Finally . . . Phenom!

Nathan Edwards

As any supporter of a losing sports franchise knows, it ain’t easy being a superfan. For the last two seasons, AMD loyalists have watched Intel’s Core 2 Duo and Core 2 Quad embarrass the Athlon 64 and QuadFX off the field. Yet devotees have chanted the refrain of the truly faithful: Come next season, baby, watch out!

Well, it’s next season and AMD’s chance to prove that it’s still a contender is finally here. This month, we bench, dissect, and ponder the hell out of AMD’s new CPU to find out if the Phenom lives up to its name.

The Scoop on Phenom

Before we get down to the business of benchmarking, here’s the backstory on AMD’s new CPU.


AMD’s “true quad core” jams all four cores onto a single 65nm, 285mm2 die. In addition to other core-efficiency enhancements, AMD now uses a shared 2MB L3 cache that runs at the same speed as the memory controller, which is currently 1.8GHz or 2GHz.

Q: How do you pronounce Phenom?

A: It’s fee -nom, not fuh -nom.

Q: What advances does Phenom offer?

A: Phenom is AMD’s first quad-core processor and is touted as a “true quad core.” Based on a 65nm process, Phenom uses an enhanced version of the stellar K8 Athlon 64 core, which features many of the same “wider and faster” techniques as Intel’s Core 2 Duo. Improvements over the Athlon 64 include the ability to execute SSE instructions in 128-bit chunks versus 64-bit. Cache speed gets a bump, as well, with L1 going from 16 bytes per cycle to 32 bytes per cycle, and L2 going from 64 bits per cycle to 128 bits. AMD also spends silicon on increased floating-point performance; a few new instructions; HyperTransport 3, which nearly quadruples the bandwidth over previous implementations; and more L3 cache.

Q: What’s meant by “true quad core”?

A: Each Phenom features four execution cores on one single, contiguous die. Architecturally, it’s far more elegant than Intel’s quad core, which fuses two dual-core chips in a CPU and forces the dual-core islands to talk to each other over the front-side bus. Phenom was designed from the get-go as a quad chip, and each core communicates at HyperTransport 3 speeds—several orders of magnitude faster than Intel’s front-side bus. All the cores can also share data stored in the L3 cache, so a core would have to reach out only to the L3 instead of the much slower system RAM in certain applications. This adds up to a chip that, on paper, seems to at least equal—if not exceed—Intel’s Core microarchitecture.

Q: What clock speeds will Phenom come in?

A: Just two speed grades will initially be available: a 2.2GHz Phenom 9500 and a 2.3GHz Phenom 9600. In a few months, two additional speeds will be offered: a 2.4GHz Phenom 9700 and a 2.6GHz Phenom 9900. While all feature the same microarchitecture and cache amounts, there are key differences. The 9500 and 9600 are much cooler at 95 watts apiece. The 9700 increases to 125 watts and the 9900 hits a Prescott-like heat dissipation of 140 watts. You do get something in return, however. The 9900 will run its HyperTransport link at 4GHz compared to 3.6GHz in the lower-clocked parts, and its memory controller runs a bit faster at 2GHz versus 1.8GHz in the others.

Q: Why are only two of the four CPUs available now?

A: AMD’s true quad-core approach sounds great on paper but it’s also directly responsible for delays in getting the chip out and hitting higher clock speeds. We’ll remind you of the four-leaf clover analogy: While Intel makes its four-leaf clovers by fusing a pair of two-leaf clovers, AMD grows all-natural four-leaf clovers. Unfortunately, the latter are much harder to come by. AMD has admitted that problems at the fab are the reason for the late launch of Barcelona—the Opteron quad core—and with Phenom. And of the CPUs that AMD is able to produce, not enough reach 2.4GHz or 2.6GHz to launch the chips right now—thus the initial 2.2GHz and 2.3GHz CPU rollout.

The Top CPUs from AMD and Intel Compared
Model
Athlon 64 FX-74 Athlon 64 X2 6400+ Black Edition
Phenom 9600
Phenom 9900
Intel Core 2 Quad QX6600
Intel Core 2 Extreme QX6950 Intel Core 2 Extreme QX9650
Intel Core 2 QX9770
Clock Speed
3GHz 3.2GHz 2.3GHz 2.6GHz 2.4GHz 3GHz 3GHz 3.2GHz
L1 Cache
128KB 128KB 512KB 512KB 128KB 128KB 128KB 128KB
L2 Cache
2MB 2MB 2MB 2MB 8MB 8MB 12MB 12MB
L3 Cache
N/A N/A 2MB 2MB N/A N/A
N/A N/A
Front-Side Bus
N/A N/A N/A N/A 1,066MHz 1,333MHz 1,333MHz 1,600MHz
Execution Cores
2
2
4 4
4 4 4 4
Process Technology
90nm 90nm 65nm 65nm
65nm 65nm 45nm 45nm
Transistors
227 million 227 million 450 million 450 million 582 million 582 million 820 million 820 million
Die Size
230mm 2 235mm 2
285mm 2
285mm 2
286mm 2
286mm 2
214mm 2
214mm 2
Price Per 1,000
$300 $220
$283
TBD
$266
$1,000
$1,000
TBD
Interface Socket F Socket AM2
Socket AM2
Socket AM2
LGA775
LGA775
LGA775
LGA775
Rated TDP
125 watts 125 watts 95 watts
140 watts 95 watts
130 watts
130 watts
136 watts
Dual Socket Compatible?
Yes No
No
No No
No No No

