AMD Zen 2 12-Core, 24-Thread Matisse CPU Spotted in UserBenchmark

[Raevenlord]

A new development could shake up our expectations on AMD's new Ryzen 2 CPUs, which if true, could mean that previous rumors of much increased core-counts at the top of AMD's offerings were true. User TUM Apisak, who has been involved in multiple information leaks and scouting for the hardware world, has digged enough to find a submitted UserBenchmark that screams of a 12-core, 24-thread AMD Matisse part (an engineering sample at that, so keep your hats on for the presented clock speeds).

The benchmark list the used CPU via product code 2D3212BGMCWH2_37 / 34_N (H2 is indicative of a Matisse CPU The benchmark is listing a base clock speed of 3.4 GHz and an average boost clock speed of 3.6 GHz. The rest of the system specs are very, very basic, with 4 GB of 1333 MHz DDR4 memory being used on a new AMD platform, based on the Myrtle-MTS based chipset. The processor is listed having a 105 watts TDP and 32 MB of L3 cache.



View at TechPowerUp Main Site

 

[goldman]

ryzen 2 or zen 2?

 

[SIGSEGV]

with 4 GB of 1333 MHz DDR4 memory being used on a new AMD platform

drrooooooollls...

promising...
i hope there's 16 cores of the next ryzen
16 cores then take my money pls

 

[KarymidoN]

hope this is not TR4 or AM4 flagship, keep playing hard AMD. the more competiton we have better the performance and the price...

 

[Raevenlord]

ryzen 2 or zen 2?

=( not on my A game. Thanks.

 

[efikkan]

I would expect there to be many such benchmarks in the coming months, but keep in mind that AMD still doesn't have the final stepping yet, so any claims of final clocks will still just be speculation.

I don't use "UserBenchmark" myself, and have no idea how reliable and consistent it really is, but I put together a small comparison with a few other results from the same page:

(keep in mind some of these might be slightly skewed due to some benchmarks being overclocked)

I am curious if the 4GB of memory had any impact though.

 

[Valent117]

my i7 4790k locked at 4ghz, delidded and 4x4gb of 2133mhz CL9 ddr3 memory gives me 118 single, 407 quad and 617 multi core score

 

[Xzibit]

drrooooooollls...

promising...
i hope there's 16 cores of the next ryzen
16 cores then take my money pls

Gamers Nexus if I remember correctly said they would be 16c & 12c in one of their videos. It wasn't a AMD source but I'm guessing a board partner source.

 

[ironwolf]

Lisa Su all but [officially] confirmed that there was more room on the CPU for another chiplet. That should be convincing enough that there will [eventually] be >8 core Zen 2 CPUs at some point.

 

[TheLaughingMan]

Single channel memory at 1333 MHz? Where they just making sure it wouldn't explode or something?

 

[ghazi]

32MB L3 cache with 12 cores is a very bizarre configuration given what we know about Zen. Rumors indicated the octacores would also have 32MB, and that the quantity per core would be doubled. If the 12-core also has 32MB that leaves few options for how it's configured; possibly half the L3 on each die is lasered off for this SKU.

 

[beautyless]

32MB L3 cache with 12 cores is a very bizarre configuration given what we know about Zen. Rumors indicated the octacores would also have 32MB, and that the quantity per core would be doubled. If the 12-core also has 32MB that leaves few options for how it's configured; possibly half the L3 on each die is lasered off for this SKU.

Is it possible L3 on the IO die?
So, 8 core = 32MB, 16 core = 32MB

 

[EarthDog]

=( not on my A game. Thanks.

It's in the body as well..

 

[ghazi]

Is it possible L3 on the IO die?
So, 8 core = 32MB, 16 core = 32MB

I had that thought at first too but L3 on the I/O die is impossible with CCX architecture, not to mention the potential latency penalty. I think what's going on here is that the 8-core uses a single die with 32MB, this 12-core uses two dies with half lasered off for 32MB, and the 16-core will have 64MB. Alternatively the rumors of doubled L3 could have been false but I seem to recall there being an 8-core benchmark score showing 32MB as well.

On a semi-related note it would be a very interesting twist if the I/O die turned out to have an L4 cache. I could see that helping with the MCM design and giving a nice uplift in games and other programs.

 

[Captain_Tom]

So out of curiosity I looked up what a 2700X gets:

https://cpu.userbenchmark.com/AMD-Ryzen-7-2700X/Rating/3958

A good 2700X system seems to get ~1300 in multicore and 120 in singlecore..... with 3GHz DDR4. So this (supposed) 12-core Zen 2 CPU is getting a 34% higher multicore, and only 20% lower single core while using 1333MHz ram and 10-15% lower clocks.

