AMD Geekbenchs for engineering examples???

[shdowhunt60]

Just now, Dabombinable said:

5. With a better cache arrangement 1 of the integer units/ALU wouldn't end up waiting for the other to finish its task in some circumstances. 

There were definitely hiccups in Bulldozers design, but the idea really isn't that fundamentally flawed. 

[Dabombinable]

4 minutes ago, shdowhunt60 said:

There were definitely hiccups in Bulldozers design, but the idea really isn't that fundamentally flawed. 

The idea itself is actually good, however, since its only capable of being properly utilised in a few niche circumstances, most of which consumers won't come across, it really shouldn't have been pushed forward in consumer level CPU/APU the way it was.

[Prysin]

2 minutes ago, DocSwag said:

AMD decided to save die space, they would cut out FPUs from the CPU that aren't really fully used most of the time. The idea was decent, but the execution was terrible. They ended up with a CPU that had less than if not only equally as good IPC as their previous generation and that, because it had less FPUs, would lag behind in some applications as well as gaming. Basically, it was a decent idea in theory that was executed horribly and may not have been as good of an idea as it seemed.

Phenom II (P2) doesnt outright beat FX at gaming. Don't even bother to try pull that one. Dab does it all the time, and i disprove him all the time.

P2 IPC is higher, but clocks are lower. Turbo is not as fine grained and as a result P2 cannot turbo all cores, only 1/2 of the cores. The net result is a dramatic loss to multi-threading performance. Overclocked, P2 can beat overclocked FX thread for thread. But a FX 8 has two more threads, and thus is able to beat P2, but the IMC in P2 is worse then FX (and FX IMC is bad), during overclocking the IMC is put under immense strain, and thus IMC capacity can be a limiting factor for P2.

 

Games has been using 4 threads since 2006 at least (source engine got upgraded to use 4 threads back then). Cry-engine 1.0 and Crysis 1 could use up to 6 threads i believe, but it was heavily reliant on one thread for most of the tasks.

In games using 4 threads, P2 can only beat FX if the P2 chip is OCd manually. If left at stock it will NOT beat FX, mainly due to turbo not allowing more then half the cores to run at a higher speed.

 

CMT was built for servers. It is based upon a mix of Simultaneous Multi-threading (Hyper Threading) and Clustered Integer Core (CIC). This mix allows for a CPU capable of immense amount of IO work while doing light tasks. Excellent for servers, and actually quite efficient. Which is why AMDs Opterons was selling really well until Ivy Bridge Xeons came out and started reclaiming lost ground with their much much higher IPC and higher thread count.

However, in desktop workloads, which is not IO based, it isnt nearly as effective. It will lag behind due to its low IPC.

Yet, CMT is also a bit clever. I said earlier it is SMT + CIC. Which is true. If coded for correctly, AMDs CMT solution is able to behave like Integer ALUs and SMT Float/Vector. Meaning if you code for FX, it can do both IO and float/vector simultaneously. However, such coding is mostly only found in server space where you build custom software. Games doesnt really use this feature and will probably run either float, or int/alu.

An ALU is a basic arithmetic unit, however there is more advanced forms of ALUs that can do more of the types of work you would do with float/integer units. However these are slower to do this task.

It is not clear which type of ALU AMD has used in their CMT architectures.

 

[DocSwag]

2 hours ago, Prysin said:

Phenom II (P2) doesnt outright beat FX at gaming. Don't even bother to try pull that one. Dab does it all the time, and i disprove him all the time.

P2 IPC is higher, but clocks are lower. Turbo is not as fine grained and as a result P2 cannot turbo all cores, only 1/2 of the cores. The net result is a dramatic loss to multi-threading performance. Overclocked, P2 can beat overclocked FX thread for thread. But a FX 8 has two more threads, and thus is able to beat P2, but the IMC in P2 is worse then FX (and FX IMC is bad), during overclocking the IMC is put under immense strain, and thus IMC capacity can be a limiting factor for P2.

 

Games has been using 4 threads since 2006 at least (source engine got upgraded to use 4 threads back then). Cry-engine 1.0 and Crysis 1 could use up to 6 threads i believe, but it was heavily reliant on one thread for most of the tasks.

In games using 4 threads, P2 can only beat FX if the P2 chip is OCd manually. If left at stock it will NOT beat FX, mainly due to turbo not allowing more then half the cores to run at a higher speed.

 

CMT was built for servers. It is based upon a mix of Simultaneous Multi-threading (Hyper Threading) and Clustered Integer Core (CIC). This mix allows for a CPU capable of immense amount of IO work while doing light tasks. Excellent for servers, and actually quite efficient. Which is why AMDs Opterons was selling really well until Ivy Bridge Xeons came out and started reclaiming lost ground with their much much higher IPC and higher thread count.

However, in desktop workloads, which is not IO based, it isnt nearly as effective. It will lag behind due to its low IPC.

Yet, CMT is also a bit clever. I said earlier it is SMT + CIC. Which is true. If coded for correctly, AMDs CMT solution is able to behave like Integer ALUs and SMT Float/Vector. Meaning if you code for FX, it can do both IO and float/vector simultaneously. However, such coding is mostly only found in server space where you build custom software. Games doesnt really use this feature and will probably run either float, or int/alu.

An ALU is a basic arithmetic unit, however there is more advanced forms of ALUs that can do more of the types of work you would do with float/integer units. However these are slower to do this task.

It is not clear which type of ALU AMD has used in their CMT architectures.

 

I wasn't saying P2 is better than fx. I was saying its IPC was better. But yes, I agree that CMT was a clever idea, but it's execution was bad and there just wasn't a platform for it.

[Doobeedoo]

That convo about Zen APUs... of course thery will come. What would make you think they wouldnt?! After main high end Zen CPUs on desktop they will come. Also, for laptops up to 4 cores afaik.

[Prysin]

3 hours ago, DocSwag said:

I wasn't saying P2 is better than fx. I was saying its IPC was better. But yes, I agree that CMT was a clever idea, but it's execution was bad and there just wasn't a platform for it.

the way you presented the case, your post can come across as "P2 was better lol". Just saying.

 

6 hours ago, Dabombinable said:

The idea itself is actually good, however, since its only capable of being properly utilised in a few niche circumstances, most of which consumers won't come across, it really shouldn't have been pushed forward in consumer level CPU/APU the way it was.

Actually, its not hard to utilize CMT properly, what is needed is for AMD to write a instruction and a optimized compiler for the CPU and release it to the public. By doing so, they could have  made sure every software could be using optimized instructions when detecting a "Genuine AMD" product. Much like Intel made a "cheat" compiler for benchmarks, AMD could have done the same with games.

 

AMD could have done this, they could still do this, and if they did, it is very possible that CMT based CPUs could gain significant increase in performance. How much, i do not know. But i would not be shocked if it turned out to increase performance by 20% or more.

 

And infact, the Cache in the CMT based CPUs has so wide a bus that it can easily handle two ALUs/Integers requesting work, the bottleneck would be in the controllers fetching and sending info. In the later CMT versions (Steamroller and Excavator) the fetch/send units where split into 1 unit per ALU core, rather then 1 unit per module. This helps optimize the cache use.

Then there is the fact that Intels CPUs has a smaller bus, but much much faster transfer-speeds.

 

The fact that AMD is planning to recycle some of the cache structure from their current CPUs for ZEN are seriously worrying me. The bandwidth readouts for their cache is NOT good. It is essentially 30-50% slower then a Haswell CPUs cache.

The only way to speed up AMDs cache is by baseclock OC. Surprisingly, AMDs FX baseclock are not tied to all the controllers, it has more then one clock. Much like Skylake does. Meaning you CAN boost its performance by running a higher baseclock.

[Coaxialgamer]

Now that i looked at the results again , they seem pretty much irrelevant since we have no idea of if turbo was used on the AMD engineering sample , and if it was used on comparable intel chips.