More Answers on Page 2!

Q: Will Phenom work in my existing motherboard?

A: Phenom is designed as a Socket AM2/Socket AM2+ chip and should, therefore, drop right into the majority of existing motherboards, provided the motherboard maker updates the BIOS—and didn’t screw up on the board design (see our sidebar).

Q: Does Phenom have the same RAM issues that DDR2 Athlon 64s did?

A: No. AMD corrected the issue that limited the DDR2 Athlon 64s to whole-number RAM divisors. This, in essence, would force DDR2/800 RAM to run at DDR2/766. Phenom CPUs use a separate clock for the memory controller, so memory will run at its intended speed. Consequently, however, the memory controller no longer runs at the core’s speed. The memory controller on the 2.6GHz Athlon 64 FX-60 runs at 2.6GHz. On the 2.6GHz Phenom 9900, the memory controller runs at 2GHz and notches down to 1.8GHz for the 2.3GHz Phenom 9600. It’s not clear if or how this impacts memory performance; it’s still a good clip faster than what the memory controller runs at in competing Intel machines, where that part is located in the north bridge.

Q: What about online reports that AMD’s CPU contains a bug?!

A: First, every CPU released, and probably every piece of silicon, has bugs. Companies call them “erratum.” However, there are bugs and then there are bugs. In the case of Phenom, a last-minute big mutha of a bug was found in the translation lookaside buffer (TLB), a small cache used by the CPU to manage memory. AMD says that under very heavy workloads, such as in virtualization, the TLB bug could cause the system to hard lock. The company initially said the TLB bug was the reason it pulled the 2.4GHz and 2.6GHz Phenoms from launch but later recanted the statement, citing the aforementioned yield and volume issues.

Q: Is there a fix for the bug, and how might that affect performance?

A: A fix can be made through the BIOS, and AMD has informed board vendors how to implement the workaround. How much the BIOS change affects performance is hard to say. We, unfortunately, could not test the TLB patch, as it’s unclear whether our BIOS had the workaround implemented (we suspect it did not). Furthermore, AMD is forbidding board vendors from letting users toggle the workaround on and off in the BIOS. The website Techreport.com tested boards with and without the patch and reported that, depending on the test, the performance decline with the patch was anywhere from 0 to 60 percent.

Q: How does AMD answer these claims?

A: AMD says the bug is so esoteric that it is unlikely to lock the system. That’s why the company is pledging to let you toggle the patch on or off in a future update of the company’s Overdrive application. We must point out, however, that the bug is severe enough that AMD is reportedly delaying a ramp up of quad-core Opteron sales until it has a silicon-level fix, which won’t be until later this year. At that time, AMD will also release a 9550 and 9650 with updated silicon.

Q: Did you experience the TLB bug in testing?

A: We don’t know. We can say that the first Phenom CPU we received ran at only a third of the performance of an equivalent Intel CPU. That chip eventually went back to AMD for examination and we received a second CPU that performed more to our expectations. We tested the chip at various clock speeds and did experience two hard locks that could not be explained. We can’t say if the lockups were related to the TLB bug or simply immature drivers and BIOS. It does make us wonder if this problem is more serious than AMD has stated.

Q: If reviews are based on CPUs without the patch, doesn’t that misrepresent the CPU’s performance?

A: As old Ben said: “That depends on your point of view.” Because the 2.6GHz Phenom 9900 won’t ship until the winter begins thawing out, AMD will have updated “B3” silicon in place, and the performance numbers you see for the current chip should be representative. Of course, if you bought the Phenom 9500 or 9600 the day they came out, the performance numbers you achieve will likely be out of sync with those in most reviews, which were likely conducted without the fix in place. If, however, AMD is right and it’s very difficult to run into problems, you can simply flip off the TLB fix (when the updated Overdrive app is available) and get performance closer to what you’re seeing in reviews for the 2.3GHz part.

Q: Is Phenom faster than Intel’s part?

A: You’ll have to read our final benchmark report for the full verdict on performance, but the short answer is no. While the chip was close in some tests, AMD’s fastest Phenom, which won’t even be available for another few months, generally lags behind Intel’s midrange 2.66GHz Core 2 Quad Q6700 chip. Mind you, that’s virtually the same CPU Intel released more than a year ago using its older process technology.