Hopefully this indeed means that Zen 2 will have more than enough bandwidth with dual channel DDR4.

 

[Darmok N Jalad]

Even with the slow clocks, that’s 9W per core, but it would be less than that with all the other logic drawing power. A hex-core r2600 runs at somewhat higher clocks at 10W per core, but bolting 2 of those together would be 130W.

 

[lexluthermiester]

Is it possible L3 on the IO die? So, 8 core = 32MB, 16 core = 32MB

That might be possible. Theoretically, that arrangement could have logistical, time and cost savings.

 

[R0H1T]

I would expect there to be many such benchmarks in the coming months, but keep in mind that AMD still doesn't have the final stepping yet, so any claims of final clocks will still just be speculation.

I don't use "UserBenchmark" myself, and have no idea how reliable and consistent it really is, but I put together a small comparison with a few other results from the same page:
View attachment 115054
(keep in mind some of these might be slightly skewed due to some benchmarks being overclocked)

I am curious if the 4GB of memory had any impact though.

It certainly did, probably single channel as well.
Take a look at the quad & multi core scores for the sample, worse scaling than 2700x which suggests gimped memory &/or awful clocks. Now with just 4GB RAM I'd guess it's the former.

 

[Imsochobo]

It certainly did, probably single channel as well.
Take a look at the quad & multi core scores for the sample, worse scaling than 2700x which suggests gimped memory &/or awful clocks. Now with just 4GB RAM I'd guess it's the former.

I suspect awful clocks and or identification for scheduler putting 2+2 between the dies rather than 4 in same die.
Also memory does impact it and at the moment there is too many suspects.

Memory latency is on par with zen+, however this is at substantionally lower clocks so we should see it better too.
Thus we see single core and latency improvements which should help out gaming and that is practically confirmed.

 

[efikkan]

Single channel memory at 1333 MHz? Where they just making sure it wouldn't explode or something?

According to the source, it's a single 4 GB 1333 MHz ("2667 MHz" DDR effective clock).

I don't know about this specific benchmark, but I would have run it with dual channel memory and at least 16 GB to make sure it's not bottlenecked.

Considering Zen+ supports DDR4-2933 JEDEC, I think it's safe to assume Zen 2 will at least support this speed, probably DDR-3200 JEDEC too.

 

[TheLaughingMan]

No. The straight forward answer is each CCX chiplet has a max of 8 cores with 16 MB L3 cache. With two chiplets you just double everything so 16 cores and 32 MB of L3. If the chiplet has defective or unusable cores, you cut the core count down to 14, 12, or 10 which ever AMD things works best for the price with no other changes.

In this case I think we got a chip made of two chiplets with only 6 functioning cores each, but full access to their 16 MB of L3 so 32 MB total.

I still don't think this was a real "lets test performance" benchmark. Thinking about it now, I personally now think this was testing the Infinite fabric between the I/O and two chiplets to test for bottlenecks or traffic flow issues. Otherwise, why would they gimp the config with 4 GB of single channel memory.

 

[EarthDog]

Why would the amount of memory matter? Data has to go through cache regardless, yes?

 

[TheLaughingMan]

Why would the amount of memory matter? Data has to go through cache regardless, yes?

It matters if you were trying to show off performance. If the benchmark or test you are running has a clear history of being affected by memory speed, capacity, or latency you should do whatever you can to minimize those variables. If that is what you are trying to show.

I pointed it out because I don't think this was about performance or these are final clock speeds or top configuration. I just think this was a function test and a way for AMD to show Ryzen 3k will have more than 8 cores. Thus whoever did it just stuck in the minimum RAM needed to get the system to boot.

 

[efikkan]

I still don't think this was a real "lets test performance" benchmark. Thinking about it now, I personally now think this was testing the Infinite fabric between the I/O and two chiplets to test for bottlenecks or traffic flow issues. Otherwise, why would they gimp the config with 4 GB of single channel memory.

If we assume all the information reported in the benchmark is accurate, I see one logical explanation; this is a lab PC testing for a specific testing purpose, like at a motherboard vendor or similar. I really see no other reason to use such an old GPU and a screen of 1280x1024. Even the BIOS date is 2019-01-10. This setup is clearly meant for some kind of basic testing, not benchmarking. I guess someone with access to the machine ran a benchmark for fun and submitted it.

 

[Imsochobo]

If we assume all the information reported in the benchmark is accurate, I see one logical explanation; this is a lab PC testing for a specific testing purpose, like at a motherboard vendor or similar. I really see no other reason to use such an old GPU and a screen of 1280x1024. Even the BIOS date is 2019-01-10. This setup is clearly meant for some kind of basic testing, not benchmarking. I guess someone with access to the machine ran a benchmark for fun and submitted it.

it auto submits.