backtop


Print 64 comment(s) - last by Parhel.. on Aug 27 at 3:52 PM


CPU ID screen shot of "Yorkfield" at 2.33 GHz  (Source: DailyTech)

Intel "Yorkfield" ScienceMark 2 L2 cache performance  (Source: DailyTech)
DailyTech managed to snag a quad-core "Yorkfield" for a few quick benchmarks

Benchmarks of Intel’s Penryn based dual-core Wolfdale have appeared a couple times in the past month. The early benchmarks tested engineering sample processors and showed Wolfdale, on average, performing 5 percent faster, clock for clock then Conroe. However, benchmarks of the quad-core Yorkfield are virtually non-existent to the public.

Intel’s Yorkfield is not a native quad-core design. As with Kentsfield, Yorkfield features two dual-core dies fused together. The design results in each pair of cores having access to its own pool of shared L2 cache. Since Penryn has more cache, each pair of cores has access to 6MB of L2 for a total of 12MB – up from the 4MB per pair and 8MB total of Kentsfield.

In addition to the increased cache size, Penryn features a faster 24-way associative L2 cache, which cuts off a few clock cycles. Kentsfield has an 16-way associative L2 cache.

Penryn
also features new SSE4 instructions catered towards multimedia tasks. SSE4 introduces 47 new instructions to improve performance of video accelerators, graphics building blocks and streaming load. Intel claims a 2x performance gain in video acceleration tasks. There are 14 new instructions for video accelerator performance enhancement. Intel improves compiler auto-vectorization performance with 32 new instructions.

Intel expects SSE4 optimizations to deliver performance improvements in video authoring, imaging, graphics, video search, off-chip accelerators, gaming and physics applications. Early benchmarks with an SSE4 optimized version of DivX 6.6 Alpha yielded a 116 percent performance improvement due to SSE4 optimizations.

Also new to Penryn is the Super Shuffle Engine. Intel’s Super Shuffle Engine allows for shuffling unpacking, packing, align concatenated sources, wide shifts, insertion and extraction, and setup for horizontal arithmetic functions. Intel claims a “2x faster SSE shuffle instruction execution,” according to earlier briefing documents.

Although Yorkfield uses a 45nm fab process and consumes less power, Intel plans to stick to its existing 95 Watt and 130 Watt thermal design power ratings.

DailyTech previously presented quick and dirty benchmarks of AMD’s 1.6 GHz Barcelona processor last June. Today, DailyTech has a few quick and dirty benchmarks of Intel’s quad-core Yorkfield Core 2 processor, in an LGA775 package.

The testing configuration is as follows:
  • Intel Core 2 Extreme QX6700 @ 2.33 GHz, 1333 MHz front-side bus
  • Intel Yorkfield 2.33 GHz, 1333 MHz front-side bus
  • Intel P35 Express based motherboard
  • 2x1GB DDR3-1333 memory
  • AMD ATI Radeon HD 2600 XT
Since Intel does not have a 2.33 GHz Kentsfield processor, a Core 2 Extreme QX6700 is used. The Core 2 Extreme QX6700 has an unlocked multiplier, which allowed us to clock it at 2.33 GHz with a 1333 MHz front-side bus.

 SiSoft Sandra XII CPU-Arithmetic
MIPS

Kentsfield
2.33 GHz
Yorkfield
2.33 GHz
ALU
43003
43299
FPU
29981
34693

 SiSoft Sandra XII CPU Multimedia
MIPS

Kentsfield
2.33 GHz
Yorkfield
2.33 GHz
ALU
257295
256216
FPU
140055
140301

 SiSoft Sandra XII Memory Bandwidth
MB/s

Kentsfield
2.33 GHz
Yorkfield
2.33 GHz
ALU
6639
7124
FPU
6639
7121

Synthetic benchmarks do not really reveal too much of a performance difference between Kentsfield and Yorkfield. However, SiSoft Sandra XII does not contain SSE4 optimizations yet.

Unlike AMD, Intel relies on an off-chip memory controller. Although AMD achieves low latencies with its integrated memory controller, Intel manages the same feat with a northbridge-installed controller. Intel managed to offset the latencies associated with off-die memory controllers with increased L2 cache. Yorkfield’s additional L2 cache and speedier 24-way associative L2 cache yields an approximate memory bandwidth boost of 7 percent.

 Cinebench 10 Performance
CB-CPU

Kentsfield
2.33 GHz
Yorkfield
2.33 GHz
Single
2400
2582
Multithread
8518
9206

 DivX 6.6
Time

Kentsfield
2.33 GHz
Yorkfield
2.33 GHz
Seconds
12.90
11.80

Cinebench 10 yields an approximate 8 percent boost in single and multithreaded rendering. Encoding a video file into DivX also yields a similar 8 percent performance boost.

Overall, with our limited time with Yorkfield, performance of the quad-core processor is roughly 8 percent faster clock for clock than Kentsfield. However, this is expected as Yorkfield is essentially a 45nm die shrink of Kentsfield with a few tweaks here and there.