Q: AMD has nothing to counter Intel’s top-end CPUs? What’s up with that?

A: AMD is spinning the Phenom story two ways: The first is that people need to stop thinking of CPUs as singular entities. Phenom, so it goes, is part of the Spider platform, which includes the quad-core CPU, AMD’s new 790FX chipset, and the Radeon HD 3870 GPU in CrossFireX mode—four cards running in tandem (see the sidebar below). Sure, AMD screwed the pooch getting Phenom clock speeds up and yields higher, but would you rather spend $2,000 on just a Core 2 Extreme and 2GB of DDR3 or a reasonably performing Phenom with four Radeon HD 3870 cards in it? Four!! AMD’s alternate spin is that, yes, it lags behind Intel today, but it’ll be back in the game eventually.

Q: Does that argument about four GPUs hold any water?

A: Until we actually test four Radeon cards in a box (no drivers were available to do so at press time), we can’t give you a definitive answer, but we’re not sure it’s actually enough to beat two GeForce 8800 Ultra (or even GTX) cards when combined with Intel’s fastest Core 2 Extreme CPUs. And in all things other than gaming, the Intel system will easily outclass the Phenom 9900. So we’re pretty skeptical about such a configuration outboxing an SLI/Core 2 Extreme box.

Q: How well does Phenom overclock?

A: It will vary from chip to chip, of course, but Phenom is not shaping up to be a great overclocker today. We didn’t get very far with our engineering sample chip and few other reviewers have either. And when you look at how the thermals ramp up for relatively minor speed increases, it’s no wonder. Going from 2.3GHz to 2.4GHz takes the thermals from 95 watts to 125 watts. Going from 2.4GHz to 2.6GHz jumps it up to 140 watts. Older AMD and many Intel enthusiast parts have high thermal ratings but only because they’re anticipating users to overclock the hell out of them. We suspect that the increased thermals for the two faster Phenom parts are more related to AMD’s issue at the fab.

Q: So its graphics cards are slower and its CPUs are slower—has AMD simply ceded the high end?

A: AMD tells us that it absolutely has not given up on the high end. Again, the company fully admits that it blew it on the Phenom clock speeds and yields, but says it is committed to turning the situation around. When that will happen isn’t known. It might take until the company’s 45nm process is online sometime this year or next to become competitive.

Q: What’s the deal with AMD’s tri core?

A: The tri core is being sold on the concept that if two is good and four is great, three is a perfectly attractive middle option. AMD’s tri core is primarily aimed at people who don’t want to pay for quad core but want some additional performance at a more affordable price. The CPUs are, as you might suspect, dies that won’t pass muster as quad cores but work fine with one core turned off. While some view this as selling defective chips, AMD says it’s business as usual. In the past, if a portion of a CPU’s 1MB L2 was bad, it could be sold as a chip with 512KB or 128KB L2, with the offending portion turned off. Like the higher-clocked Phenoms, the tri cores won’t be out until later in the year—they will carry model designators of 7 instead of 9. Since they’re the same chip as a quad core but with one core turned off, you can expect performance to fall in between their quad- and dual-core brethren.

Q: Is there any reason to even buy a Phenom?

A: If you’re a performance or overclocking freak, no. Intel is ahead and even AMD says so. But for folks with an existing AM2 board that supports Phenom (see sidebar), the CPU is a very easy, relatively inexpensive upgrade that gets you performance beyond Athlon 64. That should give die-hard AMD fans some solace. You might also be interested in Phenom if you buy into AMD’s Spider platform argument, but that’s unproven technology at this point.

Q: Where does AMD go from here?

A: AMD’s next stop is 45nm, which it says will be online at the end of this year. There’s likely to be a shrink of the Phenom core with some enhancements to get the performance up, but AMD’s CPU code-named Bulldozer will be the next chip to truly take on Intel. Bulldozer, which is due in 2009, will be a multicore design, but AMD hasn’t revealed very many specifics. The problem for AMD is that Intel is expected to make another jump forward with its chip code-named Nehalem, which will adopt AMD’s on-die memory controller and chip-to-chip communication techniques and feature four cores per die and an improved version of HyperThreading. With two quad-cores glued together under the heat spreader, a Nehalem would have up to 16 cores (eight real, eight virtual) available to the OS.

Is AMD's Spider the Model of the Future?

The PC is no longer about a CPU or GPU in isolation, it’s about “platforms,” says AMD. And the company’s Spider platform gives us a glimpse of what that means. Spider is based on the new 790FX chipset, which will support up to four Radeon HD 3870 cards in CrossFire and the quad-core Phenom—all for a pretty low price. AMD predicts that you’ll be able to build a quad-GPU machine with a quad-core Phenom for less than $2,000. If you went for Intel’s Extreme CPU, you’d spend $1,000 for the CPU and another $500 for the DDR3, leaving you just $500 for the rest of the components.

