5 ms·
Not much. Clock wall, power wall, and latency wall are here. The only thing improving is memory size, bandwidth, and number of cores. Now, we are still very f
by BenoitP 8y ago
Not much. Clock wall, power wall, and latency wall are here.
The only thing improving is memory size, bandwidth, and number of cores.
Now, we are still very far off from every worload being at its theoretically maximum throughput (that would be achieved with an ASIC).
Over the next years, the interesting things won't come from Intel or AMD. The generic processor killing all competition has reached an end. Intel could just wait till the next technology node to thwart off any attempt of adapting to a corner of the market. This is no longer the case; with so much money at stake, expect a Cambrian explosion.
It will come from specialized workloads getting an accelerator (think integrated TPUs, NICs with FPGAs, etc); it will come from tighter chip packaging/integration (think close-access DRAM a few μm from the CPU with silicon vias, early filtering with predicate pushdown directly into the SSD controller, etc.).
Now, the code will have to adapt; and I posit that the languages better suited for this should have:
* An easy way to control the memory layout
* An easy way to insert special-purpose instructions
* A compiler infrastructure flexible enough to be able to coerce code into the accelerators
* An easy way to serialize (possibly half-executed-)code that should be transported closer to the data
- gameswithgo 8y agoHave you seen the Burst Compiler in Unity3d?
- BenoitP 8y agoJust looked it up, and I am sort of failing to find a proper resource. It is about using SIMD capabilities?
- gameswithgo 8y agothere is a video by mike acton about it somewhere. But yes, the current iteration lets you use a subset of the usual language (C#) and will SIMD it, but I believe its intended to be able to send the same stuff off to a GPU if desired as well.
- ksec 8y ago>Not much. Clock wall, power wall, and latency wall are here. This. I will add one more, price Wall. The cost to design a CPU along with cost of leading node is increasing, while the scale and ASP are not increasing. Cost of building new Fab is increasing, cost of Fabs equipment is increasing, cost of design tools are increasing, they could be double the price of what it was 10 years ago, and the scale of PC unit has been mostly flat or declining for number of years. As this trend continues, 5nm, 3nm somewhere around these TSMC nodes there will be less players able to afford these. For Intel that means they will sell it at a smaller die ( higher margin ) and try to recoup their investment asap. For AMD there are lots of improvement, mostly catching up Intel. Zen 2 will increase IPC, 7nm allows them to double the core count AGAIN. Zen 3 with 7nm+ will have DDR5, PCI-E 5?, and other improvement. So the 10/7nm cycle will dramatically increase core count, thanks to AMD. Unfortunately DRAM and SSD are still expensive, and it seems the Cloud Sector has infinite appetite for DRAM and SSD.
- NightlyDev 8y agoSSDs isn't still expensive. Prices has gone way down lately...
- ksec 8y ago>SSDs isn't still expensive. Prices has gone way down lately... Not expensive, may be that is not the word I should have use. But according to [1] ,its price is barely the same as 18 months ago, and if you look at longer term, price hasn't dropped much compared to 24 to 36 months ago. And I was speaking in the context of servers, in terabyte of Memory and Petabytes of storage. [1] https://pcpartpicker.com/trends/price/internal-hard-drive/ https://pcpartpicker.com/trends/price/internal-hard-drive/