Behold the Ryzen 3000 CPUs

[Metroid]

It just depends how you look at it.

My view on these small steps forward towards Intel ST performance are a constantly repeating confirmation that Intel really wasn't all that lazy all these years, they just made a substantial jump forward and perhaps even too much of it with Sandy Bridge. At some point there just isn't any low hanging fruit left and I think the last decade up until today and the near future confirms that for both AMD and Intel. AMD's Zen is however much better equipped to deal with that reality, being much more scalable and cost effective.

I guess this is due to the memory latency, if amd improves on that then it will have a killer on single thread performance. From what I saw so far, the best amd showed was the 67ns with 3733mhz cas 17.

It's really amazing what AMD have done here, they even told us what kind ddr4 to purchase.

 

[ratirt]

1: They want to sell X570. Nobody ever promised more than Ryzen 3000 would work with existing boards, not that you will get all the goodies.
2: Not all old boards have beefy VRMs and the new CPUs might pull more then they were original designed for.
If you are disappointed then take it out on AMD.

They want to sell CPUs and give a chance for the new PCI-e v4 standard to rise with the new x570 boards. "They want to sell boards" what a stupid thing to say, considering AMD is a processors manufacturer. You should think what you are writing really. on top of that the x470 and x570 have same socket and are backwards compatible.

I guess this is due to the memory latency, if amd improves on that then it will have a killer on single thread performance. From what I saw so far, the best amd showed was the 67nm with 3733mhz cas 17.

It's really amazing what AMD have done here, they even told us what kind ddr4 to purchase.

I been thinking about the memory and why the latency is so bad. I couldn't find what ram modules where used and what frequency. I assume the modules tested with the New Ryzen are not the best fit here.

 

[londiste]

There is no clear statement about the RAM CL used in the review.
He writes DDR4-3200 Dual Channel but link in the text goes to a DDR-3000 CL16 kit that does OC to DDR4-3200. Likely runs art DDR-3400 CL17/18 which would explain the latency difference compared to AMD slide.

I been thinking about the memory and why the latency is so bad. I couldn't find what ram modules where used and what frequency. I assume the modules tested with the New Ryzen are not the best fit here.

Latency is OK, just not as small as Intel's. The added latency cannot be removed with a better memory kit but can be minimized by using DDR4-3733 memory.
The reason for the added latency is chiplet design and memory controller in IO Die. This means RAM is inevitably an IF hop away from cores which adds latency.

 

[shmuck]

From what I saw so far, the best amd showed was the 67nm with 3733mhz cas 17.

Really hope to see Ryzen 3000 with memory overclocking and B450 reviews on day one. I have my eyes on a 3600 CL17 kit and wonder how well it would run.

 

[ratirt]

There is no clear statement about the RAM CL used in the review.
He writes DDR4-3200 Dual Channel but link in the text goes to a DDR-3000 CL16 kit that does OC to DDR4-3200. Likely runs art DDR-3400 CL17/18 which would explain the latency difference compared to AMD slide.

AMD said the latency is improved over the 2000 series but somehow I can't see it. have I missed something here or is it the x470 board causing this? Comparing the graphs with the latencies for 3000 series Ryzen and benchmark of 2700x there is no improvement.

 

[londiste]

AMD said the latency is improved over the 2000 series but somehow I can't see it. have I missed something here or is it the x470 board causing this? Comparing the graphs with the latencies for 3000 series Ryzen and benchmark of 2700x there is no improvement.

What they state is true but there is a little marketingspeak there to obfuscate the tradeoff

Memory latency is improved, just not for the absolute latencies in single chiplet (6/8-core) configurations. There is a considerable improvement for CPU configurations with more than 1 chiplet. In Zen/Zen+, memory was connected to a memory controller on one of the dies and the other die had to jump to the first one for memory access. First die got fast memory and the second one got a slow one. This was the problem they needed to address, not so much the absolute latency number. This was addressed by moving the memory controller to IO die. Now all chiplets have the same amount of latency to memory. It might be worse for some cores but avoids large differences between different cores (that does cause issues) and the latency is also better at average compared to Zen+ even if this is in large part due to improved memory controller and faster memory support.

There is clear and large benefit for Epyc and upcoming Zen2 Threadrippers, less so for AM4 CPUs. 12/16-core variants benefit from this and 6/8-core variant generally don't. With other improvements 6/8-core variants do not get a noticeable hit either which is awesome.

 

[Assimilator]

My view on these small steps forward towards Intel ST performance are a constantly repeating confirmation that Intel really wasn't all that lazy all these years, they just made a substantial jump forward and perhaps even too much of it with Sandy Bridge.

Technically it was the Core uarch - after NetBurst, it was such a massive improvement in performance and power consumption, and improved so fast over multiple generations, that consumers got spoiled to expect that sort of performance. So when the curve necessarily started to flatten due to all the low-hanging fruit being harvested, people started to complain - wrongly, because CPU design is *really freaking difficult*.

 

[londiste]

Technically it was the Core uarch - after NetBurst, it was such a massive improvement in performance and power consumption, and improved so fast over multiple generations, that consumers got spoiled to expect that sort of performance. So when the curve necessarily started to flatten due to all the low-hanging fruit being harvested, people started to complain - wrongly, because CPU design is *really freaking difficult*.

The situation is not that different right now.
- Back then, NetBurst > Core was a huge step (and Intel followed up on that with Nehalem and Sandy Bridge).
- Now, AMD made a huge step forward with Bulldozer > Zen and is follows up on this with Zen2.
It is not necessarily about the other stagnating but about doing evolutionary steps in design compared to revolutionary ones.

 

[Metroid]

There is no clear statement about the RAM CL used in the review.
He writes DDR4-3200 Dual Channel but link in the text goes to a DDR-3000 CL16 kit that does OC to DDR4-3200. Likely runs art DDR-3400 CL17/18 which would explain the latency difference compared to AMD slide.

Latency is OK, just not as small as Intel's. The added latency cannot be removed with a better memory kit but can be minimized by using DDR4-3733 memory.
The reason for the added latency is chiplet design and memory controller in IO Die. This means RAM is inevitably an IF hop away from cores which adds latency.

I guess he used a 3000 c16, 80ns, indeed he said 3200 and at that frequency might have been around c18.

The situation is not that different right now.
- Back then, NetBurst > Core was a huge step (and Intel followed up on that with Nehalem and Sandy Bridge).
- Now, AMD made a huge step forward with Bulldozer > Zen and is follows up on this with Zen2.
It is not necessarily about the other stagnating but about doing evolutionary steps in design compared to revolutionary ones.

Step forward in learning with bulldozer as it was shit, reason after that they hired Jim Keller.

 

[IceShroom]

1: They want to sell X570. Nobody ever promised more than Ryzen 3000 would work with existing boards, not that you will get all the goodies.
2: Not all old boards have beefy VRMs and the new CPUs might pull more then they were original designed for.
If you are disappointed then take it out on AMD.

It is the Mobo manufacture that decides the price of motherboard, not AMD.
You should blame mobo manufacture for high price with uselss RGB. And for 300/400 serise board who stoped mobo manufacture to design better board?? Its is the mobo manufacture and they should be blame for it.

 

[mstenholm]

It is the Mobo manufacture that decides the price of motherboard, not AMD.
You should blame mobo manufacture for high price with uselss RGB. And for 300/400 serise board who stoped mobo manufacture to design better board?? Its is the mobo manufacture and they should be blame for it.

Blame them to if it makes you feel better.

 

[IceShroom]

Blame them to if it makes you feel better.

I am telling this not for felling good, but its a fact.
Seeing this kind of attitude towards AMD, after RX 5700 Series, as people not getting RTX cheaper. But remember the more you buy RTX the more you save.

 

[Kapone33]

Yeah, I second that. Why wouldn't it?
X470 is high-end chipset, it is not ancient, it is AM4 that AMD has said they'd support at least until next year.
That particular board should be good in terms or VRM as well as BIOS.

If AM4 on 1st and 2nd Gen scales well with RAM speeds, a 4000+ MHZ kit on a X570 board should mean more performance gains.

 

[londiste]

If AM4 on 1st and 2nd Gen scales well with RAM speeds, a 4000+ MHZ kit on a X570 board should mean more performance gains.

There is little evidence that motherboard plays that large a part in possible memory speeds. X570 boards are hopefully decently built when it comes to RAM traces (one would expect them to be given the price range) but better 300/400 boards are the same. Memory controller is on the CPU package, BIOS has only supportive role, it should be easy to have the same memory controller functionality for Ryzen 3000 processor on both X570 and older boards.

AMD was pretty clear in their slides, 3733 is the sweetspot. Regardless on what the actual RAM kit can do you are much likely to benefit from lower timings on 3733 than from higher speed except when your load is heavily memory bandwidth dependent. It is not about RAM speed but about IF speed. Divider will kick in - manually or automatically, we'll have to see - after 3733 which results in a small latency hit.

 

[Kapone33]

There is little evidence that motherboard plays that large a part in possible memory speeds. X570 boards are hopefully decently built when it comes to RAM traces (one would expect them to be given the price range) but better 300/400 boards are the same. Memory controller is on the CPU package, BIOS has only supportive role, it should be easy to have the same memory controller functionality for Ryzen 3000 processor on both X570 and older boards.

AMD was pretty clear in their slides, 3733 is the sweetspot. Regardless on what the actual RAM kit can do you are much likely to benefit from lower timings on 3733 than from higher speed except when your load is heavily memory bandwidth dependent. It is not about RAM speed but about IF speed. Divider will kick in - manually or automatically, we'll have to see - after 3733 which results in a small latency hit.

Well I can tell you from experience that X470 boards gave a much more stable OC than X370 when it comes to RAM. I know that there have been BIOS updates to support more timings but there is no X370 or X470 board that supports anything higher than 3600 as an OC. Below is the listing for the X570 Taichi. There must be something about this platform that has MB vendors hyped enough to produce more skus for the X570 than X470/B450 combined.


X570 Taichi

• Supports AMD AM4 Socket Ryzen™ 2000 and 3000 Series processors
• Intel® Wi-Fi 6 802.11ax (2.4Gbps) + BT 5.0
• Supports DDR4 4666+ (OC)
• 3 PCIe 4.0 x16, 2 PCIe 4.0 x1
• NVIDIA® NVLink™, Quad SLI™, AMD 3-Way CrossFireX™
• 7.1 CH HD Audio (Realtek ALC1220 Audio Codec), Supports Purity Sound™ 4 & DTS Connect
• 8 SATA3, 2 Hyper M.2 (PCIe Gen4 x4 & SATA3), 1 Hyper M.2 (PCIe Gen4 x4)
• 3 USB 3.2 Gen2 (Rear Type A+C, Front Type-C), 8 USB 3.2 Gen1 (2 Front, 6 Rear)
• Intel® Gigabit LAN
• ASRock Polychrome SYNC

 

[londiste]

There must be something about this platform that has MB vendors hyped enough to produce more skus for the X570 than X470/B450 combined.

PCI-e 4.0, fastest ever processors etc, justifying (or necessitating) a high price?

There is evidence enough that Zen/Zen+ memory controller is not the limiting factor, IF is.

 

[IceShroom]

If AM4 on 1st and 2nd Gen scales well with RAM speeds, a 4000+ MHZ kit on a X570 board should mean more performance gains.

According to official AMD slide, speed over 1866MHz [3733MT/s] will not bring any improvement.
After 1866MHz Infinity Fabric will run at 2:1 mode, means Infinity Fabric's speed will decreas, will result lose of performence. So 1800MHz - 1866MHz with tite timing will do better than 2000MHz [4000MT/s].
But if you want to brag like Intel users that my RAM is running 4000MHz[actually it should be MT/s, as real speed of ram is 2000MHz].

 

[Kapone33]

There is evidence enough that Zen/Zen+ memory controller is not the limiting factor, IF is.

PCI-e 4.0, fastest ever processors etc, justifying (or necessitating) a high price?

It is the cost of the manufacture of the boards and the components they have populated them with that makes them so much more a premium than anything before like a natural 16 phase VRM controller.
There is evidence enough that Zen/Zen+ memory controller is not the limiting factor, IF is.

I am not saying it is the limiting factor. (Ice shroom pretty much covered that just now). All I was saying is that memory speed and compatibility improved from X370 to X470 and should be the same for X570.

 

[freeagent]

The latency is a bit of a concern.. I say that because on my old z77 system, my latency is down to the upper mid 30s. Seeing 80s is like ugh. I know Ryzen is strong, and bigger numbers all around make things go by quickly. But how does it feel? Is it snappy? The last time I had latency this low was on socket 939, and I don't want to let it go

 

[londiste]

The latency is a bit of a concern.. I say that because on my old z77 system, my latency is down to the upper mid 30s. Seeing 80s is like ugh. I know Ryzen is strong, and bigger numbers all around make things go by quickly. But how does it feel? Is it snappy? The last time I had latency this low was on socket 939, and I don't want to let it go

You are comparing DDR3 to DDR4. Latency is what needed to be sacrificed a little in the alter of raw bandwidth of DDR4

 

[TheLostSwede]

Well I can tell you from experience that X470 boards gave a much more stable OC than X370 when it comes to RAM. I know that there have been BIOS updates to support more timings but there is no X370 or X470 board that supports anything higher than 3600 as an OC. Below is the listing for the X570 Taichi. There must be something about this platform that has MB vendors hyped enough to produce more skus for the X570 than X470/B450 combined.


X570 Taichi

• Supports AMD AM4 Socket Ryzen™ 2000 and 3000 Series processors
• Intel® Wi-Fi 6 802.11ax (2.4Gbps) + BT 5.0
• Supports DDR4 4666+ (OC)
• 3 PCIe 4.0 x16, 2 PCIe 4.0 x1
• NVIDIA® NVLink™, Quad SLI™, AMD 3-Way CrossFireX™
• 7.1 CH HD Audio (Realtek ALC1220 Audio Codec), Supports Purity Sound™ 4 & DTS Connect
• 8 SATA3, 2 Hyper M.2 (PCIe Gen4 x4 & SATA3), 1 Hyper M.2 (PCIe Gen4 x4)
• 3 USB 3.2 Gen2 (Rear Type A+C, Front Type-C), 8 USB 3.2 Gen1 (2 Front, 6 Rear)
• Intel® Gigabit LAN
• ASRock Polychrome SYNC

The boards are quite different, as they use mid/low loss PCBs now, which allows not just for better signal integrity, but also for higher speed signaling. This was not the case on previous generation boards and this is one advantage that PCIe 4.0 brings, as it forced the board makers to up the quality of the materials used, which affected other parts of the board as well.

PCI-e 4.0, fastest ever processors etc, justifying (or necessitating) a high price?

There is evidence enough that Zen/Zen+ memory controller is not the limiting factor, IF is.

Not so infinite then, huh?

You are comparing DDR3 to DDR4. Latency is what needed to be sacrificed a little in the alter of raw bandwidth of DDR4

And it'll just get worse with DDR5...

 

[HD64G]

Technically it was the Core uarch - after NetBurst, it was such a massive improvement in performance and power consumption, and improved so fast over multiple generations, that consumers got spoiled to expect that sort of performance. So when the curve necessarily started to flatten due to all the low-hanging fruit being harvested, people started to complain - wrongly, because CPU design is *really freaking difficult*.

I agree but laziness is the main cause of Intel's CPUs being so slowly sdvancing in performance since they had 4C/8T since 2600K. We are talking about a 7 years stagnant CPU tech with very small improvements mainly derived from the manufacturing process, which once it stalled at 10nm, Intel is panicking about AMD's CPUs being much cheaper to make and with Zen2 even performance crown will turn sides. And that's why they were forced to employ Keller (who is the Zen arch design father) to design something to get them out of the dead end they have brought themselves into, just due to teir laziness. On the other side, the customers that were milked for so long by paying $500 per year or so when Intel changed socket and brought out a new cpu with +5% performance vs the previous one have a part in this tactics from Intel...

 

[Kapone33]

I agree but laziness is the main cause of Intel's CPUs being so slowly sdvancing in performance since they had 4C/8T since 2600K. We are talking about a 7 years stagnant CPU tech with very small improvements mainly derived from the manufacturing process, which once it stalled at 10nm, Intel is panicking about AMD's CPUs being much cheaper to make and with Zen2 even performance crown will turn sides. And that's why they were forced to employ Keller (who is the Zen arch design father) to design something to get them out of the dead end they have brought themselves into, just due to teir laziness. On the other side, the customers that were milked for so long by paying $500 per year or so when Intel changed socket and brought out a new cpu with +5% performance vs the previous one have a part in this tactics from Intel...

Every year? In some cases 3 months!