[Prysin]

3 hours ago, Coaxialgamer said:

Now that i looked at the results again , they seem pretty much irrelevant since we have no idea of if turbo was used on the AMD engineering sample , and if it was used on comparable intel chips.

well, here is some numbers for the Athlon X4 845 (as i promised earlier)

 

My Setup:

PCPartPicker part list / Price breakdown by merchant

CPU: AMD Athlon X4 845 3.5GHz Quad-Core Processor  ($67.60 @ NCIX US)
CPU Cooler: Noctua NH-D9L 46.4 CFM CPU Cooler  ($56.00 @ Amazon)
Motherboard: Gigabyte GA-F2A88XN-WIFI Mini ITX FM2+ Motherboard  ($96.99 @ SuperBiiz)
Memory: Crucial Ballistix Tactical Tracer 16GB (2 x 8GB) DDR3-1600 Memory  ($64.99 @ Amazon)
Storage: Western Digital Black2 1TB 2.5" 5400RPM Hybrid Internal Hard Drive  ($199.99 @ B&H)
Storage: Hitachi Ultrastar 500GB 3.5" 7200RPM Internal Hard Drive  ($49.50 @ Amazon)
Video Card: Asus GeForce GTX 950 2GB Video Card  ($148.99 @ SuperBiiz)
Case: RAIJINTEK Metis (Black) Mini ITX Tower Case  ($49.99 @ Newegg)
Power Supply: Corsair RM 750W 80+ Gold Certified Fully-Modular ATX Power Supply  ($118.99 @ NCIX US)
Operating System: Microsoft Windows 10 Home OEM (64-bit)  ($85.99 @ NCIX US)
Case Fan: Noctua NF-F12 PWM 55.0 CFM 120mm  Fan  ($19.48 @ Amazon)
Keyboard: Logitech K830 Wireless Standard Keyboard w/Touchpad  ($99.99 @ Best Buy)
Total: $1058.50
Prices include shipping, taxes, and discounts when available
Generated by PCPartPicker 2016-04-13 10:55 EDT-0400

 

Forgot to add: Cheap'o Red LED strip from Akasa, 15 USD.... Not sure if it impacts performance, but it looks ballin.

 

Baseclock is set to 106, due to baseclock being altered the memory is running at 1700MHz, the CPU is running at 3.73Ghz stock and 4077.3MHz MAX Turbo.

I am not going to bother changing my BCLCK OC to show stock results.

 

First Benchmark:

 

CPU-Z Bench:

Single Thread = 1278

Multi Thread = 4814

 

For reference:
A I7 3960X 3.3GHz 6c/12t (Ivy extreme edition) scores 1250 single thread and 7600 multi thread.

A FX 8150 3.6GHz 4m/8t (orig. Bulldozer) scores 803 single thread and 5100 multi thread.

 

Cinebench R15 (Cinebench is very cache dependent. and thus Excavator CPUs tank hard in this bench. Their L2 is faster, more responsive, but its still just 2MB)

Single Thread = 95

Multi Thread = 325

Scaling: 3.42x (pretty good actually)

 

Firestrike 1.1 (1080p)

Physics score (Phys runs off CPU) = 5015

Total Score = 5012

 

Ashes of the Singularity (1080p High settings)

DX11 test (does not reveal GPU or CPU dependency)

Avg. FPS = 25.5

FPS normal batch = 30.3

FPS medium batch = 26.2

FPS large batch = 21.6

Driver Overhead = 25% ish

 

DX12 test

Avg. FPS = 25

Avg CPU FPS = 34.6 <- thats the max frames the CPU are capable of pushing in the benchmark.

FPS normal batch = 28.7

FPS medium batch = 26.3

FPS large batch = 21.2

GPU dependency small batch = 88% -> 12% CPU or IO bottleneck

GPU dependency medium batch = 98% -> 2% CPU or IO bottleneck

GPU dependency large batch = 87.1% -> 12.9% CPU or IO bottleneck

Driver Overhead = 5-10% ish