. That's part of the slowdown.
And that's great. But that'll also be a one-time thing in a sense. After it has moved to the latest process node, it will be on the same trajectory as other architectures.
Again, what's the valuation of Intel if (in the future) we see them churning out the same 1GB NVMe cards at the same cost with the same performance? The valuation of Intel and every other tech company is based on future growth potential.
However, the planar silicon scaling is what primarily enables cost reduction for production. I expect the (total device) density to continue to increase with the 3D tech, but cost will go up with capacity increases. Maybe not linearly though, because you don't have to pay for extra packaging.
The 3D memories will probably see some planar (x-y) scaling, they just don't need all of their scaling from it. 3D scaling will still increase bits/die area, and thus reduce $/bit. If doubling the stacking height doubles capacity but adds 30% to wafer cost with extra processing steps, its still much more efficient than 100% cost for two wafers of the previous design.
Scaling efficiency shouldn't be too dissimilar from planar scaling, at least until the next wall is hit.
The whole semiconductor industry is in trouble right now, though few seem to feel the weight of the issue.
We are looking at just a couple process nodes at best before we reach the end of the road for silicon lithography. Sure, there will be further tweaks on existing techniques which will squeeze out small improvements in power, performance and density.
But long gone are the days when we saw steady improvements in circuit density and simultaneously speed and simultaneously cost.
When this knowledge finally sinks in with the investment community, it will call into question the valuation of the entire computer industry. We're already seeing that in the desktop space. I could replace my 6-year old Intel i7-3770 desktop with 32GB of RAM, but what's out there that's significantly better at a reasonable cost? Well, a used Xeon workstation maybe, but that's about it.
That should be a big red flag to the investment community, but for reasons I don't understand, people don't seem to care yet.