4 ms·
Ok, but based on the die size and the transistor count we could have had 12 Thuban cores in the same amount of space. It really feels like if we are going to g
by sitkack 15y ago
Ok, but based on the die size and the transistor count we could have had 12 Thuban cores in the same amount of space.
It really feels like if we are going to go with more cores, why not make them simpler? Why not dust off the old Athlon K7 at 37 million transistors. Roughly 54 of those fit into the 2 billion transistors making up the 8 core bulldozer. Bulldozer has an Itanic feel to it, no that is too harsh, it has a P4 feel to it.
I really feel like desktops are done, GPUs have this biggest effect on desktop perf right now. It would seem to focus on the fusion A series parts and the server. As a server side dev, I would rather have more cores (more cache actually). I don't need memory coherency (that should be mitigated by the hypervisor).
- rkalla 15y agosit, you are exactly right and they are going that route with the Opteron's. The 16-core Opterons were announced a while ago, I believe there is GPU infrastructure in the desktop chips (similar to what Intel is doing with the always-includes-a-GPU-now CPUs). I imagine when we see the 16-core AMD's next year on 22nm, we'll have 24-core Opterons or something equivalent that is a boon to the hosting industry. Not saying Intel won't be right there with their own solution, but going the "more cores" route should be interesting to watch and see if it pays off.
- sliverstorm 15y agoFirst, power. Second, the old K7's are not as fast as you remember.
- sitkack 15y agoI don't think power is it. Your saying the K7 generates more heat per computation (more inefficient use of transistors) than the newer cores? I was about to conflate power usage with computational work, not necessarily the same thing... The power will drop using the new process because the voltage is much much lower (1v vs 2v) I remember the K7 doing just over 1 GOP (giga operation per second), so that would put the whole cluster in 50 GOP range (if clocked at only 1Ghz like the Thunderbird). I wasn't being directly literal, I will allow these mythical engineers that work for me using all the 2 billion transistors to refine things from previous designs.
- sliverstorm 15y agoIt is true the power usage of a K7 per op will drop if you redesign the core to use 32nm. That's a key detail though- you can't just drop a 130nm design onto a 32nm mask. There are many hitches with this idea though. K7 cores have limited power management capability; modern cores have sophisticated frequency scaling, and shut down sectors that are idle. K7 cores are 32-bit. K7 cores do not support most modern extended instruction sets. 50 active K7 cores on one die will generate way more than 125W, causing cooling problems. In short, there are a lot of issues with dated processors that we don't think about, because we take the advances for granted. This is not to say your idea does not have merit. If it was I, I would probably first investigate using a Bobcat core. The Bobcat core has low power usage for a x86 core (good for packing a large number of cores on one die), and considering a single Bulldozer 'module' (core) has 200M transistors, the Bobcat probably has ?100M?, allowing you to fit ?20? cores with 2 threads each. P.S. If you are about to brush off the extended instruction sets bit, consider who a 50-core die would interest most. My guess is the scientific community would be a large part of the target markets, and I believe they make great use of extended instruction sets.