AMD says Spider is just a preview though. Ultimately, the company plans to have graphics cores integrated with x86 cores, making the platformization of the PC a foregone conclusion. Don’t believe it? Intel, which currently has x86 CPUs and chipsets, is heavily investing in graphics as well, and has also said it will eventually offer a product with a graphics core integrated into the CPU.

What’s not clear is how this will affect the salad bar formula we currently use for building a PC. Will platforms that have you order meal A or meal B replace our pick-and-choose world? Stay tuned.

AMD's Best Message Is Compatibility


AMD's AM2 socket

While AMD hit a rough patch with the short-lived Socket 940 and Socket 754 platforms, lately it’s been solid when it comes to providing an upgrade path. If you had a Socket 939 board, you could easily go from a low-end CPU to a spendy single core. And when dual cores came out, you could just drop one of those suckers in the very same board and it worked too. The same goes for AMD’s Socket AM2/AM2+ design. If you’ve been living with an older 90nm Athlon 64 X2 5000+ for two years, you should be able to update the BIOS and drop in a Phenom to get quad-core performance.

That’s not guaranteed, of course. The company says board design issues, and even the size of the flash memory used to store the BIOS image, could have an impact on Phenom support. How do you know if your AM2 board will run the new CPU? Obviously, boards using AMD’s new 790FX chipset will work, but there are two other ways to verify compatibility: Cruise AMD’s website, http://tinyurl.com/yrmmy4 , to see if the company has approved your board for Phenom yet. Or visit your motherboard manufacturer’s website and check its CPU-compatibility list before you make a purchase.

AMD has learned from its prior mistakes. Many Socket 939 users felt burned when the company made a quick transition to AM2 and turned the fab taps off on S939 CPUs. When AMD moves to DDR3 in 2009, it expects to have backward compatibility with AM2 and AM3 boards with its DDR3 CPUs. Overall, AMD gets a good grade for compatibility even if performance is a disappointment.


Phenom in Action

Okay. Enough about the CPU. Let's see what the benchmarks tell.

For our comparison, AMD provided an unlocked engineering-sample Phenom that we ran at 2.6GHz and 2.3GHz in Asus’s new 790FX-based M3A32-MVP Deluxe board. We compared AMD’s CPU to the original Intel 2.66GHz Core 2 Extreme QX6700 CPU that we received more than a year ago from Intel. While it carries the Extreme tag, the QX6700 is identical to the Core 2 Quad Q6700 except that it’s unlocked. That let us run the chip at both 2.66GHz and 2.4GHz to simulate the performance of a Core 2 Quad Q6600. The board used for the Core 2 chip was EVGA’s 680i SLI. Both machines featured DDR2 RAM clocked at 1,066MHz. Memory timing was manually set on both platforms and both used 150GB Western Digital Raptor hard drives and identically clocked GeForce 8800 GTX cards, as well as the same drivers. Once we were finished with the Phenom testing, we dropped in an Athlon 64 X2 6400+ for comparison. As we pointed out in the main story, we could not determine if the performance-crippling TLB patch was present in the Asus M3A32-MVP board we used, but we suspect it was not, so our performance for the 2.3GHz runs would be higher than that of a system with the patch applied. For our test, we did not trot out Intel’s 45nm 3GHz Core 2 Extreme QX9650 or the company’s new 3.2GHz Core 2 Extreme QX9770. We didn’t even pull out Intel’s older 3GHz quad-core— after the 2.66GHz Core 2 Quad beat up the Phenom, we decided to be merciful. And there was definitely no need to throw in 1,600MHz FSB chips or DDR3/1600. When one team is getting pummeled, you don’t grind it further into the ground. You let it slink off the field with a modicum of pride intact.

The Final Verdict

After all the trash talking, all the “true quad core” pimping, the result is a chip that’s slower than Intel’s cheapest quad core. And more expensive to boot!

The sad part is that while Phenom will be viewed as a big yawner, it’s not really a bad CPU. In fact, it’s a respectable chip for encoding and most applications. And it’s certainly damned faster than any Athlon 64.

But in the final tally, if you are into performance—and most enthusiasts are—Intel is the only game in town this season. There’s simply no comparison. On the other hand, if you’re vested in AMD and own a Socket AM2 board, it’s probably worth considering a Phenom, as it gets you two more cores, bags more performance, and little more hassle than a BIOS update and a little thermal paste on your hand.

The more troubling issue is the cost to AMD’s credibility. With its CPUs so far in the hole, does it have the resources and capability to make a comeback?


(click to see full-sized benchmark chart)

AM2

Around the web

by CPMStar (Sponsored) Free to play

Comments