AMD RyZen CPU Architecture for 2017

If the real benchmarks live up to the hype so far then I've seriously considering getting a 1700 or 1700x. It's been a long time since I've had an AMD CPU but I'll happily go back at those prices.
the 1700 is quite an impressive CPU, and if you take into account the price, 369€ for a 8 core 16 threads CPU.. I am so happy that AMD is back. Hopefully I am going to build a PC with a AMD CPU next year.
 
Yes the Ryzens consume a bit more power, ~10 watts. Though as this is a desktop thats of very little interest to me, laptop sure.
My observation was just that best CPU one can get today is only about twice the speed of the semi good one I brought 4 years ago
This is because we are getting close to the theoretical limit of silicone, I think the limit is 5 or 4 nm so every time will be harder and harder to archive more performance for the same power consumption and I have no idea what companies will do when we get there(anyone knows/have ideas?) It is also a consequence of poor optimize software that in most cases cares more about frequency than threads.
 
the 1700 is quite an impressive CPU, and if you take into account the price, 369€ for a 8 core 16 threads CPU.. I am so happy that AMD is back. Hopefully I am going to build a PC with a AMD CPU next year.
Ryzen 1700 is impressive, because it is 3.0 GHz 8-core at 65W. It has much higher performance, but still comparable perf/watt than Intel's power optimized 2.1 GHz 8-core Xeon D (designed for Facebook). Assuming of course that actual power draw = TDP and both run at base clocks. I am wondering how low clocks would Ryzen 8-core need to hit 45W target.
 
Last edited:
Ryzen 1700 is impressive, because it is 3.0 GHz 8-core at 65W. It has much higher performance, but still comparable perf/watt than Intel's power optimized 2.1 GHz 8-core Xeon D (designed for Facebook). Assuming of course that actual power draw = TDP and both run at base clocks. I am wondering how low clocks would Ryzen 8-core need to hit 45W target.

Assuming a cubic relationship between clock speed and power (i.e. assuming a linear relationship between clock speed per se and power, but a linear relationship between voltage and clock speed, and a quadratic relationship between clock speed and voltage) that would be around 2.65GHz; but that neglects the fact that static power would only scale quadratically (with voltage only) and the hypothesis that clock speed would scale linearly with voltage is somewhat shaky, so maybe closer to 2.5GHz, which is still pretty decent.

Of course that doesn't take binning into account, and you'd expect that better binning would be possible on server parts. All in all I'd roughly guesstimate it around 2.6GHz—maybe with 3.0~3.2GHz Turbo, although that may be of very little importance for that kind of workload.
 
Assuming a cubic relationship between clock speed and power (i.e. assuming a linear relationship between clock speed per se and power, but a linear relationship between voltage and clock speed, and a quadratic relationship between clock speed and voltage) that would be around 2.65GHz; but that neglects the fact that static power would only scale quadratically (with voltage only) and the hypothesis that clock speed would scale linearly with voltage is somewhat shaky, so maybe closer to 2.5GHz, which is still pretty decent.

Of course that doesn't take binning into account, and you'd expect that better binning would be possible on server parts. All in all I'd roughly guesstimate it around 2.6GHz—maybe with 3.0~3.2GHz Turbo, although that may be of very little importance for that kind of workload.

Because the power regulation is per core the 32 core part could be really really interesting particularly with VM farms with clocks all over the place between cores. What amd set the max turbo at for a 180watt TDP part would be interesting to see.
 
Because the power regulation is per core the 32 core part could be really really interesting particularly with VM farms with clocks all over the place between cores. What amd set the max turbo at for a 180watt TDP part would be interesting to see.

Well, that depends. It's a 4-die chip, so is it 180W per chip with no restrictions, or up to 45W per die?
 
Well, that depends. It's a 4-die chip, so is it 180W per chip with no restrictions, or up to 45W per die?
Intel has also 16-core (32 thread) Xeon Ds at 45W. But base clock is only 1.3 GHz (2.1 GHz turbo). Also only dual channel memory controller (like Ryzen 8-cores):
https://ark.intel.com/products/93353/Intel-Xeon-Processor-D-1577-24M-Cache-1_30-GHz

Comparing Ryzen 1700 against this processor is also interesting. Ryzen has 2.3x higher base clock, Intel has 2x cores. Ryzen will offer better throughput even if we assume perfect multi-threaded scaling of Intel part. But Ryzen 1700 has 65W TDP and Intel's 16-core Xeon D has 45W TDP. So Intel likely has slightly better perf/watt. Intel chip however costs 4x (1477$ vs 369$). Ryzen core seems to be perfect for this market as well. Interesting to see what kind of server products AMD introduces. Ryzen with 16 cores at ~2 GHz would be perfect against Xeon D. Needs to be single die SoC (integrated network controller).
 
eBsYLq2.jpg


Skylake for reference:

Djvm0PX.png


Note: AIDA's memory bench is multi-threaded.
 
Intel has also 16-core (32 thread) Xeon Ds at 45W. But base clock is only 1.3 GHz (2.1 GHz turbo). Also only dual channel memory controller (like Ryzen 8-cores):
https://ark.intel.com/products/93353/Intel-Xeon-Processor-D-1577-24M-Cache-1_30-GHz

Comparing Ryzen 1700 against this processor is also interesting. Ryzen has 2.3x higher base clock, Intel has 2x cores. Ryzen will offer better throughput even if we assume perfect multi-threaded scaling of Intel part. But Ryzen 1700 has 65W TDP and Intel's 16-core Xeon D has 45W TDP. So Intel likely has slightly better perf/watt. Intel chip however costs 4x (1477$ vs 369$). Ryzen core seems to be perfect for this market as well. Interesting to see what kind of server products AMD introduces. Ryzen with 16 cores at ~2 GHz would be perfect against Xeon D. Needs to be single die SoC (integrated network controller).

Ryzen has a 10GbE controller, so Naples should have 4. That ought to be enough for at least some applications.
 
So twice as many cores in queue for memory access => twice the latency? (while for caches it should be like hoom said as they got the same amount of cores per cache)
Probably the memory controller in Ryzen is oversubscribed in this case. Bandwidth looks good, though.

Anyone with an 8-core LGA2011 system willing to make comparison with 4 vs. 2 channel config?
 
Common wisdom says they will tune for lower cost and lower power consumption per transistor and stack things vertically.
That is not really scalable, if you have 3 stacks you have 3 times power comp. and temp...unless we find ways to use grapheme or (maybe hydrogen?) to make transistors it will be very difficult to keep making significant improvements of performance. Maybe a full AI design? IDk...
 
really good, not the best review in the world but its something. Why winrar is so low? and 95W extra power with OC lol (although it looks like an ES since it doesnt have the final name in the CPUz.)
 
Back
Top