Expect Intel to begin shipping Yorkfield in mass quantities in Q1 2008. Quad-core Xeon X5400 Harpertown processors, which are somewhat similar to Yorkfield, will ship in November.


Comments     Threshold


This article is over a month old, voting and posting comments is disabled

RE: Not bad
By retrospooty on 8/24/2007 12:11:07 AM , Rating: 2
Since the cache has improvements, and is also larger, we may well see it scale better. In other words its an avg of 5% faster at a meager 2.33 ghz, but what about 3.33, or higher which is on tap for later in 2008. I am sure could be achieved now by raising voltage. It may be quite a bit faster.


RE: Not bad
By mars777 on 8/24/2007 5:02:15 AM , Rating: 4
4 facts:

1. Cache is there to compensate memory access latency
2. More cache is needed to compensate inter-cpu communication
3. Higher clocked is the CPU, it makes more inter-cpu and memory requests.
4. Clocking higher the CPU retains the original quantity of cache

Well now try to rethink and post your conclusions :)

The bottleneck is in the cpu-to-cpu and cpu-to-memory bus...
Overclocking needs higher FSB, and that is exactly what you do when you overclock, raise FSB and raise memory clock. On intel systems this is true and this means clocking higher the CPU too. That is because of the front side bus deriving from the northbridge, it dictates the memory AND cpu-to-cpu bus speed. So when an (current) Intel CPU goes higher in clocks it needs more data through its FSB, but the cache remains the same - and the front side bus too (since faster chips have only higher multiplier). When you overclock the cpu you overclock the cache too, but this means nothing since its merely a requirement. If it didn't overclock with the processor you would have SEVERE performance drops.
Conclusion: current amount of cache is what Intel calculated as the right amount for Penryn bandwidth needs and improvement for applications (encoding is one of those that will benefit much). The amount of cache is generally computed based on architecture, and scaling is mainly dictated by architecture too. I've read somewhere that C2D doesn't need more than 3MB of cache, as more wouldn't help anything but applications that rely mainly on cache (ie encoding).

On AMD systems there is another bus in the way. The Hypertransport bus is for CPU-CPU and CPU-Peripherals communications. The FSB there is something fictional. Memory clock is postulated by division of the CPU clock - because of the integrated memory controller. And this is the reason why AMD cpus scale better with speed (with higher multipliers) - because you clock higher the memory controller too (giving the same memory speed you still have lower latencies)!


RE: Not bad
By Targon on 8/24/2007 5:15:39 AM , Rating: 3
But if you plan to put these things into a server, you do NOT overclock, because stability is more important. It is also possible that at this point, Intel is having problems ramping the speed of these things up in the way that pre-B2 stepping Barcelona processors could not be clocked up to 3GHz.

Anything new, from a new design to a new process technology will take some time to get right, which is why the new chips are yet to be released. The interesting thing is that so many people were saying how Peryn would eliminate any chance of AMD catching up or regaining the performance lead.


RE: Not bad
By mrdelldude on 8/24/2007 10:25:26 PM , Rating: 2
quote:
But if you plan to put these things into a server, you do NOT overclock, because stability is more important.


True

quote:
It is also possible that at this point, Intel is having problems ramping the speed of these things up in the way that pre-B2 stepping Barcelona processors could not be clocked up to 3GHz.


However in the preview of the dual-core chip - Kentsfield, they were able to take a 2.33GHz chip to 3.22GHz with little problem. They used the stock voltage in doing so, just upped the bus speed.

Granted a quad-core puts out more heat than a dual-core, so likely this sample wouldn't scale as easily.

So I think they would be able to ramp up the speed rather quickly if they feel they need to.


RE: Not bad
By retrospooty on 8/24/2007 9:35:04 AM , Rating: 3
"1. Cache is there to compensate memory access latency
2. More cache is needed to compensate inter-cpu communication
3. Higher clocked is the CPU, it makes more inter-cpu and memory requests.
4. Clocking higher the CPU retains the original quantity of cache"


What you aren't thinking about is that the cache runs at full speed, meaning a 2.33ghz Penryn's cache is runing at 2.33ghz, and a 3.33ghz is running at 3.33. All I am saying is that I believe the higher speed bins will break away a bit more than 5%. Not likely going to be a huge difference, but it will be faster. Previous reports here had a 3.33ghz penryn running an average of 20% faster than a 3ghz Conroe. 10% of that was of course due to the direct speed increase.

Future benchies will tell for sure though.


"And boy have we patented it!" -- Steve Jobs, Macworld 2007

Related Articles
More "Penryn" Benchmarks Revealed
August 22, 2007, 2:25 PM
Intel Sets "Penryn" Launch Date
August 14, 2007, 6:18 PM
"Penryn" Benchmarks Hit The Web
August 7, 2007, 4:04 PM
Quick and Dirty AMD K10 Cinebench
June 6, 2007, 5:12 AM













botimage
Copyright 2014 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki